<%BANNER%>

Ultra-low power analog circuits for spike feature extraction and detection from extracellular neural recordings

University of Florida Institutional Repository
xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20101124_AAAABF INGEST_TIME 2010-11-24T14:19:10Z PACKAGE UFE0019615_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 4706 DFID F20101124_AABEUH ORIGIN DEPOSITOR PATH rogers_c_Page_084thm.jpg GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
8b59d58291506f07742a616fd97fff17
SHA-1
428ca20aecd6d5779ec7f62a6b8502ee59057fe6
26699 F20101124_AABETT rogers_c_Page_071.QC.jpg
1364e6062ff70b02e3e697a12a11e433
8b1e540beb3f76f12cbd58b665d546d9d5c4c8bf
37116 F20101124_AABDRF rogers_c_Page_017.jpg
7338ca48da09e621f873bec15361bffa
6c2030caeaa8ae49fbb382f4db071e78ae899c93
6984 F20101124_AABEUI rogers_c_Page_085thm.jpg
8fb08473e29734599536bb39ff297ca5
dc92158ef9aaad563c96734a5ade34ef34770806
19971 F20101124_AABETU rogers_c_Page_072.QC.jpg
dfc64da2d012a1edc014684b7801f176
06c58da95698e9ea924636d846ebdbe55fcd8e7d
91231 F20101124_AABDRG rogers_c_Page_018.jpg
769ee0f670cf4d97c60b3d4246178be1
72914c9745bf97ddebe9647b70c51a8ff00a0110
25897 F20101124_AABDQS rogers_c_Page_001.jpg
1cb1ef14d3c1e19585d355df94372f18
21b005091460809accdf04dce746dc4a83a03dc7
25855 F20101124_AABEUJ rogers_c_Page_086.QC.jpg
9a8938db9d23a96973e7d9395adcca58
b53f69fc005b00bfebad62ae9bdf0c7b3d89710d
6041 F20101124_AABETV rogers_c_Page_073thm.jpg
caa1df62a77ec58f09ceb01d4b07c6ff
6c301435ae7155d8fb13151dacf1aaa89d7b1ad5
89358 F20101124_AABDRH rogers_c_Page_020.jpg
f6b2db34678a2810bc55e99b86ca38e3
a1feac37d82055c640fdd72dd1d424646da71459
10059 F20101124_AABDQT rogers_c_Page_002.jpg
52730d082ae7efdda62fd9962693c76f
09db866a58d6ac18fa54bcd9ea6dbfacbf958e22
7107 F20101124_AABEUK rogers_c_Page_086thm.jpg
cc654ea460868e6d72562b6f239365b5
a3e32f2a4a3a37534136c5b393dfb6f9250af54d
79422 F20101124_AABDRI rogers_c_Page_021.jpg
8e3f6fc28544374e6dba080d749d2c0f
b4e88dcdb9dfa9c507ddc270b2dcd7149ae59aea
27283 F20101124_AABEUL rogers_c_Page_087.QC.jpg
268d80bb15cc93c6470ebda78b964652
4ea5a315944a79b477777488c8333342f7401bb2
7615 F20101124_AABETW rogers_c_Page_074.QC.jpg
1416f0dcff76bbd20d2e3244a6662042
01f66bc116e56044a9911c7cc7028376cf12a3b9
88635 F20101124_AABDRJ rogers_c_Page_022.jpg
ba0cc3a40593121c4c2af6bf670a508f
43871f9ac3f4af2517588201d25d97920d2e45a7
25953 F20101124_AABDQU rogers_c_Page_003.jpg
9baf87b2b86d9d9984509c63590dd5b3
2b4417faafba75eb971fd8a15f237f50d4e22107
5278 F20101124_AABEVA rogers_c_Page_102thm.jpg
f1ffffb425e16af4297982c07ed6c578
938841d330cbe812bfc474d3a43c8ea6777d6516
F20101124_AABEUM rogers_c_Page_088thm.jpg
33b417a79d629b7adef88641409fc0ff
867aa7a6771bca774f892d4f90628f720209af26
2677 F20101124_AABETX rogers_c_Page_074thm.jpg
eb6480de894f849af2d6bd3ec7c253bc
9f6edd63fc24779b2e80aff7907bc1fa2ed13a44
90852 F20101124_AABDRK rogers_c_Page_023.jpg
a4852c9b317cae8bed1cc7c30af190fe
7464d06a2611c5d8a3f478dd29d977ea1fb0386d
76196 F20101124_AABDQV rogers_c_Page_004.jpg
727c1db483d823bafe427976f49ba4ec
7170386ab7dac7f01651527635f628cd9d070774
9693 F20101124_AABEVB rogers_c_Page_103.QC.jpg
7c48bf0d2ad73f1ad7df4767fc722963
65f8b5ca55b48d725c6a7a44ed673ebf3e47afb6
22422 F20101124_AABEUN rogers_c_Page_089.QC.jpg
f6e31ffce6f50fdd1702b4c89907e64e
db1ec2af1837185f862474f4052869d33cc1b874
27578 F20101124_AABETY rogers_c_Page_075.QC.jpg
32bfecb4c8728b036db4b211bd5d621d
87e72c5cdca1db417605c592c81d70632e443cc8
86982 F20101124_AABDRL rogers_c_Page_025.jpg
8e5e51fd86277da7d24ad93be9ba1099
2d87e8ba4d8a784c0707ddf0279500d121e58cb5
32955 F20101124_AABDQW rogers_c_Page_005.jpg
720b8da13cb9e05cbb6f97ea9defc2cb
d533f29e01b4881ba51f4f1da52b7012a79713e2
27255 F20101124_AABEVC rogers_c_Page_104.QC.jpg
6ab7e9052d5e08e30c93f3b94bed974f
12ae2bc8942d8ebc5efd087617e91ac44597481b
6098 F20101124_AABEUO rogers_c_Page_089thm.jpg
8add013d9902c2098c7d8e74a5a97d61
7493754976547ff9c96b0febae7ab4506cbec4ce
24486 F20101124_AABETZ rogers_c_Page_076.QC.jpg
2aaf0c930228a7c42eea4c34f7da73a3
70550556f0ba2fc6603703fafe09d46ea55a04e0
90037 F20101124_AABDRM rogers_c_Page_026.jpg
504b0c31ba60d8b310e072ac7b00a89e
f5c3861aa8a8916bf1a24a7ba033f6f451c82718
29576 F20101124_AABDQX rogers_c_Page_008.jpg
42fd3e9a63d8137a0e4160c949398296
af84508e22de002252fa8628999d67e1e232911d
71394 F20101124_AABDSA rogers_c_Page_042.jpg
8360d26b428004257f4b40673d0e7d16
883d44f23cbbd99f26cf778f424106cbd8e57ed6
6712 F20101124_AABEVD rogers_c_Page_104thm.jpg
871c7ea2e41665b93725fa942cdb1279
9da4d6ea6b5fe67cff8b21efcd6efd7204ee5b96
11412 F20101124_AABEUP rogers_c_Page_090.QC.jpg
6304234697ba36796052d0896bfcc2a5
dc4df34b3e953247b160fa3347c473c80dcdc2ad
88019 F20101124_AABDRN rogers_c_Page_028.jpg
6e945aeb7c1e1b445b5fa62b72f3bf13
0ead0a599664cfc18671cd4d7fb1bfb9c8d79e13
31966 F20101124_AABDQY rogers_c_Page_009.jpg
fc10fb908688098c4da84064b952426b
f840d55c39d01d5bc61d876024d7688ca4816683
90250 F20101124_AABDSB rogers_c_Page_043.jpg
f8003f90efd4dc25caf9bb9f8e58a032
0d9d96decf0bb7b48615cf31620688d1b9ec0d6e
18759 F20101124_AABEVE rogers_c_Page_105.QC.jpg
6509d8af685172d9931399d55cfbf218
bcbb5fd635aae645da6829f1bb9de5bbbcc0cb4a
6554 F20101124_AABEUQ rogers_c_Page_091thm.jpg
8486d9febb642de345a1c7402a4b9eee
996de2024f9e99cc738749b1d5ffcdd15081e047
84429 F20101124_AABDRO rogers_c_Page_029.jpg
882ad5dde68122fabbd39c4dcc8e0a15
2a0778ed2749e9c7113f9bd9e44b208fe496910f
89449 F20101124_AABDQZ rogers_c_Page_010.jpg
c43821c7439128af812e86e30de1f9e7
939ff29047031112a6d702d657bbb42ed5051ba1
52093 F20101124_AABDSC rogers_c_Page_044.jpg
7b0fd1d7d65d4a3ee8cba9a9d3cdb970
4ce78e7bfb338522a0a8cf0b7af5ebdbefab7dee
5256 F20101124_AABEVF rogers_c_Page_105thm.jpg
c5d3cc925e360976f404322f6836def2
4e29ce6d6f26541d6080a63fd1a18386338c4545
19999 F20101124_AABEUR rogers_c_Page_092.QC.jpg
b7d9ca776c74069136127967e9ee7c46
d15b51813c0cd183f80dd98f0e8d77b6a4f97dc6
87048 F20101124_AABDRP rogers_c_Page_030.jpg
84fe2662b7897e7ebf34ada8e19d87a0
c4b5512cef0960211eaf238e264106a0036bedb6
39211 F20101124_AABDSD rogers_c_Page_045.jpg
5c24eb10af63759762cf5bc985ea9fc2
5dc32ad138d3be8af6f33b46ef0779d4f922f035
23747 F20101124_AABEVG rogers_c_Page_106.QC.jpg
882d60786ee5c99c12dfb5bbe8a388ad
23b41a98eb5f64c7cdfe0a08471a4af1b38b453d
11209 F20101124_AABEUS rogers_c_Page_093.QC.jpg
a329ba223f18856813b7f4c1c97a117b
d994d20ac2d75009684e79c6520a19599b3d37dd
86241 F20101124_AABDRQ rogers_c_Page_031.jpg
b36f970eea3088be1d079b89fe984d4c
7d1bbf6b3c4e9b3ea9ebf73ca416a9f051d93722
64508 F20101124_AABDSE rogers_c_Page_046.jpg
0d674c09a0812230642a6313be3b0e5e
dee0d4f61c98db5134d583db584095642a5ac48c
6879 F20101124_AABEVH rogers_c_Page_106thm.jpg
6839991148256b293acbbae30fcbe1d8
aa55a70cbd7fb32eb5c7eb36fcb962bf497ffe8a
24475 F20101124_AABEUT rogers_c_Page_095.QC.jpg
2348090eb8a5f5c96db2a04b7b4eb56a
d3fcbc599f0416e0b53d52df07fe2f7b7231a717
85984 F20101124_AABDRR rogers_c_Page_032.jpg
1e0d38a5dcaa2fa2158bbc8062867b35
a553f860a5bc68c1d6914670029543cc42769711
82497 F20101124_AABDSF rogers_c_Page_047.jpg
306d7957a1d43071c7509fb85441e5ca
b77958be698dc9878b733ba2ecff226e92bf98fa
18020 F20101124_AABEVI rogers_c_Page_107.QC.jpg
d85f617bcb1871e7e94fa6535f4f9d9d
9e755492f6e3805baa99c18c4d5d54ea0e66768f
6039 F20101124_AABEUU rogers_c_Page_095thm.jpg
5cf2ebe50d2e3dc5d83f46c00e57ba4c
211d80624c96a2406f3a44dd0edd1abd6e027a60
42943 F20101124_AABDRS rogers_c_Page_033.jpg
6982bb99861b4553556e7e55269dfa69
4ab2ed152ad904ca3a3786caa9aae53b2a91f55c
43944 F20101124_AABDSG rogers_c_Page_048.jpg
dcbd62d6bb8e52d2c71ece7c355c763e
31c7431ee1db7cfcc99b5bec03dab65c03207551
25611 F20101124_AABEVJ rogers_c_Page_109.QC.jpg
61ea77ec8db8157f493a48fb845a6f59
960078abfb32a6a3288f0dd4d07574409c230cc3
4235 F20101124_AABEUV rogers_c_Page_098thm.jpg
98b60879f50cc7c1bd67f3b724632e4a
fc40d1c3099091dc60e9ebcca19669c3f7de4cdf
83689 F20101124_AABDRT rogers_c_Page_034.jpg
ab1b591ded17d79d4e4e418eebf16c16
2fd185dc4f61f2165bbafac7110dba95a114534a
83336 F20101124_AABDSH rogers_c_Page_049.jpg
9c42a9fd3994533cac6885adf0874d97
f4c92b760ab7d6215ed8f5d0f26a479cae694c2e
6443 F20101124_AABEVK rogers_c_Page_109thm.jpg
313dc395df3ade2ca9ba60ef1c079fa2
e8cc9f644006c9b1423ceac3fb53924a826014b7
26912 F20101124_AABEUW rogers_c_Page_099.QC.jpg
3aa7d65b1bf5f8f9594dbdf3a64dda32
7af5212a14e05b8c87e36ff78d07f63aa1532a33
38163 F20101124_AABDRU rogers_c_Page_035.jpg
8f0c92805b9ea14e53e4797542632e1d
dad434a9aa2e3d1057154aab53adbeb2993db89a
69946 F20101124_AABDSI rogers_c_Page_050.jpg
6c137b08febfbbfdbed9b7959b0ff0cd
a7e1f5f525c9e25cfc81da8f15aee0889342c070
14468 F20101124_AABEVL rogers_c_Page_111.QC.jpg
48710799b87e31f0caad89caee97179e
89723a55c6e00ea73f15c96c917810a7f0c206b2
81754 F20101124_AABDSJ rogers_c_Page_051.jpg
a9319b3a6e2e622ce6dcfc32dc2308a3
71fa8064a42841c19ae25c25e97b0d64563b893c
1859 F20101124_AABEWA rogers_c_Page_120thm.jpg
2f5995bf7b6d12638259e9b78b0328bb
ff6673ec93737ef4f1cdbc826451ef46b1698f48
27621 F20101124_AABEVM rogers_c_Page_112.QC.jpg
d1226d51b86de40fdd2c1a6eaa69c59c
fd97a2d76eae506f723c8001e493b92db79fbc91
8827 F20101124_AABEUX rogers_c_Page_100.QC.jpg
1552ea102862a97c6a80fdbb57052df2
5795d650e5be386a0ee178d026fe5427b55146fb
69324 F20101124_AABDRV rogers_c_Page_037.jpg
632539d3a0ec6e8d84eb7d01307d2ebe
90a05354a85185a6ef9aef170119a2beee149359
82345 F20101124_AABDSK rogers_c_Page_052.jpg
dbc7cd9ca622c60b7e03c6e7a98e63d0
a2724b308813f6d2ee584cff67bce97741685b8a
23903 F20101124_AABEWB rogers_c_Page_121.QC.jpg
f9160145ca815d3cf8e1a83a3f040181
e2e5c0058068b07a99064320ac8bd8a9f2e94904
6774 F20101124_AABEVN rogers_c_Page_112thm.jpg
23826ff70c8f52f70d8958b677362c83
d7c59598aaffbff01679f51adb16795e5321c53f
2702 F20101124_AABEUY rogers_c_Page_100thm.jpg
a221be6f8158f44e20cc56beeb3f1dfe
47ba2007687a6446807ad159ad665103209e264d
72873 F20101124_AABDRW rogers_c_Page_038.jpg
e45bba8fcb6f6d6b2ab02dc9a5f11956
822d4b1af2badcd1ad86f61634f54cdaef405dbf
58565 F20101124_AABDSL rogers_c_Page_053.jpg
b881929f33fd67bd8299e17e989ef3fa
f954072836da3a968274180fe87c39c003b4e5a3
6855 F20101124_AABEWC rogers_c_Page_123.QC.jpg
cc86b06e1e2752f9bc7bb6132611fa41
7325cc44e6468033c8491cdccacdce5ce04d1c9a
6775 F20101124_AABEVO rogers_c_Page_113thm.jpg
55a5c701af0d2cd49e670d453baba54b
7379c706c4a9bc07b704cf0b1f1f53ceff1192ee
6463 F20101124_AABEUZ rogers_c_Page_101thm.jpg
751b25251f50ffbdc27105608c284560
79228a92b9449adde7bec593733faeb0fd964c15
54503 F20101124_AABDRX rogers_c_Page_039.jpg
2504db76bb04758d3a0c8ba057b05665
7301d59f08c8c04a682bacd1a7438ec18e531f8e
91673 F20101124_AABDTA rogers_c_Page_068.jpg
484f2d8bc93bb811d48f64617159bb47
90243dd10e2d2276e318f25d759057895c46e012
27175 F20101124_AABDSM rogers_c_Page_054.jpg
2fa18fa52ae271f776fc564c110cf136
4ae46bd643e2fa7353730892139ab702eef9fe69
6777 F20101124_AABEWD rogers_c_Page_125.QC.jpg
a61058888a093d211a903dbb5162cc60
aa76289f3a595dd142886916ea8ea80d0a27cf02
20623 F20101124_AABEVP rogers_c_Page_114.QC.jpg
f1f5131d132f81ba2f53a65c6a93aecc
bb51a078fd7df09e4689f564d132b0e3650a2593
86781 F20101124_AABDRY rogers_c_Page_040.jpg
8a9166d65427b052daa3e105c40bd565
b67b1f251418d9009b681f957b2f1ef2184b64f1
43937 F20101124_AABDTB rogers_c_Page_069.jpg
398049c9046d39ccd15eb06e7bab402f
c1b893d12ff9ed95259372af9f745a112b421c7d
64047 F20101124_AABDSN rogers_c_Page_055.jpg
4a865c05c76271c26df105ed0f3ec447
de130ac2141fece58bbd2f03ee86d90a45390831
24301 F20101124_AABEWE rogers_c_Page_127.QC.jpg
f7a8202aa19dcb09a4c1fd1469c37bf2
d6e8eac40cbdd46111cdf8d76df541f86d4db12c
5824 F20101124_AABEVQ rogers_c_Page_114thm.jpg
469e9049b45f4de8200178051988cc9f
ff9d00b8cb9f6164f1285160f4b9f29c4a10f14f
80371 F20101124_AABDRZ rogers_c_Page_041.jpg
067939ecf68dfef89bae945a6aa72a60
0a29e69b90a316c7173cc4fdb57647f1eaa20e30
86074 F20101124_AABDTC rogers_c_Page_071.jpg
7ec6423f473c9ebc3b9a0874b6b1c824
d422bf2ff163fecd1480c8c6f9d6801da171af7c
47203 F20101124_AABDSO rogers_c_Page_056.jpg
94063d8bae830507b304ff2c9f3e6453
7c415c9f085ad5ccff106d89ab9c6d576e1788f5
6331 F20101124_AABEWF rogers_c_Page_127thm.jpg
bc8de3754b0ee47bdf341e8c0c51b950
cde5d4cbf7011cfce328ec70da5cf11a861b3e47
22903 F20101124_AABEVR rogers_c_Page_115.QC.jpg
b5a2834416302c8ac242d13d1e7d7199
3bcf2028efb1166b81e16381219d4a9b8de6715a
64436 F20101124_AABDTD rogers_c_Page_072.jpg
fffd0432663a1c2f9f5b62c145e01025
542884b2d38833d67cef56bb678166dfbaed7226
26910 F20101124_AABDSP rogers_c_Page_057.jpg
7de2aad69e0476f8189fb32727cf26c6
ad97d76f48c9d603b64334ea27e56a285ab9a5e0
26069 F20101124_AABEWG rogers_c_Page_129.QC.jpg
50709ed5751bcff92d7471e7b7003df0
a8ffe15f34b226ece956f679f484c3cd992087c6
6362 F20101124_AABEVS rogers_c_Page_115thm.jpg
6c7409645b66111639266c40014e81dc
7e74b0e013d67460f59c4f5cd37f522f01bac23b
70219 F20101124_AABDTE rogers_c_Page_073.jpg
9014aac19b55d77d2f9e6e1e32e01f38
b79666fcd90968687c88b25e4df1145de52e616a
88930 F20101124_AABDSQ rogers_c_Page_058.jpg
e2fa271df63fa0ecf93bc6f81e82c03d
9f09dbe698004b00e0a5c5c92e76ed674e4b8a13
6305 F20101124_AABEWH rogers_c_Page_130thm.jpg
aadde6e5ed3e5d83c43f4f89597e8b1b
bf6f6c16651c352be9e2a86a584f36c414c94871
18897 F20101124_AABEVT rogers_c_Page_116.QC.jpg
12888c2e97af866957561fc6fda8a6db
9bd339094dfedfec8cb727f3629a911b16d5e1a7
24684 F20101124_AABDTF rogers_c_Page_074.jpg
60528fb256151415ceb9fcd2874fd367
ded88a7e4b477f2ccb76d38c2b8ba2593584d065
68596 F20101124_AABDSR rogers_c_Page_059.jpg
0340e99638883ebeb66437aa60d7fed1
c615152c82c1e27ae40fc94049838c742cc34039
16116 F20101124_AABEWI rogers_c_Page_131.QC.jpg
1e74c9d14cb911b529bd10d0c60ede0f
f1822676b17fb92e2da4a06f46e1fecb283ae21d
5757 F20101124_AABEVU rogers_c_Page_116thm.jpg
482cab6abed848ef31b83ce17ef330c8
11e07ffbe6ea1b905b1469a7c24b5132ec65cf36
90067 F20101124_AABDTG rogers_c_Page_075.jpg
a98e3e24ff4a23c4344dec3f650a7dc7
54a74fa1edb6adb975ebb99abe25b7a1bd34febc
36783 F20101124_AABDSS rogers_c_Page_060.jpg
ca79100b17926401643af2f35d8667b7
bfa19c528652c472664daf4fc05a89380353b2fb
4286 F20101124_AABEWJ rogers_c_Page_131thm.jpg
50aa10abe221d76e6febcd18a1ca6495
b76c08b042b132720a83c8380e863c3cb9cf0cb9
16612 F20101124_AABEVV rogers_c_Page_117.QC.jpg
8afc710b50edbc209207bab3607314ed
e60eab12bf58e1e1bc7125c8d266d40a0f92df6b
77331 F20101124_AABDTH rogers_c_Page_076.jpg
c9d3199839c50dd5665e31caf6ff118e
bea723d9b11e125c168826260eb34acda9a33042
80970 F20101124_AABDST rogers_c_Page_061.jpg
223a22a508b266ce3a76262d0a9a441c
86a41801ce4800fb86ae9e42c64596d199367db0
4282 F20101124_AABEVW rogers_c_Page_117thm.jpg
b77004451ed23eb2009a9cc75d8e8172
2f56b7347f163ca87521554777e5f0f22f0dc2e9
79064 F20101124_AABDTI rogers_c_Page_078.jpg
e8a5f8dd1070d2bd1f496413ea8765fa
7a6b7dd05493525b6f0a1ea8525372922eafbf27
48259 F20101124_AABDSU rogers_c_Page_062.jpg
f3c7b90840f8c041179d7c46afdb30a6
4ad0cfba0e08c9ff0bfafa26e6e545c46063d9a9
24039 F20101124_AABEVX rogers_c_Page_118.QC.jpg
ecb25b24fd6641959d74c6ecbadf2d74
ea4fe69d7cce73b6561bacb65f284b0b5c951039
59156 F20101124_AABDTJ rogers_c_Page_079.jpg
0f892bc468152a1b4e0b3921c2a6d84c
0ad2dbaaaa8bf7a4e68611f53efcf103caf2007a
27586 F20101124_AABDSV rogers_c_Page_063.jpg
768cca235b4932feff7f56c3c92790c0
a0ed598f0afb667dd9679438ddb84b73fa2a16dc
68162 F20101124_AABDTK rogers_c_Page_080.jpg
4b4a54663112e1e8b5213146c8e49400
fd8dabf6c498df877138f08956bf6e3656de0f4a
25369 F20101124_AABEVY rogers_c_Page_119.QC.jpg
cfb2cc99ce46ebfa4b4f9b0e908e9df5
6737ad982561300d8658b0cba8ef1699b2584b29
41230 F20101124_AABDTL rogers_c_Page_081.jpg
84ce12d5a5b434e8f49c43c0eabb74e6
df585eec329d7603b13e08ddb76442b6f1eb9fe7
75996 F20101124_AABDSW rogers_c_Page_064.jpg
e1dcbc88e34e095796772e0eb21c86cd
4036546427a524297ee13be4e83bbbeab1d5c317
6236 F20101124_AABEVZ rogers_c_Page_119thm.jpg
f870ebc11cfe6f259b201a362f0c2367
37f750e293043ba7b4aec35d5c960d68d77b8f91
66124 F20101124_AABDUA rogers_c_Page_097.jpg
a3a8e5527acc47d3b0e179b96f997932
fec608dc9331268a9640a5854ec33838f97479e7
61556 F20101124_AABDTM rogers_c_Page_082.jpg
7db360f73741b09dbb488639bc5f055d
11d81a40682c0d2b583f4576cbfac5a33d01198f
33697 F20101124_AABDSX rogers_c_Page_065.jpg
69753d3e0498786e0272b152ab00a312
b58a791e504245b697f3e5653b8940260bc2b7c4
53313 F20101124_AABDUB rogers_c_Page_098.jpg
1396518a8c7b1db4f61a52f941f5583c
0746cff376eaa32ca4202a0a2780ac41692afd93
58393 F20101124_AABDTN rogers_c_Page_083.jpg
d13efeeddabfa732f2cd01d69cf9fe8a
4f6dde5126bab350a3309405fbe12a3232b876a7
38483 F20101124_AABDSY rogers_c_Page_066.jpg
b9ef544cce6c7a34c4a5f61d82938b3e
cd76fd0f36190b80246a503d695587464a67fbf1
83951 F20101124_AABDUC rogers_c_Page_099.jpg
f52dd2bbbe9d83d01dde14cdc64ec4a7
aac179a8c439d87f89ef87bf4e9751f7ddf1cd82
49143 F20101124_AABDTO rogers_c_Page_084.jpg
1bf110af687ce18a01515be4fb01e761
edf7a7205fbb7b68bb2d86d920a61e65f1d98cf0
49552 F20101124_AABDSZ rogers_c_Page_067.jpg
eaa0f5c6fa02f50d462af58129c4b8ae
8deabf3b29fc75ba58c74fa8e48c829bbb8a6f49
32021 F20101124_AABDUD rogers_c_Page_100.jpg
24d1b6e5679cbc8f794de1585b056f50
542353210d239973a2d9c6d7262cbef28f351e6f
92487 F20101124_AABDTP rogers_c_Page_085.jpg
c37bc285a7af19b0d55e0f6f72dd9f55
9ae2a6b500e6a1735bf60f1b17e0b73d52787533
81038 F20101124_AABDTQ rogers_c_Page_086.jpg
89a21516d2c72bd214050c82ec142e18
a8722390bd21893eadbdd2a8c116d635134a9882
87968 F20101124_AABDUE rogers_c_Page_101.jpg
d9ebd3e6fc45e0d1f9e7e75e8560a7ec
cd278d2ba766e9daa5fe1ebb1d601c1286d9574c
87556 F20101124_AABDTR rogers_c_Page_087.jpg
5ed811c3eb8c96aeedea1520128002b3
0e324a26bfcc4b1ea8747023b8ada1535a26fa41
1053954 F20101124_AABEAA rogers_c_Page_005.tif
12fc8d9a6104afad5ae7d00d843390ba
e15f011fecc8e9b8ec97e8bdafa0cdcecc734db4
28948 F20101124_AABDUF rogers_c_Page_103.jpg
6a67ac3c2281e2d1520352ea8e9dafcd
15bc6b8835acf76ee4e16b3d255d4db6a938723c
79833 F20101124_AABDTS rogers_c_Page_089.jpg
a4cafd649dd1ef9911b00f9a98e1fba7
e6a479fb0e982721aab70b62d33b712c148fb58e
25271604 F20101124_AABEAB rogers_c_Page_006.tif
3bb250921524e8234939db99b2db3cbd
f025572d9758ece8073e6aff159fe649588a0b09
87287 F20101124_AABDUG rogers_c_Page_104.jpg
53db10171059172e8ffd64b1f3630525
bbf9962ad42626014d2a0eeefc1d6692a9737fe3
39671 F20101124_AABDTT rogers_c_Page_090.jpg
aa6d462706c62540d5a9ed8813d1d86b
37f5b016855e7f6a8932d7b9ab0b94699bd297ba
F20101124_AABEAC rogers_c_Page_007.tif
5ab452f97c06b7a030a0e518d230398d
557781cf9472475aca9c890f7eaa1167ffa7df98
58564 F20101124_AABDUH rogers_c_Page_105.jpg
a0b2c940abcbcf1540d726f8f7f07511
f931cd99518f6d1dda7ef0b56d4357f141e0b1b1
85312 F20101124_AABDTU rogers_c_Page_091.jpg
be1a17d4815cd6a4b0ab04f7151e98a1
675bf811798cce6a49bd7c1e49af7ca7a3b89b39
74693 F20101124_AABDUI rogers_c_Page_106.jpg
93071b17f082547287c53f61ef868878
b7528bccb560979116b3cc44dd07892e1a5bb11f
64032 F20101124_AABDTV rogers_c_Page_092.jpg
72125719e476a3246c800fc771590840
376f8e16639c83fab30bb513e2a7d14a34d950d0
F20101124_AABEAD rogers_c_Page_008.tif
451abbf4de7d2da8a01d6e8ea00f6d05
c2ce9d0b122cda243cf53e80bf3a9e4949dbc08f
59603 F20101124_AABDUJ rogers_c_Page_107.jpg
7f2bd28a5c5e95578bc04c8c8afae172
5a0c10411c167e2161313392ed3cb956daeadf9b
36348 F20101124_AABDTW rogers_c_Page_093.jpg
fada526f5056476515142a5c9f3b802f
2d4b0b09c7c0a7690a6c13121de9917a36dfdb58
F20101124_AABEAE rogers_c_Page_009.tif
b57a93f83bb59fcc11f21bf8bc28e336
04eafc6a5ff10c1d63ff4a66abdb6483693de613
66597 F20101124_AABDUK rogers_c_Page_108.jpg
47a6bac8fb1c784b6138ed76559ef8fe
b5c292096f791c9b8390e90a0223e62531080f97
F20101124_AABEAF rogers_c_Page_010.tif
d5e7b45b82e0e8ad6a89d1d390bf9dd3
933a0f70c6190b09cc1675c27bfe3f739a9fcbad
80838 F20101124_AABDUL rogers_c_Page_109.jpg
645781172bc3a20aa73ef0d566ad6f71
ab6401b4aa3392ddac0e759b6ce876a9003fe939
71291 F20101124_AABDTX rogers_c_Page_094.jpg
972b2d1e3acc7ab9208958a24f487304
9dd60c10ccdb2bd95de27d47c6b01f355ae20d24
F20101124_AABEAG rogers_c_Page_011.tif
030ec5c1767f79118e4ed5a10dd8da0e
5c6c920d784199d24f4bf578d5c07bfac9d266be
107679 F20101124_AABDVA rogers_c_Page_128.jpg
5a7940a3a518a24caa079e0ed42b98e6
1edc5588b834ba423c02814b5719e1d3c70fbb8d
47986 F20101124_AABDUM rogers_c_Page_111.jpg
a5f77290e3165aa5742252c8970c920a
823aea79b56c9b69d9f8932f20909caf31d43e19
76385 F20101124_AABDTY rogers_c_Page_095.jpg
86794d875a77a4856d5e60609e831a95
993aac21f5389cd9c93b9283a8fa098fa8f649dd
F20101124_AABEAH rogers_c_Page_013.tif
194c559c4918d8d1de128976ad717e0b
94368c7c904bcaf1931d38a90b73366ef1634d79
89144 F20101124_AABDVB rogers_c_Page_129.jpg
0a3e53bf8b5261241ea4d7f26aa5c757
64638794d64b3b08bde7c6494851d966e09f3cb2
88775 F20101124_AABDUN rogers_c_Page_112.jpg
b22995d5ad7a50822561f8eb71af300e
d42b585f08f454cfc4ddf72dfad2ad45937538cc
77638 F20101124_AABDTZ rogers_c_Page_096.jpg
bef1e20eeb979fa787a887a084ee81fb
4b14bf4b2e36694f330338a0bebcf684bad42b22
F20101124_AABEAI rogers_c_Page_014.tif
efc1dd98c4984f2281e0497dc946a064
d9b57e389300c6eb88e4016ad90800c140e18255
85114 F20101124_AABDVC rogers_c_Page_130.jpg
6d2dc716761da2d7b86dea757053710b
a095d03e0104681e362e1e64412412063ea0b17c
75314 F20101124_AABDUO rogers_c_Page_113.jpg
5cb81174d6ae8cc41e37435031e061d0
fa56b38e1a8391c477a29de6854661c7c5dbbf59
F20101124_AABEAJ rogers_c_Page_015.tif
27567bff6197b172c69a5f3edf2b9fd9
64b0e991e0eb053aa1d5a8d5c7c13d787ea65670
27971 F20101124_AABDVD rogers_c_Page_001.jp2
185ee1eddfe0c3b98c82326e0823f3ca
933ee16405cfc76fa4501020fdf25ae5f357bf8e
65875 F20101124_AABDUP rogers_c_Page_114.jpg
c8354cb07cd7b267b7fe40c507012370
3acf0690eca4778205204e599044c0b62e201604
F20101124_AABEAK rogers_c_Page_016.tif
b9e5aa42f97d3d9077f9968be2f2ba03
d20430d08d963c875ec112ee201f172abef617cb
5724 F20101124_AABDVE rogers_c_Page_002.jp2
929f2b841461699e1aceda1c3c7c9364
3e24b57034eaadc916f1733e5b4fd03968f370f2
73566 F20101124_AABDUQ rogers_c_Page_115.jpg
1adabcaddac0d8ce9de829f1ca1f4085
f3d79f9573d8878ffbc3e2bdd248eaf74bd37d28
F20101124_AABEBA rogers_c_Page_037.tif
086aeaa0023c0bcbd0248bd2f89752ec
b348c2914e32f1b5a0d76edcb9dd3e1eaaebe3dc
F20101124_AABEAL rogers_c_Page_017.tif
5edafd88903de67e548ffeb2429fb8ea
9953943f5654d09c88cbf3bb1654cd37bfb3b1f6
33669 F20101124_AABDVF rogers_c_Page_003.jp2
ce93c55bbd9d4850ae20bb1e0d65401a
74533a2398b4133a9226ba840b44342e856e4bab
60330 F20101124_AABDUR rogers_c_Page_116.jpg
beb48a402e4462d63dd589abce16d5e2
7d842912adb9bd4cab8134c0698baeac437187ae
F20101124_AABEBB rogers_c_Page_038.tif
1c4446a6434c76e21f3658d3f47aa840
2f4599f28bd9cafee2acf2871080419dbf2574dd
F20101124_AABEAM rogers_c_Page_018.tif
f82a49dca532bb8646b3d16316941643
1885fef42dca92cb67aace07b36f9501da7d2767
119557 F20101124_AABDVG rogers_c_Page_004.jp2
aabed60b0b365b29e6343ad99beba1fa
7dce1607987a8dd1e10cbc9d13ee908454ac8877
52709 F20101124_AABDUS rogers_c_Page_117.jpg
392c0b2610a4f3ae074a7254b2a2fff8
d4742644330b3f5c16025a26be71ee8c45ea948a
F20101124_AABEAN rogers_c_Page_019.tif
a9ce0292c68e00ac31d9bd45b2c345ef
c7ae3d10151a56bc728dad79274f438491df8b38
45709 F20101124_AABDVH rogers_c_Page_005.jp2
f3fa7aca7f78309f4db113479d5509dc
748937d8fb1b5371eeb9d1627a938f10737a802d
75050 F20101124_AABDUT rogers_c_Page_118.jpg
9e6d0b622f31c8dcb0255cfff025708f
0f483600333b6081f33582b1ec424b107ba53961
F20101124_AABEBC rogers_c_Page_039.tif
6d0e90d33ccdf41cfa39dc351b1e98bd
69481b2f40cdb806ddfae6933d40353bbf26a92a
F20101124_AABEAO rogers_c_Page_020.tif
789a9016f97e39f9b4e0e879fc8e96b1
b1838bf17ef47ce2aeb86984546f08127352725a
1051967 F20101124_AABDVI rogers_c_Page_006.jp2
d7d4c4cfc893759128347be461b59fb2
cfdc4cd30919f5e4aaf62a03fc3313048c2a4593
82202 F20101124_AABDUU rogers_c_Page_119.jpg
33ff0d14f4ac70b0e021dea13ca362f1
c8ad49c059fdb8b0b7f3dfa9028ecf6dc181ddba
F20101124_AABEBD rogers_c_Page_042.tif
6f8f5d22234d2c53d261233d0e85977a
4658d60b4aa64f09e56b5791eb54cef6f360a99e
F20101124_AABEAP rogers_c_Page_021.tif
a8babc755c6af97398a157594a807876
33a9ed72f2204712c2909be49d695836361e06d4
1051982 F20101124_AABDVJ rogers_c_Page_007.jp2
6d2435087c721eb71c8e5db38c40617f
8d1f1fad230bbedbe8095e1d65f2570e88219b54
15544 F20101124_AABDUV rogers_c_Page_120.jpg
83aca20e140ec043167e69b7b812af6a
1fe43bbb6eb4a38d24966c466e36cd724439c728
F20101124_AABEAQ rogers_c_Page_023.tif
146a0f24cd52393ddfa0445069f358b5
ad295c67a73b86088844785eeca72f25f195bd11
562266 F20101124_AABDVK rogers_c_Page_008.jp2
b663549d5af103c5f5851113f47bee4b
27b5a6896b364fee329615eb075065309018c585
73765 F20101124_AABDUW rogers_c_Page_121.jpg
989a574b8f16fd406422595d862b7e92
dac80b44adfa6c5db8539a7d686dfacb04ff033e
F20101124_AABEBE rogers_c_Page_043.tif
e920b18f4a18cd531323562c5a612230
29eea7056fa2c35ce7dfdbda2b67211a2e568b62
F20101124_AABEAR rogers_c_Page_024.tif
9226efdec864d39da0f319b3151ed960
b926f2c5a6694c1e469f5afcb7bb5d8e7d52d99f
675479 F20101124_AABDVL rogers_c_Page_009.jp2
af7436ac40ca9daa2fbab52c16bdc147
edfd740ad536b2e3693d84746c4fc80e0b228353
67817 F20101124_AABDUX rogers_c_Page_122.jpg
090be1fadf6bdc911e49e09b6a26e7c5
d789c291fb557bfcd80704fe8ddd79d9b5f38b23
F20101124_AABEBF rogers_c_Page_044.tif
5709c9649bbb764496e0e468d4e994fa
4e9a61a3d9d14036e9a17ea44d210b3eb59b0f9a
1051886 F20101124_AABDWA rogers_c_Page_024.jp2
85c510a9dfb9c9531f5bb8e197d65f43
0d37889da6ec124f5ec58cc5823ced450cc2a50d
F20101124_AABEAS rogers_c_Page_025.tif
8b8e915ba5ca5ee6193f47854a161875
606cb24b11166eeabb4d4b7294611c461bc59448
1051984 F20101124_AABDVM rogers_c_Page_010.jp2
156a96452438086ee30ee1fba36918a2
059898e5b1740a3f49851d225535ab869cf6d5b5
F20101124_AABEBG rogers_c_Page_045.tif
6e26bde0376bccd68f80e9dfcd4eb73c
c0f870f5c58a25cc0eaef20cf29b1a8ef920abcd
1051951 F20101124_AABDWB rogers_c_Page_025.jp2
eeaf2c847fcacfebe55e8f11926cd78f
51f2f0fc911f7bad1f592bd7d5d5c07a6a2c0d45
F20101124_AABEAT rogers_c_Page_026.tif
47291230b20f1973d8cb8cfd074f93f0
23a31323205cd60abf12f8f89b06a66db9d5d467
1051979 F20101124_AABDVN rogers_c_Page_011.jp2
8a8e8a997f4d58737e3974f8ddc1ddc8
f296246150d44deab6d7cd792eca148964dd56a9
30487 F20101124_AABDUY rogers_c_Page_124.jpg
326955ec4e187d99fb44cb17b44c1330
d2000dc5a42e753f82390890a4c90644fbba3b94
F20101124_AABEBH rogers_c_Page_046.tif
a7f5856935a2a25a792d98fb3b35fe8b
dc9408792126e287345b0670b20531105eb7a808
1051956 F20101124_AABDWC rogers_c_Page_026.jp2
2e265ea9a8558bbd98ff5c0701a7faaa
b3320b4093dc0c6211605c39ca9654097502db58
F20101124_AABEAU rogers_c_Page_028.tif
7f6e5061c7151200787bacfba04942d5
9df11a13844aaeec0c89a3290bc46bb928c7b7f4
1051971 F20101124_AABDVO rogers_c_Page_012.jp2
308e59bd281b2deaf39b733a8fa42857
b157606d9ec2cd4324c437e2b4590a0fa9487f37
20177 F20101124_AABDUZ rogers_c_Page_125.jpg
0c9a1524d53657c5c25db53ea33bbb1e
0e8da25ac400a77207d1879c6cb5dd12826e4355
F20101124_AABEBI rogers_c_Page_047.tif
df99542c280007ee5724fc0f9f63335b
c47abc518c3546aece2ea04dd7e9327539e0a569
1051986 F20101124_AABDWD rogers_c_Page_027.jp2
e27df0a4d13d53f787056dda6a70952a
db917a72f20b476f146104f7f20783c90bc1dc3b
F20101124_AABEAV rogers_c_Page_030.tif
a8cd45434d6766b566fde989a1758050
68da321ecdcd33d3fef1ae0519a1a287b2c1bcb9
106268 F20101124_AABDVP rogers_c_Page_013.jp2
677767d974fefa721cbf90b8174e6b78
8ee5bc4935aac691b5c6625270580ed6a27e6c60
F20101124_AABEBJ rogers_c_Page_050.tif
cbb560eaac4b54806445eda0a13552bc
6264a8321b1b7c832a25c06221334ad4de18bbc6
1051975 F20101124_AABDWE rogers_c_Page_028.jp2
4cac0dc4c4d3a5cda8a67489e367f4f5
d347d74a1bda16ed02a260386c5c70eb9d31dbbe
F20101124_AABEAW rogers_c_Page_031.tif
36b3ddec184e1ff8b09e06f1ac85ff47
f5c206c75cf428ac3359181698ec159a76fecd35
1051977 F20101124_AABDVQ rogers_c_Page_014.jp2
a0ce519ec2ed3177921ce27e45823e72
8b1564ff6832b887d5547b4c78a5a49bf263c37e
F20101124_AABEBK rogers_c_Page_051.tif
c244ec554665afb57827eb813880514e
653f06de92aa2f19c5cfa2f97639a7aab0595327
1051976 F20101124_AABDWF rogers_c_Page_029.jp2
0744175adcea8a0613cd3d80cf648a72
4593f8f7f71f21cdd7fa3221c5459f8edb18b8f9
F20101124_AABEAX rogers_c_Page_032.tif
5b3e003da16a85929f1bad3317a2230b
7fee6c0a33231073be285a51fc70b1d24ee4a034
1051970 F20101124_AABDVR rogers_c_Page_015.jp2
ed4d73ded453228f0ae939d41934d8b8
ab77be8d6a1f64c61d32cddbb9212ecaf5c37514
F20101124_AABECA rogers_c_Page_070.tif
cc8c648ccfdb1f93148bb770d4f01fc6
ea87446a7c4e3ba912b3c341a6752be8059bda34
F20101124_AABEBL rogers_c_Page_052.tif
46a89cbba3c4a59acfc7aea901355f16
ff5fa0c3a6f95ddd9cfd02a604e371e1da49d52f
1051973 F20101124_AABDWG rogers_c_Page_030.jp2
05ef7193d4339e32644b52980c81c6ec
0896c2777947e9640a3368ee3a0acb8bca5a2f16
F20101124_AABEAY rogers_c_Page_033.tif
9effba60670786128ed8713fe9eabe64
1313efd6cf42a58ac03fd79c873d4ee4a87f58ff
702896 F20101124_AABDVS rogers_c_Page_016.jp2
9d19b69c26ae27428b5e677885c2dcaf
db5cf8a2fac1f0e189a66a0dc7e2e27176d82ce2
F20101124_AABECB rogers_c_Page_071.tif
9c24189b6c981f10fc404f2c643be433
837032e84ffe89bcfdda2d9118c01ccec1d3ccfe
F20101124_AABEBM rogers_c_Page_054.tif
32635468ca94d779dda2627f770b2f53
0b1dc9518d1a2947af1ed4c3c0d82cef0b20c4a5
F20101124_AABDWH rogers_c_Page_031.jp2
dff2184b412b4c02030d4205c847b8f4
2c687191495e65bc53e070b71709cea7582b843b
F20101124_AABEAZ rogers_c_Page_034.tif
4f2b83832f88d3ed6b223a282e13aee3
d67418a3f46a484064f55e219922c4246973527d
480498 F20101124_AABDVT rogers_c_Page_017.jp2
34499055e951c66c9a46902f76334b9f
0bd40475751f73904a74654e0054d11ec41931de
F20101124_AABECC rogers_c_Page_072.tif
d404147d7e7d301e6dab59c48d8e5944
69dd7800e192ab25e9dc2ffc2add535a5eb0d522
F20101124_AABEBN rogers_c_Page_055.tif
6c905a70f3a6099447139d3a649a1d8a
e5b89a5585d17d73234bdea6729d65e3453442ec
F20101124_AABDWI rogers_c_Page_032.jp2
7cbcd401f2cafeaeb42876fc5ea9dd6e
c269d7739aecdbdc0ce559c3ea568dcea155cf37
F20101124_AABDVU rogers_c_Page_018.jp2
cd8653b62ad1878560a497a6898e27bb
0421751e3e0dd803acb628e0a036d0acf55fc49f
F20101124_AABECD rogers_c_Page_073.tif
0d6481662c05670c78c85e2ac07fd45a
379337d40e94b1a28982320aee18e28ccb8b74f7
F20101124_AABEBO rogers_c_Page_056.tif
dbeee67a91ee5feefaa7241ecae91e12
ebd63d081b25aafe09b58ad66234756666cafd83
900655 F20101124_AABDWJ rogers_c_Page_033.jp2
20dd0dabcbff14d3d4d2b60e383c85a2
09470bff390a7c3ac26bd269f0b836278b10f5f1
F20101124_AABDVV rogers_c_Page_019.jp2
183278910912c81d5267482fe002b60e
0b3108d24831cc5f9213fe70ecbe18836b89b814
F20101124_AABECE rogers_c_Page_074.tif
a8cdc26ca4e54780cd868a8529357f11
820ff7c0b949091e1283306e53ca9f87f7b3aa3b
F20101124_AABEBP rogers_c_Page_057.tif
2e6e5430f6710380f65b24a6c26360d0
28248aa74377dfc61f982781095c0ae1aa640ccc
1051947 F20101124_AABDWK rogers_c_Page_034.jp2
9439a7d2274377cfdc833be10d1d0f43
bebe1be7c1c44c335f1892dbc9076bd2c1bce71f
1051959 F20101124_AABDVW rogers_c_Page_020.jp2
e78b5b1df6a9165b1edb0b61351e54bb
2bcd58a8ce226b2959115916dee1738f94f15572
F20101124_AABEBQ rogers_c_Page_058.tif
a07d2a15aa6e65156f199208af9bcc02
d1e1cab989377c31e5a5d08bc3878dab5bfdce4a
493863 F20101124_AABDWL rogers_c_Page_035.jp2
60bd423aa87a55626f8990ccc5d21515
b43047c4bb33232ec1e08b83fd8da1bf2260167b
F20101124_AABDVX rogers_c_Page_021.jp2
b6051ffabce8891203c28f9efd5289c6
1a6111c78fddbcce3ce46a02ee0eda97a1168153
F20101124_AABECF rogers_c_Page_075.tif
83fabe8f8a96cd61520e1a804ce9428e
bfa4ab2397710af340bb6cd470e95d5c69850f6b
F20101124_AABEBR rogers_c_Page_059.tif
4aeb1cf6c1062d3fad520e73082fe93b
388af30569353bf6985450acdec498a5e755e23c
946806 F20101124_AABDWM rogers_c_Page_037.jp2
a38edc404fdd50c50b1cf2e7bb0a56b0
1cbc32cfd5fe4db7636718e5a56a62c4f814e85d
1051968 F20101124_AABDVY rogers_c_Page_022.jp2
351e6d6b7cde8c695d1ed05172b05176
3bb991007c363bf0ad7f265ee471d8220b02cada
F20101124_AABECG rogers_c_Page_076.tif
52f609196ad96af10813b41ed2d60fc3
fef87d47806dfd70dd3419386b473ccba5dc7757
F20101124_AABDXA rogers_c_Page_052.jp2
ac48a1a2004a39b0de3cb72aef1e0f39
479e4515520cb36dd14d7fa9f6aa239c74b7807e
F20101124_AABEBS rogers_c_Page_061.tif
f23ff2a70d61b8d2f5373e44064e1af5
ad253b9c3a70e52423f9cc25456f2c8a9616a278
764574 F20101124_AABDWN rogers_c_Page_039.jp2
c6ec8b47bc0d352c940347cfba7d5157
026cc8dba9f534386ec78a2bbc8ffc881073ed0d
F20101124_AABECH rogers_c_Page_077.tif
14ac66261e1fec22c266ac938e713eee
5aa2717a67f0f3d7aeeffba6eaf6c3e55accb9e0
967902 F20101124_AABDXB rogers_c_Page_053.jp2
982e6e5b2bec42c7ea02aeea1bc20fc5
5804f58095e77fd2b7ea7b0c14f28acc19bc7b6a
F20101124_AABEBT rogers_c_Page_062.tif
51e65a234e0342833ba4957fa0279b4a
64ef03ea2f86897e58317a9c45835c3a03e99b1a
1051954 F20101124_AABDWO rogers_c_Page_040.jp2
a740a31009e3fd2a1248eea1de6a7335
c3a57ea930753959a189bcf3e4498dc591761e59
1051966 F20101124_AABDVZ rogers_c_Page_023.jp2
e8150cce8036fa53faeb73c3c85bf3da
ee6c37a9fb4480ba0b750b6233993ebb751d653a
F20101124_AABECI rogers_c_Page_079.tif
20afe84af78e5a3f4ff7ff033e2afbb8
83d8946e4b40c21857e7682fc2f8819f7030d4d6
997361 F20101124_AABDXC rogers_c_Page_055.jp2
58b898040aec7cb2374d18a89ea08800
6eff75caa86eeaaf5ed48e47e8c65061969ed85a
F20101124_AABEBU rogers_c_Page_063.tif
0712b66bd48d1b478c385b582a6d52d2
d92a6d63644f9d712602596fbbe506c07640fa0c
F20101124_AABDWP rogers_c_Page_041.jp2
50d737dbdc827543da4fe1b5e9f700f1
bf542cf2a910b3ffd964e005cacbaf96e548afd9
F20101124_AABECJ rogers_c_Page_080.tif
afb8bcb112ca7e7946aec20341cb5b04
1601d11f82aca40ad66796c5ba5eac4c5903b571
1011385 F20101124_AABDXD rogers_c_Page_056.jp2
acdd03e2cf325ff5d68e82e48e3465bc
5fa4dc227004e6b11a076de3ccaf3144ac1353c3
F20101124_AABEBV rogers_c_Page_065.tif
8ece424cd3c35f1007cb15b4a08250d2
06305cfd7d86fd7e513896ab7fe5898146975b82
963234 F20101124_AABDWQ rogers_c_Page_042.jp2
bb6d1cf4b296f1de87ac3cc63747f3c0
fc57e822d1085812ee6a77928f53410a229e02d0
F20101124_AABECK rogers_c_Page_081.tif
8fcd840fcfd7820bdf1b5e8ec8a4c905
1e7158cdb6c7dfe950353d4016573a1203c58811
435557 F20101124_AABDXE rogers_c_Page_057.jp2
6da927790f16e5a30e499d56fe3afd91
06d645a01ac20d137d29f2632385d5b9bacb38d7
F20101124_AABEBW rogers_c_Page_066.tif
50c49201430cbab473d8f68454ffd831
e31f5a40f550ddad274bc03710b0bbe17fec6d50
1051980 F20101124_AABDWR rogers_c_Page_043.jp2
05fc0a0c7fce8a393ae88c74dd8f5ac2
c6dfb4ce1a224a45cac55d2f0395f74ca9def1bb
F20101124_AABEDA rogers_c_Page_099.tif
f8e556fdd769a3b0b080d44ca5f8dd23
73c5cbb00ab31a75aff389e222046fc303de74d8
F20101124_AABECL rogers_c_Page_082.tif
7c85fb2c99b4d0f576bba035bc597f10
c362d9fcda8c02400dabd8723b6c3dfb7cfa9d25
1051985 F20101124_AABDXF rogers_c_Page_058.jp2
8e417e6c069852b61257008fe14fd042
e1a50574361eaa4578995614a0a970a9ddba253b
F20101124_AABEBX rogers_c_Page_067.tif
d26db2fe32106c21d080e6ba9b455ed5
ccce014b4a851716d8aabcf1b93fcb74698742f5
857348 F20101124_AABDWS rogers_c_Page_044.jp2
47175bfc60b22bcf44cf2ce9818a5c1e
2165fbe2340e43ef31ae7a24e10cc10582b351b6
8423998 F20101124_AABEDB rogers_c_Page_100.tif
8ad9315e2bab28ad819ee9179d9dad7e
31c0a3bf829bbe53248c959f19cbccf9f794949d
F20101124_AABECM rogers_c_Page_083.tif
e62752467999bdce965bdbf0f7f317ea
d2e0dbc3088b432a85d7046709768c0e24815989
F20101124_AABDXG rogers_c_Page_059.jp2
212193343ede815a2661034c92090e5f
f69653536816cdb71ed3aa35468a9d6457fdbf23
F20101124_AABEBY rogers_c_Page_068.tif
bc9be69ade21b19aa2f5a3c7a4119c19
8b7d3946a520f6b9b855c2668541a850f0e14172
633312 F20101124_AABDWT rogers_c_Page_045.jp2
973fa1238cbfbc4e1385e1078dc4f573
c490c25181b9553b6526c2096e92041638ac5f34
F20101124_AABEDC rogers_c_Page_101.tif
63009663f1819978b7fd52e26facb9fe
20be52bc102efc51816c0112ff701062a20be6a6
F20101124_AABECN rogers_c_Page_084.tif
5b4347f38390c6406242779b81415dc0
53c6f296a815f9c2c3e38fc1e84aa04880d552d0
676484 F20101124_AABDXH rogers_c_Page_060.jp2
71a6a041b87c1fa004458dcc0affa398
86fb095484465b1e6f9a0d04082f4a6b52b380f8
F20101124_AABEBZ rogers_c_Page_069.tif
af4b04e674f5114d73ebe28c97e17872
a31432f73a656e7a3bbdf3fe1d68c929a31956db
902838 F20101124_AABDWU rogers_c_Page_046.jp2
abeae1491f5dcef16a0c53911c38532a
0c198ad597ff4cae98711e801d077608c84c0b94
F20101124_AABEDD rogers_c_Page_102.tif
10d3db746acf1f7ac3ca1de967a75fc3
6ebd8b549f8605911ee5f684fb174c05505b8bc2
F20101124_AABECO rogers_c_Page_085.tif
951444aa5e44127a90daa6498694d69b
d046b73951d6b0ce089b98f320bd208d86d54e87
1051957 F20101124_AABDXI rogers_c_Page_061.jp2
7839c1aab3323c3391fb5329421acfed
61cc18152349ee76a1e80baef0504f94a3596311
F20101124_AABDWV rogers_c_Page_047.jp2
3dd9281772cef77952c1ca9443b4f78b
374d2fc41a78a2da4294cb163a43a030212d757b
F20101124_AABEDE rogers_c_Page_103.tif
3e8bd90250c2772be486e8615fe4078b
cd90e039fcacad98b540284c61f651f2b6cdbda3
F20101124_AABECP rogers_c_Page_086.tif
580cbde2adfeb5485c8fa50ecbfacdab
d946c53af3ae2d6cca2fdc8d96bea9843c5c099b
939738 F20101124_AABDXJ rogers_c_Page_062.jp2
4e303f4e2634be3fc7e6a1b90688ed8b
551b5991aaacb99680a0e9e68591c953be132ef8
517616 F20101124_AABDWW rogers_c_Page_048.jp2
dd00f53fd9d45a97842fb9a4b04f4cd1
af64c9f2f88e9b66045691b41448a4095a58cd5d
F20101124_AABEDF rogers_c_Page_104.tif
6ef82b49a9993845eead5e96063f4dc5
38e17aa30f0722851e75db3a407d10b2a614e784
F20101124_AABECQ rogers_c_Page_087.tif
c66b61e027779399bfa016c9c2958f46
084cadc4f9e8c834016abd8d5d14ff9d5ae0b6e8
369705 F20101124_AABDXK rogers_c_Page_063.jp2
aa70708a6ce1dd0c5d944f629683e3d3
9c46fcad4c003d1c30789419318247e844f50ebc
1051972 F20101124_AABDWX rogers_c_Page_049.jp2
672ad1b9fb77929e0ff999487973e677
4fb1c072fe6060b004f2f25ccc32994c22e35b35
F20101124_AABECR rogers_c_Page_088.tif
b846cf1f6fe8b7839d6544e6de9a3226
576a0d822327512f8a67e346c340c8ee03d448b4
692778 F20101124_AABDXL rogers_c_Page_065.jp2
644afc6de0e87ced0b8c3288cbf3fa32
30d7733a7aa6a09d7688b28001b57659ff1ab373
F20101124_AABDWY rogers_c_Page_050.jp2
2855a72424a7dec3447cc88c195909cb
6971f7d5554f911909f7fd0d9341a8c62ba9325d
F20101124_AABEDG rogers_c_Page_105.tif
81cd2a541d46f1934089184a3c21d9c7
f747f4052882ffd079a03d6238aa9b63a36ce486
49870 F20101124_AABDYA rogers_c_Page_081.jp2
321e9ea258255a9480bf5bca95d4cb2b
a3911dbf757f7cef10f4ee29ec9b55876c1978a1
F20101124_AABECS rogers_c_Page_089.tif
09a0f62ea508d2435e6777940be5bcaf
7dbb10004e16805876dc8f64dc3dc7e98c06ab2a
861175 F20101124_AABDXM rogers_c_Page_066.jp2
20fa2dbd638ae586e01192c0d060018f
80bf2bc24f0dbcfaca61dfb842fc89da57ccbeba
1051981 F20101124_AABDWZ rogers_c_Page_051.jp2
c50c449ebdd064d8f18fe9a9f405bc70
f03d3ea805e74f94f85c997b27e9089c1c14473b
F20101124_AABEDH rogers_c_Page_106.tif
1cd1fd0f02d896bbf4792ffcc6716170
7ba9541bc112be34a26838156aaf8df1e0389b76
778061 F20101124_AABDYB rogers_c_Page_082.jp2
3fefd8d775cbe3a1b5b95141e2d95fc3
59a3e87d3df751448662b16cbec583c97cb87db3
F20101124_AABECT rogers_c_Page_090.tif
76524ea04c6811989b7a1d1060ca50b8
d5ef5101b42034bc178f2016a77c02227e1fc453
F20101124_AABDXN rogers_c_Page_067.jp2
7ab617ceadb3c2955ab21233978689b7
99c75f640cd076335db93b4b580153f39040f098
F20101124_AABEDI rogers_c_Page_107.tif
ba4d77491ba4e8847dda17d3c3b48361
b4de519cbc5163073c3e3916412077744bc2e721
632362 F20101124_AABDYC rogers_c_Page_084.jp2
70ada2bc80dec7db5f1a82b18e76dbc4
62e9635f22f3b4083191c937463060a6bf3d08f7
F20101124_AABECU rogers_c_Page_091.tif
a4c413141881488209fab6ecae797c2a
995881da1dd00e06356459c3a9c4f0afdbe99992
1051930 F20101124_AABDXO rogers_c_Page_068.jp2
ff5033ea4305207564716e4ff934bce3
7bcc22699d6dd0345d743d108c31fc52bcb10c24
F20101124_AABEDJ rogers_c_Page_108.tif
f4ade4449fd52b7a3c025b1cf55385b5
12f485ae39fd470e32df29c782cd16cbae0b5549
F20101124_AABDYD rogers_c_Page_085.jp2
7efd1b69163931f0b94f6cd7cbb43f0c
615e32ae16fd97e10ee70b52ad9c70fddff890bb
F20101124_AABECV rogers_c_Page_093.tif
c1fe1b0cf8c1c405ea9fa74728d95184
481ad2e4525d4b346afeb7874cbe184a0e496022
9295 F20101124_AABDXP rogers_c_Page_070.jp2
cf0b9e37c2afb14a321c7a280b398a52
a7294e4f9390215e7a02d3154ad0df3ef105e65c
F20101124_AABEDK rogers_c_Page_109.tif
5b6db074f0a8226a39aa3533b797b728
b358bf11f619034d5add8df152cae489288ab2d3
F20101124_AABDYE rogers_c_Page_086.jp2
d8e5ca6385f8ce6d11826e2e3b421adb
dfde4d7de931bb0b8492dc5ea424a9ece1fc64ef
F20101124_AABECW rogers_c_Page_095.tif
46774fc2ff087806353b38f0eb9a2cff
be668ce24f815292e5d18f36f7ca0410e009e442
1051983 F20101124_AABDXQ rogers_c_Page_071.jp2
08b737238857568d17459b8bea8eba34
c6a689de253b094a8936a57266b2ee92ab2c37e5
F20101124_AABEDL rogers_c_Page_110.tif
2de2992c5bda1f09ee8e22343b7ab443
af5ebd18ecab67056dff8347940ef67fa18eb52b
F20101124_AABDYF rogers_c_Page_087.jp2
7020fe89d9df0b9494c4d41725e1d65c
14b69c349fe5d7c0c89b85553f579d0cc01da232
F20101124_AABECX rogers_c_Page_096.tif
ca7416e0289c0ebfe5e02b388a546324
208157197a221dd0785e343648ad89502c23380c
918552 F20101124_AABDXR rogers_c_Page_072.jp2
19b219db87b4e49afab7452022d413d4
8e7899a134fc69daaa67eb042aa5b1d16797166f
F20101124_AABEEA rogers_c_Page_127.tif
d3b8f1479c162bcc52d33b75a6ce9f67
8ff787669cf10e4dc98d4aa968efc549f400a859
F20101124_AABEDM rogers_c_Page_111.tif
d87b6ffdc3299e0511c70ee4ce925575
7c30b268075f8773df0c619f0fe03b325091c478
1051934 F20101124_AABDYG rogers_c_Page_088.jp2
e7ffe7e1bfacadbee8caf38d0bd764cb
94184002508193348969a7b702dcdb378e7b3d67
F20101124_AABECY rogers_c_Page_097.tif
53ac0571f05cd14cd0187d17fd800eab
f2975354c0c98ea854527490a60868c980cf2af9
F20101124_AABDXS rogers_c_Page_073.jp2
8c2c80c31f6cda8eb08a65faee52087e
3ed51725e48aec1c20e185a52db6f0db0c788bbf
F20101124_AABEEB rogers_c_Page_128.tif
667adef32756b33106a230c61aa1876b
795e18eb52a8e89676471ff2b3202fe933bc66dc
F20101124_AABEDN rogers_c_Page_112.tif
7d72df22d3f61d0568199a60a677f483
b7b7540560439f1df4ffcdaa67f2cec19b8af5f7
F20101124_AABDYH rogers_c_Page_089.jp2
20edbb1e6ef705eec38f5038b3ff22bb
ad5b7f4e44fd76741e2e447691d4b05bae347723
F20101124_AABECZ rogers_c_Page_098.tif
bb4a598f73b66a91cedd4cf7e9c8c17c
f7a297ca30e2a7c8e976d0d8a24dc791af0eedbf
343460 F20101124_AABDXT rogers_c_Page_074.jp2
91b2794fb49bc8a71b9f802c85517386
e6d393f2fbf8571b1afa7fc489bed8bded965247
F20101124_AABEEC rogers_c_Page_129.tif
4f8dd69eb77dc50b6e637fbe2b296a8b
27ecc8b9783f9b49352b018873ab326a2d6f470d
F20101124_AABEDO rogers_c_Page_113.tif
742c8adbf2fc98157bbae5dc623a4e64
db106b4899e2cd561c5e98027cd2895e7b0dce09
463310 F20101124_AABDYI rogers_c_Page_090.jp2
57933397055164daba13e800d5e4468b
ef057a68fca3a4763be338443ab1402751a6d3d6
F20101124_AABDXU rogers_c_Page_075.jp2
38785e00598d4c6b5e6b84484b8211f3
bf7fda20a39433638a929c4b5687c068f31e4ebb
F20101124_AABEED rogers_c_Page_130.tif
47fd2460d23a0154a8f9795bca09e37a
ccd6e9c679f70d51c88c55d496244b7433c4f67a
F20101124_AABEDP rogers_c_Page_114.tif
88cb87be6ec42680185dfed6bbcb5666
06aa84b59caed1b8eea9684a213836011ce6a4bc
1051931 F20101124_AABDYJ rogers_c_Page_091.jp2
64aec6bcd8baf10ec3305262ada0e571
1963c401fa8e958e678e36b9d3ae82e7b57026e5
F20101124_AABDXV rogers_c_Page_076.jp2
3c0fd3ee0d527ef37b2b14096001eeee
d45f01584baa7e63127fc229e286644210ee47af
F20101124_AABEEE rogers_c_Page_131.tif
c3b58b56c137281fa6388fad76c35430
249552eafd03c5b57b453dbda32558376bc170fb
F20101124_AABEDQ rogers_c_Page_116.tif
1c4e4f9317572b72ac13fea8a968a2ce
97c8042a035df68c22d9eeff300613f454635079
965934 F20101124_AABDYK rogers_c_Page_092.jp2
bce7fd79f9e1661d22394d935938923a
4f35a1aa682ae62018ccf536b87418b2dc2dea60
62448 F20101124_AABDXW rogers_c_Page_077.jp2
f7246a1a0cfaf8943c8a736d21a79c52
2c347a90e7e791c78a5bb6a8a8d0adb77499e18a
9226 F20101124_AABEEF rogers_c_Page_001.pro
8288f2ec2ce2ce514e9e1e71ced3fd75
c3abcd55fccf3680631513b018ff89edfb2e22f3
F20101124_AABEDR rogers_c_Page_117.tif
01d63693c074de8f69e7b438cc03b06c
a53dfebece536d93c8a69857afb167e917f8562d
37915 F20101124_AABDYL rogers_c_Page_093.jp2
6e09720bdba2ed3b403f6d2901b7d5c6
8edfe670f61c4fad3a5b1642a10bea68809dcc1a
F20101124_AABDXX rogers_c_Page_078.jp2
b5b1149c269c2c132a8d01b1adaf228d
c6102050dffe27955cde1cf560eb63903969d68a
986 F20101124_AABEEG rogers_c_Page_002.pro
b907e8261250fcd466d66d10d375b1e2
1ac5d216f9718937eab491155a2c72efe37d045b
F20101124_AABDZA rogers_c_Page_109.jp2
e5763f751211967de4ee72741956b522
16e7f6d7ec12d73d1c9dadb6551169a0e4835ae0
F20101124_AABEDS rogers_c_Page_118.tif
84a5062506e90797a53f749aec813aa3
d79c142fd74c39b6a29b391cd93cc45ce693c7af
1051943 F20101124_AABDYM rogers_c_Page_094.jp2
37778d17082bcd5ed394447681688989
b64e6f139598fdb89d4d589b56fffbd0b92ed2fa
751450 F20101124_AABDXY rogers_c_Page_079.jp2
9c653b75a1f568ace14af43d2686f6b9
908dec8de6e9e87b899dc41134085e867c137ac3
F20101124_AABEDT rogers_c_Page_120.tif
79a54368f0aa4f552c5476cf7040c174
7edfdf7724d9fff8bcf2ebf740c8f5d39c3ba233
121509 F20101124_AABDYN rogers_c_Page_095.jp2
b851f5687b4f9b1fbf19d067a3f3f57e
2a61868e8ece2fdc5a2f3609e94a466ec30e088f
900333 F20101124_AABDXZ rogers_c_Page_080.jp2
48582738c3434ed691f5c72e1abb4550
b9be14987822b02c5015b1dd9082d4a41025ad63
14596 F20101124_AABEEH rogers_c_Page_003.pro
e364f4fd66c49a06523968570828b166
a2106e74e44f8396c514d9467fbc9f4c83387e10
920870 F20101124_AABDZB rogers_c_Page_110.jp2
9adf7a19fd97d2c04b37048b13873711
911cf0149a5c786a51a05f079d78c242f9a30fff
F20101124_AABEDU rogers_c_Page_121.tif
70a4f223f8466432c2c83c241bf446f0
479a5f71a3e4bb648abbefd55bfe897c6ce5cd27
872057 F20101124_AABDYO rogers_c_Page_097.jp2
22f13b5558c6084c3edd0ab91331f32b
56f72fb46bac78c34d4011fa11b5fcedd55bfb73
57496 F20101124_AABEEI rogers_c_Page_004.pro
37dbb9e6ff6f03b94a451f1762784be7
d5e0c9e9c7e1730f7ccddb70b8318a9dd8921044
965472 F20101124_AABDZC rogers_c_Page_111.jp2
ad045f57f955e5359e831069df96550b
ffd3766a0a2718ddba72e671a123a2a74f7a30a8
F20101124_AABEDV rogers_c_Page_122.tif
aa565b52130550dcbac94d633291ad00
025083c682478f6b63b53e76d02d481d68b8d562
1051938 F20101124_AABDYP rogers_c_Page_098.jp2
57f21bcc052d6e7e432c36998e8c3550
bbfd16a68305cd9fecb228c87a1ee5413e9b10ae
19825 F20101124_AABEEJ rogers_c_Page_005.pro
e85237a40777302f4c004c4acff5d336
ca151fe835c228519589f8606c4a9b4af60e1a32
1051942 F20101124_AABDZD rogers_c_Page_112.jp2
3be52c9f607f831f76ca7e842b6a4c02
613c82e30e05d002ceac15ad9b0801acbd55c93a
F20101124_AABEDW rogers_c_Page_123.tif
17653016c880418f26be4a3e723d80c7
5ead63492eed1eaf74cb49a01d9fe978ef46f943
F20101124_AABDYQ rogers_c_Page_099.jp2
03bca6dab8653fce949027f4b32d6ee2
aeffc5c0ce9f485d5b50a3e447877d4ddd7f596e
31468 F20101124_AABEEK rogers_c_Page_006.pro
3b20634a77ed4a1d9e285a78a22215a1
31a88664848d09c4bbbdcaef66c473750ddab66c
1037219 F20101124_AABDZE rogers_c_Page_113.jp2
e0426c31a493f6eff0af7f7d5ce74c2d
d76bf87b4382072d19e0615faecaa9d9888cf535
F20101124_AABEDX rogers_c_Page_124.tif
304a815ec1e6353c91c7fd83e7681a02
aac1109170855fc7ae36390e53b8d63ccb209e17
544120 F20101124_AABDYR rogers_c_Page_100.jp2
613469c3ced0496b414a85bca6bce773
0a119ddad488205bace47c41a589fb7dad1d3e47
55990 F20101124_AABEFA rogers_c_Page_024.pro
b2b4038208eac4f651f65d60e5a0b1d6
6ef199691463d73d9fc642b963b6033411c037fa
36392 F20101124_AABEEL rogers_c_Page_007.pro
ece8df908d76a78d2949aacb2d34f8a8
dd6001d20d833e95bdb6c9e5e5dadf48631246b8
1050528 F20101124_AABDZF rogers_c_Page_114.jp2
cc80a70ac2d8f77836f9fd45689259cf
9fba44c6207c9c8652c07a97ba70db89092e4074
F20101124_AABEDY rogers_c_Page_125.tif
5a95fbde4908944d894d76f72d5165dc
cfb97cbdcf927f405019fcde0eaa82b5cfafda4c
1051939 F20101124_AABDYS rogers_c_Page_101.jp2
534159c5343803ff673585a56cc24da2
7b3c74c035f3cb3d89d4649b6115bef737e6cd45
58906 F20101124_AABEFB rogers_c_Page_025.pro
bdfc4bdf222b3f2b7b757f72e5226a87
7b8fd0a434fecd11f25a62fe94751cc4f8ad6b6c
12518 F20101124_AABEEM rogers_c_Page_008.pro
740636ae5726037f1c7b961f5a924c9b
5abc0d379c675eef6953b530cb1d3829803934ca
980405 F20101124_AABDZG rogers_c_Page_115.jp2
64fb52072d4a51b6b5b735531ecb5e79
4370e906709fc1ac38365a36fdb63c2e47969715
F20101124_AABEDZ rogers_c_Page_126.tif
6ecdd13184f80dfe75ca934a3f3974a6
9fdaa3f899f7a7d5ca7de12345aac3d90a35185d
817538 F20101124_AABDYT rogers_c_Page_102.jp2
730d8e74588be215a30b3d0985888e5b
6c9ccd7752c37875f3432c5c060939243637fdeb
60226 F20101124_AABEFC rogers_c_Page_026.pro
c6bf72f5c95786ad28d7a8e834b49ac0
c82cc5255c2dee00fc6fdcb2cf1a0a5335ed8962
17112 F20101124_AABEEN rogers_c_Page_009.pro
179480b74bde5709f2663d9070f6a3c8
ce8b58e53d14424d410eb0b2b6282d09fef0ac5c
1051969 F20101124_AABDZH rogers_c_Page_116.jp2
b16934cc4eeb1ecf413d5add3c07b5d3
9fe276e4596d0d6e7f4169dc145d07b93a6c7ee0
365343 F20101124_AABDYU rogers_c_Page_103.jp2
e2fa84009b2d72aa4d658be6c56f2157
3fe682372465a205d48beee3797cc79677f570ee
59350 F20101124_AABEFD rogers_c_Page_028.pro
97567717312b8a0146d7aec373d9ab39
cca09c051df904f65e8eeed1c62eedcd181fc80b
58523 F20101124_AABEEO rogers_c_Page_010.pro
dcd8b34b255751156898003b396bbdbe
8295aee3c8483cc833bec2a32f60972cbaa71003
708165 F20101124_AABDZI rogers_c_Page_117.jp2
80569ea02c1b741275023eb1cca1677d
17bd5db6cdb50897d46b358dbb61e98b25304024
F20101124_AABDYV rogers_c_Page_104.jp2
c8b11793fc31bba97cd81059d4f7de9c
b12400d5bc829d48817f84e6e0cebfc82294c611
55504 F20101124_AABEFE rogers_c_Page_029.pro
66d44d1ecac370aaff87957dafeb03bc
0a040cc1f12ca6b6014ba94958b09330bb9ee270
41770 F20101124_AABEEP rogers_c_Page_012.pro
64987e8c0c10e04212877401590726f0
2088a8c383bfdd732ba6f8aaa1ed30e59a773a49
119813 F20101124_AABDZJ rogers_c_Page_118.jp2
5bc8fa79db4b9397f2e457a957c3c900
106111413ec4a4b9a9133892bd62ff5266eb4de2
834971 F20101124_AABDYW rogers_c_Page_105.jp2
cccd9533c29d3ce2b475f68f21fccaec
7a926e06d2a0bc82ef7a47772de56f519d91721a
58459 F20101124_AABEFF rogers_c_Page_030.pro
b65e9314aa1b11958843f011a1b54ee3
8f074c1e0ffb25ffd0b591ff6164bd8f8f4590ed
49774 F20101124_AABEEQ rogers_c_Page_013.pro
f32e644a3feac89ead75395deaa9eecc
085f093aeb578a08512682b5903da860d6166155
1051914 F20101124_AABDZK rogers_c_Page_119.jp2
a45199c0c7c08e96356b24c6689ccd4c
a5f4c9113519613f88883cbf8ae1205736c56211
992499 F20101124_AABDYX rogers_c_Page_106.jp2
781fdeb67f8957bb163560d4d555f693
168db1e653cc1a77d33c6e6cbfe98f26ca034a81
57490 F20101124_AABEFG rogers_c_Page_031.pro
ea1006c288560a390334b393c3ddd7ff
d049b66213d36ce3b01c9fdd4cb0267ed54423d5
56696 F20101124_AABEER rogers_c_Page_014.pro
0059e9fde81a8d4ae4b34b9a15559a20
9b8d5fe9366dfc84b0c3d9258bd56aca6e1952f7
18129 F20101124_AABDZL rogers_c_Page_120.jp2
2ea8e8e0a89802b4ed7f045c940af84c
eea06a63e6d93e3cb7d92fd386968365e65c03d3
953777 F20101124_AABDYY rogers_c_Page_107.jp2
4d897d0d447979d1ec858c712e2ce82a
6ef3b0068d8b6bb342b9a7ad9594d3dc2d614776
56341 F20101124_AABEFH rogers_c_Page_032.pro
94482f596ea07e6dcb02d73991881208
94aab4cc0ceecaf9ce498e4504e6bb044fc1e3a7
58312 F20101124_AABEES rogers_c_Page_015.pro
5171d8b62ac3035346aae8b5320dc4a5
f24fcaf3484702f48411093319f257444a314374
992196 F20101124_AABDZM rogers_c_Page_121.jp2
0cff721ee7604a46eaf55eea91eeb81e
31f6da68b8b6e11db3257eb3e005c2739c6de34d
1051955 F20101124_AABDYZ rogers_c_Page_108.jp2
d32a455a5dfc712d9dbbcc0dacc08cac
7b7ab188e94fb58debcb7f42af4ca2cbd8f35742
8776 F20101124_AABEET rogers_c_Page_016.pro
d2996a16511327e2dc9f69f3be7dee50
cd4e0d930a11ef1ab355dfb56779908833970573
924074 F20101124_AABDZN rogers_c_Page_122.jp2
89b3f9a90a584cceccc6f6819c6d2a54
4e01ddb289889c4fe7dca1af49e4f8b7fead10cb
10595 F20101124_AABEFI rogers_c_Page_033.pro
088b8e8d93c2b045d24e46d2bd587608
96548457f89c9f07f7402d4ec950ee5a92996844
7277 F20101124_AABEEU rogers_c_Page_017.pro
8aac9d86de1cf7b08c82176337465e95
89a84d71a379067b493c0dc42a6b42b2b4edaa74
230401 F20101124_AABDZO rogers_c_Page_123.jp2
c7a7b9cf7b9567f08d4cfd7e2a75886c
b8f0d9ab5dda52efec66b89dec7edb2b4867a90d
54507 F20101124_AABEFJ rogers_c_Page_034.pro
312deff5606ef9574fc3fe05ce1d1b9b
967dfeff02a26f0bd6d308e0a956a1175eb5de14
61314 F20101124_AABEEV rogers_c_Page_018.pro
09f6989970845a7a98007cf39ca9d891
5c952522e29955a4166179c0e654c9fcfbe0f473
41762 F20101124_AABDZP rogers_c_Page_124.jp2
6eaa09b36f718f4e77407add5f8ab92a
8aaa30d2ef44c6347faeb16486921b38930d727e
6324 F20101124_AABEFK rogers_c_Page_035.pro
75564563fc431b168a8dbf5523c4c913
ea033dde271b482bd4946c51a4adbadb83f7de04
30718 F20101124_AABEEW rogers_c_Page_019.pro
e5b520af516c4d2def2607b8ae07ffb6
5b8104f7bcdbc25fd656ef2cf437903a93ca681d
25168 F20101124_AABDZQ rogers_c_Page_125.jp2
af855da6c8b6747c3d9d074c21698a2a
60eb507ee5cb14b70391df286b2fedb130c64b9a
35040 F20101124_AABEFL rogers_c_Page_036.pro
77a19dc64a3dd2438fa07ba4702a4cc5
3e29b989e3cc16f6318d7313511686170b4ad702
60192 F20101124_AABEEX rogers_c_Page_020.pro
b64b89e8554a025023bd1875139bc5de
c90c4b89fbd72d8425bab92f6a8ea6a1c64ff578
134791 F20101124_AABDZR rogers_c_Page_126.jp2
b53a729aa8814d6984cbd5612a4faadc
be47eebd9e924a1b44d516a90268a56490a19d62
41259 F20101124_AABEGA rogers_c_Page_051.pro
fb6c3f198abe447a8ca6a3afa7b185ab
7d9a82877afb3f4289781f5abd4b30895377f9f7
39774 F20101124_AABEFM rogers_c_Page_037.pro
1422e2635258fb923abc437b6085e2b7
4aafeed8602ded87bbd9ee1832152f29337ef15a
58506 F20101124_AABEEY rogers_c_Page_022.pro
3911825ad03399f9b059e07d58c4fcf8
fd9cbcba9e5fc1fca450591a7dfed79779bb1200
131963 F20101124_AABDZS rogers_c_Page_127.jp2
d6dd11141bdf5353f5a398c3e9318ba8
5c37963f6744fd24dee66e6c86d8d8124f3e0c2a
53440 F20101124_AABEGB rogers_c_Page_052.pro
d04265956db706a8135dabf2d9249eec
4a2a667d631efdaee2358220d5d0169558d5342d
35446 F20101124_AABEFN rogers_c_Page_038.pro
775d90baddfee65269430b75cc076dbe
90e1aca0f0ba66d33156723c7001269754438025
60460 F20101124_AABEEZ rogers_c_Page_023.pro
2b124364d6ff981d1a6b01064eb208db
c78d967f29e1631c1b81a9e08fd698f38b8d472a
F20101124_AABDZT rogers_c_Page_128.jp2
f9a22917a4747320ff4634df460a6a30
9e3d02a3aed3dc03f2d1389165d2507a026af8fa
4269 F20101124_AABEGC rogers_c_Page_054.pro
9af264e4610c5098d128cadba0d488f0
d7f517b3fc7ed4a5faf3d40cce9d8a3519f65cde
33927 F20101124_AABEFO rogers_c_Page_039.pro
cfa385da179a459591138546b2b22b22
4dbe5cf5672b1a66d0218f55442ffbcc0e3dcc4d
142974 F20101124_AABDZU rogers_c_Page_129.jp2
e8994230ec5e8aae86b7bb42b707d73a
947b691680cac77b7c3e6ccb3713f3d2ec7f7af5
29589 F20101124_AABEGD rogers_c_Page_055.pro
c6fc9619d759e79247448757ce39d3c3
fe095b0a82ac20242100bf8d77816d5a4b54b254
56619 F20101124_AABEFP rogers_c_Page_040.pro
57860b21bf2f3d13b843fe700fd0558d
d47b4ab728df107c1d0349c6dfccde259cae0e9e
F20101124_AABDZV rogers_c_Page_130.jp2
e1c0ee6d021710cc477be8627a894f80
c0cff47a40f1ded465899cd5d5b6ef410716c57c
9591 F20101124_AABEGE rogers_c_Page_056.pro
e90ae4eb304b70e43b898d307a7b2b66
effeacd5d5297064c1b41a00a8b59c4ff3296705
52390 F20101124_AABEFQ rogers_c_Page_041.pro
3d93bd92bcf8d8d661ccd630a84e25d8
426deecfc3b018e47c27e9e9a7a39e88a50c4e50
73676 F20101124_AABDZW rogers_c_Page_131.jp2
73e5db46655ed47e62d816b67e0a71c3
7f03024e4f6389abeb26007d662c1c46097f7623
60342 F20101124_AABEGF rogers_c_Page_058.pro
7a0fad269f6b57bade9d8f0bc00a6c45
1577217d112b434e3aa02db50cfe0036661337ec
40868 F20101124_AABEFR rogers_c_Page_042.pro
11dc9b731a28c2700818438d86124a82
159372936394237ff67e93bf536b89b4bebdb98e
F20101124_AABDZX rogers_c_Page_001.tif
f1b5e3c2807b34056e5ff306a50cfb85
c78eb013cdcb13ca0de5a737e0bfeff9a83fb727
6136 F20101124_AABEGG rogers_c_Page_060.pro
16abae167fc89612630502e534505e37
47534bf0f3ec0687e4b1dcbf843eae600b9ae210
60555 F20101124_AABEFS rogers_c_Page_043.pro
9bd1a89ee4f9b1e488a9ee976bfc2cac
9bdbbbc3cac461994f2c415219a86cbc9781ac76
F20101124_AABDZY rogers_c_Page_002.tif
97dfc714fe49650d3ec4ed816df12617
cf87f0dd0411dac2fc525a03b1b7c7c053c1a1b3
18031 F20101124_AABEGH rogers_c_Page_061.pro
7d3a589cf50f9ce3460b27f9680c8361
6556c8cf9ff0749a5ff8b1b8196eb68afea2a1e6
12591 F20101124_AABEFT rogers_c_Page_044.pro
731714c63a72235ed0a08287b12d10f0
c0db83d75538d8a917363668504008dc802747b7
F20101124_AABDZZ rogers_c_Page_003.tif
ca6f57a8113e741db81bcbf1572fb094
46f423f830138ff2691ca7a92e498da96a27c813
11332 F20101124_AABEGI rogers_c_Page_062.pro
80614c51c9413228456f7b57d5e4cfcb
f09a3fb23a4685f28c7bd13ca5584e56017bbb92
2655 F20101124_AABEFU rogers_c_Page_045.pro
24a0e2b6ce28341ef0b747b60405ed6e
fbe89fb41d1515a844e593d983ec0f6ee673505f
39607 F20101124_AABEFV rogers_c_Page_046.pro
dc77dea18ab25955f67e57bddc293cba
2fdae220227e07655cc05d48af4a91286b6f6be5
5788 F20101124_AABEGJ rogers_c_Page_063.pro
d8697461f16b8923482f65d30b29a43d
f0f86fe449c851820dc375aac8a8ae86ecf2a81f
54418 F20101124_AABEFW rogers_c_Page_047.pro
b85434b2f1dd3bdf683422364164ff44
490b0584618995e96e2e20afbf33245a7f183a49
49376 F20101124_AABEGK rogers_c_Page_064.pro
9766348f6d609fa01ad5ce0152aa5139
a2aa63b8f21ad89415e2ba84475f0a36b3d90757
10104 F20101124_AABEFX rogers_c_Page_048.pro
3c87148b88d75079dbafbdc7fd1a7972
b41a95319d6c427630cda43f0f2a8c1dba30db9a
38724 F20101124_AABEHA rogers_c_Page_080.pro
860188ec021f28232ab8673a018ea319
209ea0dc87905162a1db4a0acedadc851e49598e
8401 F20101124_AABEGL rogers_c_Page_065.pro
40bc0917d73be365c62b7d93aac45c02
6ef761452451efc3c9233c95596b2fb03320c7c2
53612 F20101124_AABEFY rogers_c_Page_049.pro
7556f635db91493670ce3fad351bc442
4c27ff649222641f083d43bd481abaa6965a03be
17985 F20101124_AABEHB rogers_c_Page_081.pro
f3dbfd7d67c7db318f6e241cb54ed878
f21fb3bc2e9cb689f0b568239fe5b09ad823d014
7184 F20101124_AABEGM rogers_c_Page_066.pro
5b3da26073c84eb961d5b4b9391b7be9
3cba2fb3779079721edede6d2cdd5bc3318953a9
20587 F20101124_AABEFZ rogers_c_Page_050.pro
aa88db3f8f2dcb2407986cf45f761991
835ecb0ac414f494b752a32473ed93071703e52b
28434 F20101124_AABEHC rogers_c_Page_082.pro
59f9e1abb6369b9c49e2a634ce4fbee6
596d0b2ab4ff6e2da8288e2fe3bcc969806dfa25
9550 F20101124_AABEGN rogers_c_Page_067.pro
3fe55b1652e2119bb5f12860dc377b3b
de3a19598789b5601998532b1c1c9c2ffff81beb
33641 F20101124_AABEHD rogers_c_Page_083.pro
006e46069d66c66b891d946fbfc930d3
e4c0f401a76c87e24f5e258dd93197bc5464691a
60352 F20101124_AABEGO rogers_c_Page_068.pro
d3aa4b8a0c73bced3e81610b444a25d5
d2391f389b9cac2a6bd26bc50c02c3d18ef2146c
16834 F20101124_AABEHE rogers_c_Page_084.pro
87b6efde65f5257ea5dfa0476ea76e19
51482cbd2baf616f6cc44c02d67c7ff21e69bdd9
9842 F20101124_AABEGP rogers_c_Page_069.pro
85e959d2dbb3e5acc7cbd218e1abaa99
496c571d502a71e5b889708c2495acf5334f36ee
35177 F20101124_AABEHF rogers_c_Page_086.pro
2a78efc2b0a54325a4c2e20b909a1169
046c84d6619caccf757d926160577ad3ad327c87
2717 F20101124_AABEGQ rogers_c_Page_070.pro
892953950dc4f1e758891bf25301a8b6
d6639d3469a69bbd6eef1c2a6bfbb6b2060930c5
43662 F20101124_AABEHG rogers_c_Page_087.pro
5ceb49d4301069be0c625daa55ffc5f1
76ca6cd87f956ffe3fe86e403d08dd3b10b0834d
56358 F20101124_AABEGR rogers_c_Page_071.pro
cb2f1044e3c7b9ae567b1047665ec507
a3e29894ea1b301f4da808e3f26406f16385fc65
40728 F20101124_AABEHH rogers_c_Page_088.pro
45314cea1004e29d13e139a9b1911a23
40375731d3e161f659ce8f6b50a26788e6699f0a
31064 F20101124_AABEGS rogers_c_Page_072.pro
f5b0fbfed793449678b28ea1583925d3
e7fe5c5c0bb451b8e365e8e17cc6c9203f26cc80
49564 F20101124_AABEHI rogers_c_Page_089.pro
049ee1897f832207018cb81cbffd1864
19ef0ae8db97360e424e0dc2b691a5afc4bf3e1d
36583 F20101124_AABEGT rogers_c_Page_073.pro
093cdda49f58dcd85c1005d94d250749
db1bfda73848336aa572effb150d73d07d2245ad
19910 F20101124_AABEHJ rogers_c_Page_090.pro
f720ab62f14298e9e2029d7dbed71d3c
54316dc314d380ed1db5ebbdff9022984f655c1a
6852 F20101124_AABEGU rogers_c_Page_074.pro
3b63bb1f4fa161b298c8040cdf623378
7f642c7d0c5253bc47145f488f3f44a7efd1d09f
55586 F20101124_AABEGV rogers_c_Page_075.pro
87accf9ff409043878a13f1d1dabfc82
020e8f1fda8761cdbd830677f4c8c71d81531078
54005 F20101124_AABEHK rogers_c_Page_091.pro
33a262bc2ab2926a2a978309411ddd46
8e9b0fae6ad54d3adb98a46a6aae4c7e94fbd8f2
49937 F20101124_AABEGW rogers_c_Page_076.pro
49c3702f8ef0b047f52f73156ca8fe3e
b9e549bf9df6a7c8de66b2e927f9ac2317bbc9fa
28335 F20101124_AABEIA rogers_c_Page_108.pro
e34e32c29ea7fe75da44a03f85ed40b6
a4222589ba414d3837dd122b67d69c612196c2a3
33565 F20101124_AABEHL rogers_c_Page_092.pro
8cb289bf65d02cda89e2472950192c36
93c9bd581a999c75f1ea7f45f3eee2cf86203004
29978 F20101124_AABEGX rogers_c_Page_077.pro
187ec59d2b977799e4d7cb7b053ce0e4
8285125d75a32f091745476eda830bc857715873
54167 F20101124_AABEIB rogers_c_Page_109.pro
389db3bd4ee8d4ff1efc521c244f80d7
c0fe20133c1761f0a914c469c255a771f6833ab4
6974 F20101124_AABEHM rogers_c_Page_093.pro
f19abdf24203be0123b7064db319494a
42162d7d0167a95cf70542a8f436238914f44e03
49297 F20101124_AABEGY rogers_c_Page_078.pro
16a2e458c3fcf379e1a0f7d74852b191
f51e9ea93ca7ef97362e5e103103870fb0806bc6
33260 F20101124_AABEIC rogers_c_Page_110.pro
6fd9e14816fa85b8f9af33f352a7ff60
bfd9923efdd9b9958db2763614ce39ca0d353ae7
39799 F20101124_AABEHN rogers_c_Page_094.pro
a4bc79f45929dbf2c632327e8bda8f8b
c440e5d2d427153c91bb50deeae354b7e5b28880
30056 F20101124_AABEGZ rogers_c_Page_079.pro
b865596d8ab9d834c7e91bcf5b07bb4d
7e7b75236f989ddc319d3896a184c5af5a47c397
17891 F20101124_AABEID rogers_c_Page_111.pro
c34fc37e3def49405aee5a5689e66d03
0f31bc1a7b38c82d11b2c6b7466a3c09a935ae90
58651 F20101124_AABEHO rogers_c_Page_095.pro
41ef8fdb9ab6c26c0a6e5327dd91e8e7
0dcbc5b5aaca534bdb59cc24156227bcd36f420e
57616 F20101124_AABEIE rogers_c_Page_112.pro
2003bc8f86318f9f9a0bba9d25543e93
a2bc29dd385d1dcbb97be6e4c07f02115b7c0a66
49501 F20101124_AABEHP rogers_c_Page_096.pro
c744793d635c914670525e7791fc1e47
b5b5f91aaf2ced3d38ece9214e4652057ed59a20
37526 F20101124_AABEIF rogers_c_Page_113.pro
bd799d0ae532c1e8b32b65bcb78f8165
b2cd759c194cd41ed4e248d81bf2476a8c2942f5
34579 F20101124_AABEHQ rogers_c_Page_097.pro
5f2c2d129890da10cae9981b26214dd5
448f963a3fbc937fde782b1a7a85acea97041b0f
34609 F20101124_AABEIG rogers_c_Page_114.pro
cebf6f215ef51e79ae2c3c87c639bab7
0d78eb80cae493e27489c51ac709b05502766aa0
2252 F20101124_AABEHR rogers_c_Page_098.pro
457c9b618278790fa4ee061b03e9eaf3
b57eba51b7da4f50239c8ac90de84139301f4a04
37725 F20101124_AABEIH rogers_c_Page_115.pro
97ba8dc4a11dbb9b1f81503caad7139b
4e42440b4091d5dcf31d77252f485b0a2ccd7820
55717 F20101124_AABEHS rogers_c_Page_099.pro
4a100d6fc2dbc890305f7b280d111410
5bcc8553ae110318edb41ee759dd11aea06e7240
32447 F20101124_AABEII rogers_c_Page_117.pro
5049bbe79eca171f0d41255f395416ef
17cf2a3f1722a7a39cf5c34eee0607767bcb3021
6214 F20101124_AABEHT rogers_c_Page_100.pro
133e69b9026c54a352a6fb5e76457589
c64eeb528cd0d97d131795b2574c548251d4c6ab
57293 F20101124_AABEIJ rogers_c_Page_118.pro
0ef197e2d52df9c88d17291695cba40c
e42527368e8943a671c01ee10a9900fd503e92d8
57322 F20101124_AABEHU rogers_c_Page_101.pro
5e36ec3e8191d1dbd58bc44cf6a5aa8c
6f3c365db01152c351fc8150a9b37855631c27dc
51913 F20101124_AABEIK rogers_c_Page_119.pro
bbe8d3e48b36ee8bdd411e22af1a24ce
47d3c08d8d3ced4bb82a49856016a989d77c2d89
27021 F20101124_AABEHV rogers_c_Page_102.pro
01dbe3747e957403c97cd025dee01ced
1d8aabae8c379990a6e99a0b134f08692d43668a
6085 F20101124_AABEHW rogers_c_Page_103.pro
c4d21ebe08591f1dba3a5c952cab505d
1fe56ac8831c59be0a64e95bf1b7ce45bf73205c
6946 F20101124_AABEIL rogers_c_Page_120.pro
5eecf7594537fb8c4a1b3305b2aab24f
3079424c14773ecc112940ac158e6a3ceef2c9ee
55625 F20101124_AABEHX rogers_c_Page_104.pro
3d4d38c4b9df7fae67a38f804597ea45
33993554dfe724c7f791e677f9637848b727d77c
794 F20101124_AABEJA rogers_c_Page_005.txt
a58bb102e28159d9208337fe237ed632
ab684a622fb5433cba3f7427b3c63ac3e3371398
39288 F20101124_AABEIM rogers_c_Page_121.pro
4d351271e52f46825f6681a6374e2c3b
70d976f4c69932722a9afe0fa3f7a659b68e7b66
35025 F20101124_AABEHY rogers_c_Page_105.pro
f3e91791e39388ad2ad73c0ebde0e20d
a9dcaeb91a1c1233dbdfb61cbaa033953b0dfa3e
1567 F20101124_AABEJB rogers_c_Page_007.txt
350ca5ade6aa5655d6eb7d622f32b9e5
e056de67a87657499da3afa790b6b3040aab1fb6
35381 F20101124_AABEIN rogers_c_Page_122.pro
9793262423c81f41dee31c124e04ea71
238872a00d1722c6a6424b7ced41c7a257a2b302
30122 F20101124_AABEHZ rogers_c_Page_107.pro
8fbaff3281a25684e80eb566aa31527d
279bd1007bde5b270fac1c710dfba95883a14c06
540 F20101124_AABEJC rogers_c_Page_008.txt
75ec57ef58a20b442b77b2e5f5aee190
f98b180e406b5c15eb8946efb9d28eb6e7bfe137
10016 F20101124_AABEIO rogers_c_Page_123.pro
0fdd0bc038517104bf987338612ffdb7
8e75aa9649bf1ecec33a4102c4c516ce8cf4dd04
772 F20101124_AABEJD rogers_c_Page_009.txt
2feeef70a42d4d0e43f2ff2a8d0b3d84
9eae43be046080d4fd2f7378f6a1e60eaf0d944b
9710 F20101124_AABEIP rogers_c_Page_125.pro
73a573b991c5226b7abf090275145b2c
66060edc0aaaac50be275d1cde9caeaf77a4c91a
2439 F20101124_AABEJE rogers_c_Page_010.txt
32edbb22ca2c64d4ae2e8350f6e207ce
27f4ce60933ff0ac2f7f9d097968caaba4c12b63
64958 F20101124_AABEIQ rogers_c_Page_126.pro
4e09480d92df38c6dd6108674a66e702
c51ad1718e95796eb3bde4186b631eadb480dff9
1760 F20101124_AABEJF rogers_c_Page_011.txt
5f867a0c6077982f4dc8cba668dcfb93
0dd22dcd6afc71cccb2797d12b0508ca67da6d06
64196 F20101124_AABEIR rogers_c_Page_127.pro
7b115657256b3d1d57626354bcd0cc56
f9946324eba84a2316dd2f106cc4b5c3f18e152f
2129 F20101124_AABEJG rogers_c_Page_013.txt
da4660358d48240679c309895e1d2509
40f763ac588dc5e0408ead8bf279b2f64812bd1b
71804 F20101124_AABEIS rogers_c_Page_128.pro
1d7c6dd4651c688af43b4122284bc9af
d76390d452cd49b19ad6def52c96208d2f6fe19d
2305 F20101124_AABEJH rogers_c_Page_014.txt
bd9bd5c11b4d6150688e0e276f1562dc
bcd6504c177e0c8c06af554a757b172fafd9c700
69830 F20101124_AABEIT rogers_c_Page_129.pro
8df91e23a702ea4cd1fd075f7d97a25a
7fafc47ed7d911ad0d706ffd7e67329da60344c1
2317 F20101124_AABEJI rogers_c_Page_015.txt
9151e2ed8a0610ce57b4d8206f591464
67ad85f7f9a94e9018619a36db500c0983b6038f
54722 F20101124_AABEIU rogers_c_Page_130.pro
143523bea12a327bfac6288b7872ae5e
1a70b576cf5d88c7071f361b18d0a730f5dadf60
513 F20101124_AABEJJ rogers_c_Page_016.txt
f34c3b557a1b9139b444b6c72dae99d2
58036f7b3f8eacadd3473914c090b16efd2c62cd
34032 F20101124_AABEIV rogers_c_Page_131.pro
210a1677c9f09e8caf74978433431079
cd189a784e42096ce92ef2bd5fc98100f542a6b6
366 F20101124_AABEJK rogers_c_Page_017.txt
c03d48985f69da2986091491881d8c47
a496a0bf6cf570fbe374b229bc59a1d95c046f4e
498 F20101124_AABEIW rogers_c_Page_001.txt
401be381eba9bec8989b1aa7717301f4
cccae1c9c9ec259980b63b26995d4476592d8c8d
2408 F20101124_AABEJL rogers_c_Page_018.txt
4e4581a7196b2cc2c0425b1c7d2223f7
6c657a81db515f9bae3d265b617e861589b1de4c
2206 F20101124_AABEKA rogers_c_Page_034.txt
d64b275478adea53d86845adda47bde2
5fc743cc7de3a820963d42b83da4a93a3bf16bca
92 F20101124_AABEIX rogers_c_Page_002.txt
0a5356439faeddf71d396e164165102d
3c5fccbfc32c3a9849ed80fd819905c9c1e041a8
369 F20101124_AABEKB rogers_c_Page_035.txt
af2acb2335f887e338ea63e2864bd0c3
d224814f39f6ffb91f96e5cc7f990033fd32fb0f
1418 F20101124_AABEJM rogers_c_Page_019.txt
1038b323195dbc6c4a411ec9d6943316
48ed22092142cebbffdefdd17b18872f19aa96d1
656 F20101124_AABEIY rogers_c_Page_003.txt
3034066d4f5d4216b1dda609470af9f9
0d9b5aabb37b192fcab17efa4c997a58054bbacd
1382 F20101124_AABEKC rogers_c_Page_036.txt
cd245feb2d7f0b7d345c971ae8c5e478
6e12ebca03ee46514d66ff2731ffd9427c2fe1ea
2365 F20101124_AABEJN rogers_c_Page_020.txt
6547031753d7053d7a63474ddf19215a
ab05fd7ab2d2bc74e185dfc8ff54df0b0856a6df
2302 F20101124_AABEIZ rogers_c_Page_004.txt
72134fd20e1d2e6041c00be4f1e9c284
299391d3d578d60c68f9151f13823252593043ee
1574 F20101124_AABEKD rogers_c_Page_037.txt
5b244030f137905fbb07e2e5e151c548
ae42c104e0099b15b6c590a6214ff8938d81dd25
884 F20101124_AABEJO rogers_c_Page_021.txt
478d43fb6b14f51cfde277e7b37cf05b
0fb679348f2192c402cc8701a8297e20a520990e
1471 F20101124_AABEKE rogers_c_Page_038.txt
ddc2d190a344c3c4e5b7d67a3a44fdfd
5d3a72579d1711250caa3aa9627b9f0d803f161c
2308 F20101124_AABEJP rogers_c_Page_022.txt
1a38b19094ce84072856523932381c34
8ced56e8558554317e54f0733529096d44be6fbd
1386 F20101124_AABEKF rogers_c_Page_039.txt
7ba908494139f7c7a353e439fcf35064
3267bf129be8d18182708c71a046b9dad222cce6
2374 F20101124_AABEJQ rogers_c_Page_023.txt
a0a22590f15a5a9b2f65ebfcf7c0b511
c0cfc1f4b1a8186117d853880cfbc0d0ac5d7686
2313 F20101124_AABEKG rogers_c_Page_040.txt
a3c10bc2bee4e65262705a9a93a00f4b
2a06405fd5bc59c00f2212272fcd3d88aa5a3e50
2204 F20101124_AABEJR rogers_c_Page_024.txt
9ec974adc6910990a7d5b0646b873d30
5e954a01b52cc6eeef5dee822cfd8e0e5c58dab9
2269 F20101124_AABEKH rogers_c_Page_041.txt
0ea34ce71891d050658bb576972ae3fc
7a5f14cbff73c40d327a215a6cfbf44bff7288cd
2316 F20101124_AABEJS rogers_c_Page_025.txt
23006670b800696e6ed65acc40bc17f6
92427f90424bf44bb0ec4286fd91805d33688aab
2381 F20101124_AABEKI rogers_c_Page_043.txt
f7155c45327ee280db963115caa7efc3
75da5f80845cda9764f3af23fba35431de9cc598
2370 F20101124_AABEJT rogers_c_Page_026.txt
2c29381d18264af6ca24399e5bffb547
d8234fd8fb81af0372107af4dc19dcf5acae4e9e
260 F20101124_AABEKJ rogers_c_Page_045.txt
1d25e54401099a99fd5248d47014c3d1
410a5b5cc4d7d0c1699400e878148d350945ddfb
2189 F20101124_AABEJU rogers_c_Page_027.txt
24c404e93569a8e6aed7bed75749b452
f0c3962f9c49ef2f4178dc85ac6b875c5d4a3881
F20101124_AABEKK rogers_c_Page_046.txt
f76760b345a7220be329889d59e98c4a
649cfbd28b31b863c4327267f827b9f1b7e7e5f0
2338 F20101124_AABEJV rogers_c_Page_028.txt
a667f1a157b96ac66cbc06bf8db35348
044e04aa27e247e3afa0a6cae7859bc8fa3f20c7
2172 F20101124_AABEKL rogers_c_Page_047.txt
28c78852d48c84022e16036d6b08e4ac
7c3afdcd1c2d193485c5648ac68b9c1608c8f5c5
2304 F20101124_AABEJW rogers_c_Page_030.txt
ee0e940e2e1c72e1a4f916d64d3db53b
496ef0540ad1fd115484f37e78a0c11bb549fe80
726 F20101124_AABEKM rogers_c_Page_048.txt
5a081f8a2d4b44c59920b6617b373abc
5d1e01c9928044fffcc3c263fb1bfb217d84497a
2267 F20101124_AABEJX rogers_c_Page_031.txt
4d0c752bd2a82273610ca5358831e4b1
12290f6ae6ca415e9bd01aed5ed16490cc948fc6
2073 F20101124_AABELA rogers_c_Page_064.txt
762eeda9cc890aa2054c332b97112862
b53d15a9530ebb95c4e07f117549cbc40c25d526
2226 F20101124_AABEJY rogers_c_Page_032.txt
f2be92e5464eee0df27ee2c72c6709c3
519fe9b5747bae2984255de48f1b1ee51a0b3dc7
546 F20101124_AABELB rogers_c_Page_065.txt
f87b837fb930e35245f99ff54412da17
5a7dcee05833e07b276ca614b25bc4ce2bae02e5
2121 F20101124_AABEKN rogers_c_Page_049.txt
d5f56ba47acc61cc0cf5ebf5a1629cde
9f768ef83c147769cd48b66b0df58834e6bb8c2e
568 F20101124_AABEJZ rogers_c_Page_033.txt
0c891fc9031c0fdd9c9177b208a41c5d
4e9427225409f7e6b0bb3de0d35bf3cff9143d1a
494 F20101124_AABELC rogers_c_Page_066.txt
fa1aa3679388061a20851518e35b3402
efc3ac78b4c30d834e68d8e06e35a9a2b8a05b29
900 F20101124_AABEKO rogers_c_Page_050.txt
62a48b2c2966e6ca7da1c48228049cbf
db93cb8a6177839760f76ec18e06a81fa9223bf6
576 F20101124_AABELD rogers_c_Page_067.txt
c84e4380d1f6fe4830d04f65893f4b02
d23604fd521442f199e2ddbbbd0a63a3fbe786df
1657 F20101124_AABEKP rogers_c_Page_051.txt
e40c4217b563baae727525cc51a361c5
af771059f6c3dce2397c2fdd6dca7d7a632ccdcb
2413 F20101124_AABELE rogers_c_Page_068.txt
c870167358ba610545b37001e9489e74
ba1e3413244349379bc9224b907a80ffee996ada
2199 F20101124_AABEKQ rogers_c_Page_052.txt
aa67c1cf930f92923682595c8c406ad3
d9faf5020d749972e9ad657b7be4881e27e837d7
F20101124_AABELF rogers_c_Page_069.txt
ed43f18cdede0558d67890d96997d311
3e6b84440faad1145c06b5da4093378b086e241f
795 F20101124_AABEKR rogers_c_Page_053.txt
b6fb9cd4b541a87094969e05362a1814
30bdd002155e362b0e21ded7c05f1052baf8d3ad
113 F20101124_AABELG rogers_c_Page_070.txt
c91dd8fc0d872e89ff85b83085882609
5c659ab36f96a4025e9de5ba5390035085aa061d
1417 F20101124_AABEKS rogers_c_Page_055.txt
0258c95b469336a7e5b3c88302c18832
6e6393251845b21591d7d064acdf93eb39e4b461
2301 F20101124_AABELH rogers_c_Page_071.txt
6925a58da09883538221227103f71370
9592a1f7cf72cebd25208cbfc5a75d462d2f747a
618 F20101124_AABEKT rogers_c_Page_056.txt
6459fc9d72d9b66e0cca618b01cdd09f
47b51cf1fc4aee152985b82347cb771869f5dea8
1354 F20101124_AABELI rogers_c_Page_072.txt
f4dd37f4b1ee9c53a26d89a4a78e40f8
373e00819ce2b695c824bfcc51c9567efca4c334
443 F20101124_AABEKU rogers_c_Page_057.txt
4d4cb68a7ea12282835476d8d1ced656
70e29094c8f63a2c9ac57e8797305588a01b2837
380 F20101124_AABELJ rogers_c_Page_074.txt
46c3d31cd9c4bbb66012c190fa2c4689
6df25df59d605058281ee71598e28ccfd247e077
1628 F20101124_AABEKV rogers_c_Page_059.txt
0c681467aa436e06eb45637bf5a588be
cff0c8baae0ceaed9bc5f4e40acf2b3a9b74213b
2259 F20101124_AABELK rogers_c_Page_075.txt
529bca19709051be8a65facbfc9f1fc4
e4fa29d160b9f703cb800d62ba5d19657c2ff5ca
555 F20101124_AABEKW rogers_c_Page_060.txt
de109b00ce3406b202a57d81e73bb969
1753170587921e725877a9ec7376da1b4190f193
2139 F20101124_AABELL rogers_c_Page_076.txt
8b5b332394887ac211333c80e249c918
6794d3cff590caa94ed9ce0cdd4660477881ab81
820 F20101124_AABEKX rogers_c_Page_061.txt
4ffbd4ae2fccfb41659a1176b887e28b
b1da1c7277732df9d43755ef34ac969e35e50e58
1336 F20101124_AABEMA rogers_c_Page_092.txt
ebc2083ed899e96e317e6aa66cbc44f1
f8dbf49676aa3c36cc08cc4f2513f66ba1de38e4
1221 F20101124_AABELM rogers_c_Page_077.txt
3eedcd8a9724a2c9bee71a8024d38282
11e34406fd3408e659c9de10b1d7d10c331a6baa
558 F20101124_AABEKY rogers_c_Page_062.txt
0e50bd5afdfacd4af46ea9365f31bb07
ce73752dcef3ac0902e2c8083ca1c3027a5a9218
1611 F20101124_AABEMB rogers_c_Page_094.txt
59228925b50ef240118b2f7f338400b6
a53f752a687c9b5d3b6046542e61342e7176db12
2059 F20101124_AABELN rogers_c_Page_078.txt
8360dcd4eff0ddd65aae3e0c0d47bdee
0a935fc96efc482fc5ed158662bb3b01f9dbaf96
505 F20101124_AABEKZ rogers_c_Page_063.txt
ac0a78b87a22567435be69299e482f3b
8504df7634b04a1ddd75221429bb1ecef6116d75
2315 F20101124_AABEMC rogers_c_Page_095.txt
9daaa0593c6e5135ef1ba62463309c97
04b858d286075065cbbdd5dfba741a512e5c2347
2091 F20101124_AABEMD rogers_c_Page_096.txt
646c3ca52e94a243604bce9fde66e6db
0a469861f3c2f8c3f0d58a6f48a218113cb65620
1200 F20101124_AABELO rogers_c_Page_079.txt
5a8063e3a759b087d3fc4f6eee556458
ad1f8e7a1104ec81a2b3c690b39bb7dfbe00c36b
1502 F20101124_AABEME rogers_c_Page_097.txt
380bf2a94110efd04639684fcb91594d
7442db8a65bf16a9087c07d5de4de6650c0ffcb6
1659 F20101124_AABELP rogers_c_Page_080.txt
f6fd4c587d1b39681d5f4c92e17cb2b4
0545515e9512c9087c853ecb8fd4d69f7ccfbb02
F20101124_AABEMF rogers_c_Page_098.txt
8636909a29d107666d3647694cc79b94
5d6d7349c46ae6d20781f6eb5b5786ca56fff346
1149 F20101124_AABELQ rogers_c_Page_082.txt
056dfb3d665ff4e223f18ae5e9af85f9
275beb73d6f0d43c59621cf2972a3f093acbee72
F20101124_AABEMG rogers_c_Page_099.txt
f68849511b0e805cd86c711e98167997
e2c7cfab006b9b17647f2b87e42b2d48c9ea6ed9
1428 F20101124_AABELR rogers_c_Page_083.txt
cd7c6cedd5cf7d344b91fb1c7b9d187d
ffaf17f4c5e569eab461c84fc0eedc81b0539d35
324 F20101124_AABEMH rogers_c_Page_100.txt
cb4e820eea5a73c7e1220b1a6aa8f7b8
f5e4f6fdbca7f4797a96c471878ee591577e9b62
640 F20101124_AABELS rogers_c_Page_084.txt
8c5bb47cb99669321f6c1769117c6319
30c27fac11297c7752dea20b324e1c4094a5c8a5
2258 F20101124_AABEMI rogers_c_Page_101.txt
c7b341ddf8eca036ee19ec62562325ac
1a443f7fef8d3c4af9d86518d7260c4cc6f4d551
926 F20101124_AABELT rogers_c_Page_085.txt
431d298e27cb7eaf588cbe863da9e54a
37edd2835377d30e2f1291d77a5c5302f10f6533
1139 F20101124_AABEMJ rogers_c_Page_102.txt
53e99b9a0f014ec04d2c324f09d95f4e
173224db843113aeeb11b41654d9cc75a301be3e
1412 F20101124_AABELU rogers_c_Page_086.txt
4a1c710573867befbcfb39f27c3ca9de
129977643b3f3f6d3ea12ebcb6e59ac6185cc7a2
1661 F20101124_AABEMK rogers_c_Page_105.txt
23bfa97e0049e9b3585870f7172a06e5
46c49e7a218faa281f440ff9c15e5050da4a110d
1737 F20101124_AABELV rogers_c_Page_087.txt
b73b5f8772761a6f874f4a204cb151da
c43cf3acead4450713f4ab9948e5dd8116aa0229
1545 F20101124_AABEML rogers_c_Page_106.txt
fd970ed6a09ace30957005bb36794a8c
c5d2dcb67a41a02ff952821bc906338eff31ceca
1632 F20101124_AABELW rogers_c_Page_088.txt
53fd79fdecd2e000d41e8d64e655e7cd
1195e2f69a5081d22cc53dbe3be143cdbf5eacba
1426 F20101124_AABENA rogers_c_Page_122.txt
e33092594cd5b8c9962e6fb808ec5f7a
4b65c5401ff9ee3a5f41ef481a6dcad3821e55f4
1317 F20101124_AABEMM rogers_c_Page_107.txt
27b47c32fd5f8b5b240a4d2d27d773cf
fd79ef551272ed984b93912cbc3c94e0fdecd907
2068 F20101124_AABELX rogers_c_Page_089.txt
96885042ec55791bd87e52d233531bef
5064cc31ffcc54f6d1f0455f57ba23deffba17ab
407 F20101124_AABENB rogers_c_Page_123.txt
e4c0e282bd4bbd16595fe209d3b06050
9bc5b3a280fb59096c9233abead89b7cb2770c19
1366 F20101124_AABEMN rogers_c_Page_108.txt
662d04edceb4e2b33dc10e40083783e7
a70b7e8177962d8c46baf27c2df712e83953ab41
978 F20101124_AABELY rogers_c_Page_090.txt
6dc20004184272ee3e69b624b6f66e07
38249ef45b9e46864a8e68b0b2abc4bc93612303
745 F20101124_AABENC rogers_c_Page_124.txt
1aad11cd626fd8bad74e0d79f858c6b4
bde0d55cfb3ac2a2df16ee772fe507e7c22fbfd4
2237 F20101124_AABEMO rogers_c_Page_109.txt
a0f8d4ec804db4e2db1da34593c60d94
da3e03842ee5a23a1def6f7c511735e910ce0bb0
2262 F20101124_AABELZ rogers_c_Page_091.txt
49cc0fd19d78455bc578a1f1a2a4149b
f811c92aa521bc6bc552b80e5354393fdf735cd3
463 F20101124_AABEND rogers_c_Page_125.txt
d54a5c6cca35e629a3ce08dad5bdc010
8e2285d9a0f781097cd1a23dcbf65a69d912755a
2606 F20101124_AABENE rogers_c_Page_126.txt
8615d0373c8fc8da2e5081e9fd09eb12
55d2705ea67e32fab6ced85527a11b5ea52532f3
1497 F20101124_AABEMP rogers_c_Page_110.txt
4cbd4e10e9144bb4bfe9eaca0df2b00a
11408d17362af28a470eb21c634f5fe4a0eb5a6c
2580 F20101124_AABENF rogers_c_Page_127.txt
52adbfbdf6f1ce8e976196d63d92358f
fd7af38312e96d6ba10f854ac310e4d15321ccee
911 F20101124_AABEMQ rogers_c_Page_111.txt
c1d9b6a3fe3b2be09e68602a9943e989
aa42fd7a00a91ae15cb294f953a4b92e3d00cb86
2872 F20101124_AABENG rogers_c_Page_128.txt
3aa129d39f339958ffe6ea5942eb42d3
7acc2153d3c43a53433448e3ba4c3afd1b7e32d0
2275 F20101124_AABEMR rogers_c_Page_112.txt
4d257d3cba0ddd7a264f5e54ee5d33b2
7a353c18aa43a919335847e9655e7536837c8a3a
2795 F20101124_AABENH rogers_c_Page_129.txt
05031a96234ef2b98663001dd7dea0a4
ec96e90c2c5ff55ecd3d15baa383bca26975dd19
1606 F20101124_AABEMS rogers_c_Page_113.txt
861676b096e7d8499956d077ab6fbd37
b620a9be089717e04a98e59a1cc18718b15505fe
2205 F20101124_AABENI rogers_c_Page_130.txt
534cd6a5c9e60ba3ae46701a317afd78
284e2528d0c6f738a5cc5900c126284ab41463eb
1677 F20101124_AABEMT rogers_c_Page_114.txt
4fc30ac60eee29c688aaeb63057e4ff0
f0c16948a365f7544fc87f76b9670b5fb348022d
1396 F20101124_AABENJ rogers_c_Page_131.txt
5ba658531e75ef8c82272152e0241751
9d3b71cbbe598fb0cddde827af8c1ffca1f1214b
159 F20101124_AABEMU rogers_c_Page_116.txt
427631d4fe251f85ee18048051b49cb8
8ba1bf61e0810986aac253114c07f21d047d8ef7
3934042 F20101124_AABENK rogers_c.pdf
22be3e70a4297033205aebc08a677cbf
6ed0cc5eab12ff7995290d019e9dcee7989c5921
1328 F20101124_AABEMV rogers_c_Page_117.txt
98154c07d2b09d1f2a80d70df10cb3a5
52b66296f1dcc3dab7c4ab29e5b18fb664e9a244
23662 F20101124_AABENL rogers_c_Page_113.QC.jpg
e121d028190589f3301bb5f5465f4d8f
0aa72d85aa4a7f82758811a4901aeedb310b3790
2356 F20101124_AABEMW rogers_c_Page_118.txt
e34dd1e374d4ca39dfed3dba6aa3ae98
609a9a65af092365e31c053ef769d7eecda6dba1
22900 F20101124_AABENM rogers_c_Page_042.QC.jpg
727b459ddea6db5704a09bf961e26572
d24eb7fe3ef6cef6198e5780e410c2adb08430a1
2093 F20101124_AABEMX rogers_c_Page_119.txt
26fb1a1c8cd84592360d46ff742f08b3
ea6f30ab7cab58e5aa6e39cab51b5e4fc0b4ffdc
2999 F20101124_AABEOA rogers_c_Page_005thm.jpg
b0a1d9e45e178be8978cda63fdccfe02
7bfd276ed1b40ca2bd5e0f7f72878ad6aafb004c
6893 F20101124_AABENN rogers_c_Page_099thm.jpg
3b6fbac2aae68dc9a578e6d2d2b3c644
248d69b67f635700769424d957eb3c28eccb97d4
284 F20101124_AABEMY rogers_c_Page_120.txt
defac60b57df9b10aa7f877506b16e09
f3b24253f82050002df405311713800f2ca667f4
22735 F20101124_AABEOB rogers_c_Page_037.QC.jpg
a0e1c4f03f03fc2d5827ca025c36ed2e
593e40cb962134235eb9269091a188b95fb470d8
5182 F20101124_AABENO rogers_c_Page_044thm.jpg
5b6d48ec0e5d21ae7faea2af845fad84
d938e6b72c8c23057fc3868e9989399a77c89a9e
1768 F20101124_AABEMZ rogers_c_Page_121.txt
56f8680648d5ecfd915f7f677a01b17b
8f98f43a21be31f5b6834daa64b668b7053a6511
6685 F20101124_AABEOC rogers_c_Page_022thm.jpg
7324573c9b23acb2bc421306231f3e95
b25ad804fedcc8e0d99085fa3516d06076aea4c5
6298 F20101124_AABENP rogers_c_Page_059thm.jpg
d775fbc2209e5fe4125938a3db5ac48d
b17fc5de915f348b40907ad9af35a2e5913f0cb9
F20101124_AABEOD rogers_c_Page_030thm.jpg
1ef92bf2ee2dbe6b35a24fd4f35f12af
23e5e849d91d9700fd62b373092ba8192e5246ad
6837 F20101124_AABEOE rogers_c_Page_071thm.jpg
a0de4c88ecb1de3e5f062c93a534afd7
21664091477cb5142886a9bc669d72b5e38b8919
5512 F20101124_AABENQ rogers_c_Page_110thm.jpg
d289ebd3ba2b96f0bdd7ad179bd3514c
9af98c33d26447239b983272dbd6bd5d958a91d0
21624 F20101124_AABEOF rogers_c_Page_006.QC.jpg
3f4a359cd3b01be95acef342dad2a2d5
3c4b1eb744d3b58fb39e58c194fe96ec779b33d1
26924 F20101124_AABENR rogers_c_Page_101.QC.jpg
cc3b54f1ecf4694465db08b5e7955f7d
48800fd255c72ad863836f87aab7c88aaefe0c20
9318 F20101124_AABEOG rogers_c_Page_063.QC.jpg
9db97ec8b5bc40d7b759552f955b5c9c
0641d278054fd0070c99d1d9b4465b884514b991
15693 F20101124_AABENS rogers_c_Page_084.QC.jpg
84063fcb250d0486fe7237e64ad10c27
9f2368224c9fafd12877aa5b38366bd627a8629b
5101 F20101124_AABEOH rogers_c_Page_083thm.jpg
f06e5e0b87727e04fe3ab37988d577b7
64ffd149f492318510c583cfdf279f276abb5e23
6171 F20101124_AABENT rogers_c_Page_126thm.jpg
e6c19ca61ccc6cc51a7a932f20c2898b
0c09e35a37c4207102fef60c79b4d0403e8dee98
3998 F20101124_AABEOI rogers_c_Page_048thm.jpg
7600881591195896738a32a2b69e1a0f
726eb79daae3ccaa3ce179ab4c79e2a38c76f82a
18016 F20101124_AABEOJ rogers_c_Page_102.QC.jpg
aefbf2263ae9d0ea619d903b9a84997e
9207e93c29c89abd0297bf794765825607d7fdbe
11869 F20101124_AABENU rogers_c_Page_077.QC.jpg
0e6ab1c5167eec90ff95e3945d6c007d
dcdec0b4e8aa00c59659790e7126d26180ae2738
6691 F20101124_AABEOK rogers_c_Page_014thm.jpg
c5763d6f54b680169e501fe797a81a83
bacba53f77ead6e005549f8ba24e4a482d1c96a6
10134 F20101124_AABENV rogers_c_Page_124.QC.jpg
526b019d36769dc4185ad3b31fb3330c
7d7f3f730302274cc31acbb163428525f9211670
27751 F20101124_AABEOL rogers_c_Page_032.QC.jpg
92882f28f0cb32dca7366231b113afc3
54aa7c2c8ce0ea96a52579ad02fd54db8e1d0821
25329 F20101124_AABENW rogers_c_Page_011.QC.jpg
0670e1a3261528aa1757eb41dd4e254d
f852c60be2791fb18e68aeb034f5d40edf28d643
22143 F20101124_AABEPA rogers_c_Page_097.QC.jpg
e5595ff64b1eec82bf0d5a0de9ca43af
550e2aa617012b6ab98844d4ce95a132770affc7
16557 F20101124_AABEOM rogers_c_Page_016.QC.jpg
f5a7483c1bcff89aa8fb8df7fcb8d3ef
ca43c58a67e2881125f94d252aad4829b30c7329
3402 F20101124_AABENX rogers_c_Page_090thm.jpg
3e12f4686980d694852ed95475160a13
acacffba793a4c0cddf5048ccebd8fc19d20ac11
6694 F20101124_AABEPB rogers_c_Page_029thm.jpg
035f5e4b5fe2305ee43e78de40d84d08
7a0f2cfcda6fda88ffaecb4995e023adad59045f
5157 F20101124_AABEON rogers_c_Page_120.QC.jpg
652271481e7723b0eba749dfbc6cdc10
77a170c61834558157d63f4bf70cdad1acfd4b06
6265 F20101124_AABENY rogers_c_Page_096thm.jpg
41e858dc0707d57e8ad0da6e3fbdbc09
32bcd191f30eb663f1999668d64af93042e8923f
6560 F20101124_AABEPC rogers_c_Page_068thm.jpg
e002e43ff74788f7f490317faf407160
896d3158c94098b3a8ffc2982b6fffda8f4bf2d0
25340 F20101124_AABEOO rogers_c_Page_021.QC.jpg
0f70085b3b5489d777241934a5406e47
e3558607219afd3c2f0ddad7e0b9f13fd20543f1
4481 F20101124_AABENZ rogers_c_Page_062thm.jpg
ab9feb590eb9644adef11050781f99a3
bbcbeeb35fa91fd1cd1db63f087c4c5263164c68
25795 F20101124_AABEPD rogers_c_Page_088.QC.jpg
f054f2b872467db4c5cb00f672949880
fd61b18ea226e59841fd08afa2f61bf69973ce5e
14373 F20101124_AABEOP rogers_c_Page_098.QC.jpg
334bcf261ae7ccc637f7d850ea001e7a
1b2530ac45a3a508acae82de66ca450dccd458be
22615 F20101124_AABEPE rogers_c_Page_019.QC.jpg
9f3bd0bf37744b276ec063db88618a91
60a0b8b5dfc40703ba5172212a43f3d46867e971
6917 F20101124_AABEOQ rogers_c_Page_021thm.jpg
cd3a8d6c3694bebab935c42f70805807
515a0f7f40c1d142f7e2217b8dadc5e5cf8a2a23
6962 F20101124_AABEPF rogers_c_Page_023thm.jpg
add0b4679ed8ffed3daf954779b1f300
487911caccda8c911165aeda93c8c7d9c42a3fb0
19746 F20101124_AABEPG rogers_c_Page_110.QC.jpg
b54026b6b2f34bdc799b5b75a68e9b2d
b2d6f7474c73c4cc62a442890eac1586abc38a71
21556 F20101124_AABEOR rogers_c_Page_073.QC.jpg
cd44ce024b2d97fa3e3c786ea0cac60b
4f61563681f9bd7e9f74903d0c8b210bf7d63be9
5776 F20101124_AABEPH rogers_c_Page_072thm.jpg
2c4af78c2efc8497552135e7668e2635
03f6e3e606a8578d1417b7a34f2dba0ca1be48c6
4289 F20101124_AABEOS rogers_c_Page_065thm.jpg
2f027a94b9d19dc39ef578a0d8ed924c
9195d866cb838f763fcb9883c4495123f55133fb
27776 F20101124_AABEPI rogers_c_Page_022.QC.jpg
e380b2b77d81abe2aaf657d93dd8ed5f
b81cc0da4ffc69b230da0ddfb1c99426d5f7ca96
6480 F20101124_AABEOT rogers_c_Page_129thm.jpg
ebaebf3b7e32ace938e769b09c1abfe9
4d5608a8c328b461229a799842f942e0b1bc4ed9
12765 F20101124_AABEPJ rogers_c_Page_066.QC.jpg
a8e19d3f0948a38c101a0d7cd7c3e6c5
3f6edb876bcfd3146e248ce406dc3214e3ca0209
18034 F20101124_AABEOU rogers_c_Page_053.QC.jpg
6a1ddc697b4d10ae9ac1ec82f527f793
bd8e19bc9fa9331eecb24ecc9d03b6a9d6c6bb4f
11506 F20101124_AABEPK rogers_c_Page_065.QC.jpg
e7e134b87c59a7f8d4af92f09031bd67
4debad9bea38aabb35819d905369ddf02e973180
3268 F20101124_AABEOV rogers_c_Page_077thm.jpg
52947a87abc6e00ae2b2b9a00e9b708c
167330a38d1e67a38c1f9bec2ea0fcc013c82e86
6920 F20101124_AABEPL rogers_c_Page_087thm.jpg
3317a103cf26e5b2d1612122eb7d445d
f881beb5700045c4b106af145525e877ab8294ae
24241 F20101124_AABEOW rogers_c_Page_041.QC.jpg
c8accc1574b599f6dcadeb576ca43f0e
c249c73bff57f9b55cf770e7202ed0a5c1eebf09
9722 F20101124_AABEPM rogers_c_Page_009.QC.jpg
1043b72eb931db1166573e7b02fd4e02
01e04b63a914f3247328cc06ee81410333a3b871
7774 F20101124_AABEOX rogers_c_Page_001.QC.jpg
4e5d511a5131968726f33e6a25bbd813
b1a76a49bc095662a8791dc1684e55a73bae32db
6818 F20101124_AABEQA rogers_c_Page_061thm.jpg
81dc77188102e5d9446b70333ab81e41
93c8379a2c3000cc0d41feb1999fca646dabc4df
6600 F20101124_AABEPN rogers_c_Page_034thm.jpg
4c56074ec979dc9025cff60eedcea0e3
5d73147027a3d2879be7303ac1ed0a7c48c73aa3
20088 F20101124_AABEOY rogers_c_Page_046.QC.jpg
33df26627d95badee58bd0ae04717cbc
77205ee03245bbe4505bdb1e58b61385b3e287eb
5996 F20101124_AABEQB rogers_c_Page_011thm.jpg
c31c39857e26592792baa5eab5cf8c84
c202ff2fba41db8aab0f65a5650479b1b8b0a443
26504 F20101124_AABEPO rogers_c_Page_091.QC.jpg
969448cfc4c9c42eaaf862dd962e6965
876338551e62d73eaf46843a71a385fcbed43ff2
3306 F20101124_AABEOZ rogers_c_Page_103thm.jpg
6c99f0768572091e490aa12cf8fd5394
3277c91ebedc0e8d0f4dfcf315c984de35470cac
1474 F20101124_AABDNA rogers_c_Page_070thm.jpg
74e5b3a19c036340d72ab2565b4e4550
53e3a959e2cac01d3b84f97bb522f8358d0b287c
2726 F20101124_AABEQC rogers_c_Page_008thm.jpg
5f9f42d2686c5c8459520b1f15342980
eb37d35d24ac3d8c5df4940faaf454d265f1d392
F20101124_AABEPP rogers_c_Page_078thm.jpg
ec933110a085f88c9a37bee93786a308
c4c8ca5746570ee1c5ed2db3efb7601e1dd42899
693 F20101124_AABDNB rogers_c_Page_044.txt
1ea39b6aee938266e35b70b3f159775c
654c04a33bb0341930d4083ba119a6de44e981c4
5789 F20101124_AABEQD rogers_c_Page_108thm.jpg
a2163deebe1d3d5e93ba4c88996a348e
4212201e16f43a97044986fadae439c69d8888bb
23726 F20101124_AABEPQ rogers_c_Page_126.QC.jpg
944cc05442f5a96c13a2922266b99573
4cc57513781d69bd3522ed1c5febd913f88eaf08
88536 F20101124_AABDNC rogers_c_Page_011.jpg
95b9f336614ba65daedef1f79e44dbd3
1940b475d2c427808f974a23e532ab61d7d11dd3
6165 F20101124_AABEQE rogers_c_Page_004thm.jpg
102c5c35a40b1cd8d9680a4505599edb
b53079b550b6785ceed8045d0932c7d7cef58482
6262 F20101124_AABEPR rogers_c_Page_097thm.jpg
3ca256c29d8f201e8ee090cc9a068a1e
d5d079f02a91053086f8ad4b9cef809955e22a08
27705 F20101124_AABDND rogers_c_Page_020.QC.jpg
70981556e7579edbfa44c08701de1d28
a3719a8f8c004f409b6e01f873e9eb6a03c2ae31
6606 F20101124_AABEQF rogers_c_Page_052thm.jpg
8ec3599f44d8a29e0c6fb922569751dc
130157f036ded053cfa2afb0971a5a74bf8ec7d7
4301 F20101124_AABDNE rogers_c_Page_056thm.jpg
708230beb48d30cce9dc1e24d459c014
8eb8fa2bc580f599d95e2655e4ad90ae4e3164a8
11046 F20101124_AABEQG rogers_c_Page_060.QC.jpg
a93d08e0158cec07f45e4010b7c30d61
bfdb71267b301764dcc6d63c27be3a75ce4d2bd3
27372 F20101124_AABEPS rogers_c_Page_024.QC.jpg
c00aa5f7fd57616fbc957fd51255815a
9be75fcf8e44a3610c2f718a661a66649958c6a7
21518 F20101124_AABDNF rogers_c_Page_122.QC.jpg
c0d9ce5635696c1941d0c42b751f2294
a95625191b6e4084731737d2c752c197fa904542
24323 F20101124_AABEQH rogers_c_Page_130.QC.jpg
059aa242f56e82f737abc9f6c9ea57c9
34fd41d1da0831c84c2faa7182af2c646b29e3d6
3486 F20101124_AABEPT rogers_c_Page_017thm.jpg
baee0730da772823a410a4ae9a652d4c
e5d5ea4223c89e1b0dd8197399d892e66ca89611
F20101124_AABDNG rogers_c_Page_060.tif
dfb5e74e9744d29e80ca433e852736de
5c5d1329ad842fbc2c9264b82a32b49732d6e776
22218 F20101124_AABEQI rogers_c_Page_094.QC.jpg
6693411a045c5a16b56f1fa6e89963c2
e1b13bf553e0fd14e4d01f74130a57ffff887c53
5905 F20101124_AABEPU rogers_c_Page_094thm.jpg
ca7131b803d5647f58dbd65126578afb
7c03206d897f45ff30b1d7c72427ffad1cd70131
43012 F20101124_AABDNH rogers_c_Page_077.jpg
f494b61251ac4be9883706a38e202855
769d57fd3438f22f165b92e83d0f380dd2e54570
20994 F20101124_AABEQJ rogers_c_Page_036.QC.jpg
390cd0294c2d1932087b1ab1969041c7
4965a30a4a79c6ac3bfb2952348fc631f4884629
26315 F20101124_AABEPV rogers_c_Page_047.QC.jpg
d0720368c32a6acb54358335d847ee3b
c38330cb1fe59d653ce1d4ec8d701644778baa2a
F20101124_AABDNI rogers_c_Page_029.txt
30ceaefe3a7c32d73dd45eb7efafcf00
b8b74a74f6e15aba044b934b030ed9275c80375a
6277 F20101124_AABEQK rogers_c_Page_050thm.jpg
8d121bdfcb1234edc04887a34d31fb4e
f65ef4309ff668945cf4cb05bf73e9df8c8166f3
26938 F20101124_AABEPW rogers_c_Page_031.QC.jpg
17de7e8ffcf0517d31197f0408b181d3
9d9ad7553048659847a309dd5aac25f95fa8b50e
5062 F20101124_AABDNJ rogers_c_Page_067thm.jpg
993964435bdf525c0ec04c7a883ccdd1
90751e2a23f5b66601d8237486e7b050ff5a39c6
6513 F20101124_AABEQL rogers_c_Page_121thm.jpg
43cfe189a011e92ee0198df78ff1d080
5f6bcd2ef0d01386b833887e8148c08e932eebbc
6755 F20101124_AABEPX rogers_c_Page_043thm.jpg
1cca6524f648e927dd3656a06cd2d90f
4babb66effd4bbe5ff16f7ce1c6d6138197a25b0
25819 F20101124_AABERA rogers_c_Page_010.QC.jpg
c6aae54b9ee3f985629151213fdc57dd
b415eaace2a8bc0a5ab159cdd69fee7babbb1b6d
19498 F20101124_AABDNK rogers_c_Page_021.pro
d104528f3859a859952907399db3ed45
8c743724ef70a0dc8f8cbfd86d6f172cedc6505b
8640 F20101124_AABEQM rogers_c_Page_008.QC.jpg
f0fee253431374297bfccb63cbc27923
597452a4c52a36e239c906eefdccc63164f5a680
7072 F20101124_AABEPY rogers_c_Page_018thm.jpg
3ed5f5088822c7a27da3ae77bd061fa7
4fe8a04ecca734d8799aca4ae28d13747d6fe432
6399 F20101124_AABERB rogers_c_Page_010thm.jpg
5103da96095c7cbec3ab39c47e2ec1a4
1fd4d0286041cfa913d10532c955e6517b4a12e4
F20101124_AABDNL rogers_c_Page_040.tif
cb66fe0274dbc638dab9cf8d5ebf455f
8393d9f779d61fbddfbe9e716336d46c3b054135
2296 F20101124_AABEQN rogers_c_Page_123thm.jpg
952bf8192c5cb20a34fe1cca2d3d8bd2
f0d713f505dcb35ea86f777abdcfaf713fed092d
26483 F20101124_AABEPZ rogers_c_Page_014.QC.jpg
d11cbc894de319267882ea5201ee62bc
d8d7aaaec3a4453b6227055396ca85da5ff3468d
81629 F20101124_AABDOA rogers_c_Page_027.jpg
5753ca1fa2368edf3fec02d0fdc570fa
45b518660387f02db93f2c20fccaf85bed4d2081
21728 F20101124_AABERC rogers_c_Page_012.QC.jpg
2e87ee79adf51f4569964bf5ea57f213
bf13e74ee0e667b9fa454259c005b1dac40759ce
770498 F20101124_AABDNM rogers_c_Page_083.jp2
0e094c5bfa05a70b24712cb1c6597b0a
6ff72c98e7975d62677ce13caba24cf9d47f5b02
25500 F20101124_AABEQO rogers_c_Page_027.QC.jpg
5cda6e429c25e169cfdd39f119f3d990
bedf9e3828d898f2344d39f43e585581dcd470cf
F20101124_AABDMY rogers_c_Page_027.tif
66d257a171283becb464b5f31377d7c5
17e37a4ae6ffa50e0cdc8a82375b99dfa2c9b845
F20101124_AABDOB rogers_c_Page_053.tif
eec70ea810e0091fe4adda0d70230afd
070af39b2e8bc41f636acdce95dfa48ba4ac1e67
5654 F20101124_AABERD rogers_c_Page_012thm.jpg
1350adcb7299d71a7f3177396e4e11e2
e4da1a2624a71f5f6a30d98710cd31a3d3d1e218
1980 F20101124_AABDNN rogers_c_Page_073.txt
48b7c017626fc089c7ee89f00e714ce3
bc202e232aaf68648a4bef8aff0664f663e1f61a
18300 F20101124_AABEQP rogers_c_Page_083.QC.jpg
d14253c85400dd2e23c530840c391f41
811c231a2ec6072f5a6d09fe7ef899e76da49321
21357 F20101124_AABERE rogers_c_Page_013.QC.jpg
37cf90b34b663521ea76d2cdb0fae112
c3ae7ce789d51e36ed9f2a32c84613e5eb51e9e8
F20101124_AABDMZ rogers_c_Page_078.tif
07416ef157e4f2175df601c99e2e76cd
cdcea4f7fd4dadf8e3cf7a7decbe651a0782f3ec
8359 F20101124_AABDOC rogers_c_Page_057.pro
50102a5de3786f43660db669cee75d31
5fe43d43483d8cdcf19003c6ec5722abfca0ce21
1312 F20101124_AABDNO rogers_c_Page_006.txt
d4d05196eff04bf2e10110fb9894ef8c
f1bfd772e1899d5b4432f2ed3a5f8d8cb12980a6
2992 F20101124_AABEQQ rogers_c_Page_124thm.jpg
e98307b9e7314a73f2d3cd0957a211b6
08dbe44dbba5dc6ccc3a4909c5a2c8e4bff95181
5677 F20101124_AABERF rogers_c_Page_013thm.jpg
c0e96fe68cefbf819f9c7ad03c4a25d3
352ee13a5971570c34427f3786c8919a9a9e5336
43410 F20101124_AABDOD rogers_c_Page_011.pro
966d861d860e0ed82a41593c4f33176a
321d3f4353dd6612913bc1864367085779c513fe
70048 F20101124_AABDNP rogers_c_Page_019.jpg
7ee5ece1362c4eaba81e5ab9dbdbba00
f7c09135681682fe07891dffe2d35bd90084bd37
6972 F20101124_AABEQR rogers_c_Page_075thm.jpg
e34773bf5b0135bb77cdc34f0a7fc669
c0a9bdfb0e27619eff19f97a335e626d3c95ef8d
28218 F20101124_AABERG rogers_c_Page_015.QC.jpg
202963321d6aee6aca5692c2c3f8eb63
6283fb2f185b4940c42bdff61f2221f51d81ae4d
7154 F20101124_AABDOE rogers_c_Page_128thm.jpg
a0bc86c15824366a061ee1208849ca56
50f1b57588a27774a0f64fbe547605eeddbe4a47
1651 F20101124_AABDNQ rogers_c_Page_042.txt
0ac2b16007cca0ff1772c842c521117d
f8ea7fe2be8b14267c914ae6b1abcc70476c7546
196454 F20101124_AABEQS UFE0019615_00001.xml FULL
a225d3251967cb9f31aa0bc59b0b9484
f831f364f84897e6f94af806ec52334d5456a2dc
6811 F20101124_AABERH rogers_c_Page_015thm.jpg
71b5a0dfce5848a114389304489eef5f
770f6c3b6e3cc6b4b8e6f3b811547a00047788bd
5214 F20101124_AABDOF rogers_c_Page_006thm.jpg
dda45541fc348d89c488b3dd41e87806
83616ddd89c53dbb9f1c88f6044ae67eec3488ab
5529 F20101124_AABERI rogers_c_Page_016thm.jpg
e348149019e7edeaf7922005b7af97cd
ab5441858e0b066161691071882fd99e7d69d623
F20101124_AABDOG rogers_c_Page_041.tif
993f0f698856cf1484093f5485df0958
9aca4260be6fd16e609503ff17f2f088950d7018
1821 F20101124_AABDNR rogers_c_Page_012.txt
994f0da261be64e10bdaadf8bd095b3d
204956dc93c5dcf6561e3f1c7f1b60f70e72b5aa
3285 F20101124_AABEQT rogers_c_Page_002.QC.jpg
759ba63abf5103f96d460375b57824be
e070cc34a5a6dd3cf73420ffd17f233846631661
10959 F20101124_AABERJ rogers_c_Page_017.QC.jpg
7de2d08431f59ced4b2be00cdbf9ad2c
3aa936e1323ea436e0d268c08cf6e594d52b8864
346 F20101124_AABDOH rogers_c_Page_093.txt
425a161254d9a9f041b98ecd2235c73e
8605336efdd5df964593902d163f4a8271520806
F20101124_AABDNS rogers_c_Page_036.tif
ad9999c914e1a9259bb3f5a2ad2f9966
3e8c700ca0418aa0c7eb95dbd24c4e32c5b2f66d
1359 F20101124_AABEQU rogers_c_Page_002thm.jpg
54d7a9aa924f5ea4585427b66f42280a
bae2ddca5aead9bea8a7a0036745da8b0fac383c
28840 F20101124_AABERK rogers_c_Page_018.QC.jpg
0d3b5a54ec6889c75cb904200aa3d8df
6fe84e1e75acd143034088c45ddca543ff716374
5714 F20101124_AABDOI rogers_c_Page_092thm.jpg
cb3ac6fa7f77b4beedf668adb5fe9244
4554587303ffa30a988a7708d1475c360ea6a618
3928 F20101124_AABDNT rogers_c_Page_093thm.jpg
f76d34b41eaf33ff6cbe3d767dcb8415
1a675836c9432821f4b77af3c33d7463e3a42975
8411 F20101124_AABEQV rogers_c_Page_003.QC.jpg
515b483abe433cc1219da0a18702c4a0
3b89e46b817ad8138f40b256390adc5eec66c797
6260 F20101124_AABERL rogers_c_Page_019thm.jpg
49e4fa7e88f7fde87af7f3146941c94e
25993693eea411759b34a5f929770aa9704e7506
F20101124_AABDOJ rogers_c_Page_004.tif
79a8a489700f641ccf04ec4c8c96ca6d
1e91f226f3358e3d47116dc47b83f2ba776dbc2b
19586 F20101124_AABDNU rogers_c_Page_108.QC.jpg
6243ebdc236efadff5c5350650742e00
10df3f729b26f863d9f4021033f96fe3fea329fb
10529 F20101124_AABEQW rogers_c_Page_005.QC.jpg
33ce981c40b500cafb29d9c69325b19c
8b0d6c448bc403fa9fa5ab390b84eb4628fc404f
25992 F20101124_AABESA rogers_c_Page_034.QC.jpg
95bc031be09acd0f406b95e388f13cd4
4f92da07afd423b88e243b5471ff377f66cae6fc
6791 F20101124_AABERM rogers_c_Page_020thm.jpg
2f0ec18ddf9793fd73a2fec6e79c18f0
edd68515e88c7ec4661b9340759aedb7cf6016a8
1051963 F20101124_AABDOK rogers_c_Page_096.jp2
3e800c05d69315bde9a5716dcdc53928
2feb387d767769874ca82d60c155efaa3fcc551d
F20101124_AABDNV rogers_c_Page_012.tif
fe219966cc027642fe15589be79c83e8
0680470256870697e2719225331cf809f026c087
22797 F20101124_AABEQX rogers_c_Page_007.QC.jpg
9eaa6b0839940f3c53b4c5bc12936342
6334a6f36c36f7128055666813ba4d5aea662551
12150 F20101124_AABESB rogers_c_Page_035.QC.jpg
8be1fe48b75b156250345593e882a8e1
15de870eadfbbc18eac970cf227fed5ead6af50c
28078 F20101124_AABERN rogers_c_Page_023.QC.jpg
68c1e147f316c793d60bee1d55f47a67
92224024d470fbf7dbb77f8fd21a45ba4f883b79
6831 F20101124_AABDOL rogers_c_Page_026thm.jpg
5b037a0cfdae46d4b6a6bf2fd9a00d40
ac10932c39ecd542310953c1414edc6a71c250f3
2442 F20101124_AABDNW rogers_c_Page_003thm.jpg
eb836f2340b6d33f7d051eb889eaea32
5858b6d2a72a5dd9f2d03b4e2f001dfc9bc4a41d
5377 F20101124_AABEQY rogers_c_Page_007thm.jpg
5690d291015671d8a0d6df4626f88f70
73526b265e032a601be3487a901081d49aa2d8bf
4291 F20101124_AABESC rogers_c_Page_035thm.jpg
ba0e4c90b8beec4fcb727de0c04003c7
3ce88115ff791f00f91fec79de6d4a8adb2a6ae8
6745 F20101124_AABERO rogers_c_Page_024thm.jpg
c02e633ee0e7f8ec9df98501a99c8571
8ebf141cd3f494d7cc1493ef074e42b15dcb7a46
F20101124_AABDPA rogers_c_Page_115.tif
afc15d4e0c9d1dfa90713afe267d76be
dc52aaa3d1604ec1949c9843c300edaeef3174c5
57671 F20101124_AABDOM rogers_c_Page_102.jpg
1fffd41272f23f772ec001446db58806
db7c7ab373152f8f34456c16d4962bb552e621ac
24347 F20101124_AABDNX rogers_c_Page_078.QC.jpg
1690bb5855363a3692dd5c6a04d4c82e
c0d26e32f14de833e3dc1a9197de0290986fb8b8
2773 F20101124_AABEQZ rogers_c_Page_009thm.jpg
ac56113f33c3dc697552904d3316fd77
aa82d4a152f2ac60a5b77b86025f8f8792c630e9
5960 F20101124_AABESD rogers_c_Page_036thm.jpg
da02011209c6130275f0a001ee00ddd3
b4a6904391ccc1d0d4fead275ca81dc229c1dce8
27655 F20101124_AABERP rogers_c_Page_025.QC.jpg
f5c7b8004355320e4c05811c5a0c57e9
ab1ce2218a36c104945c6f91794974f1d4d6f214
1051936 F20101124_AABDPB rogers_c_Page_064.jp2
91ce3870e39c21d1f59500c6050988aa
bbe12376e376ddd9c996eb60d1e1b4cf57ca0745
F20101124_AABDON rogers_c_Page_069.jp2
5c08e16eb48ee13bc343cb3e74838134
8eb5b050626188b0f931bae9c871b114168f5e62
61232 F20101124_AABDNY rogers_c_Page_110.jpg
4e138d79a47756ae8d24738afc3a890e
ee15fdf0cc970f0e6362913d472f8cf90de527c7
6229 F20101124_AABESE rogers_c_Page_037thm.jpg
24f0ebc759e68e540b05b40232b1d84f
4cc355b60b64435cb6dfe61ceefe85671ca8d66a
6900 F20101124_AABERQ rogers_c_Page_025thm.jpg
47fcd8d21b1a72f4d270d3df79e687fa
a37ca3d19741aad8a665eb3d2e12fe25984e40aa
21544 F20101124_AABDPC rogers_c_Page_085.pro
847265f6a452c160e1cc8b9c59030899
098e3f0590e49de3ca9e003e598a2dd2ba83c2df
84477 F20101124_AABDOO rogers_c_Page_007.jpg
d0c178b994842b8c6fb1e682f889413d
5e16df735da1d6bac1b38947c603ee115b8f465d
38646 F20101124_AABDNZ rogers_c_Page_106.pro
7747d27beb8a51a7473df0b40ab645fe
a1f2352d660297d764ba8b4db27259414bf03438
22416 F20101124_AABESF rogers_c_Page_038.QC.jpg
6c371d0bf8e55aa9c6179d0002ec632e
7e5589b3c9b6151d5680da9bcc97a43b32c127a5
27771 F20101124_AABERR rogers_c_Page_026.QC.jpg
18d999ec5a9e684a6b198f086e49ac9d
0b5ec1d7e1277d24ca6a73e3d839b1a63c927a9e
26316 F20101124_AABDPD rogers_c_Page_085.QC.jpg
c42ce5981d72d5e86bbe7b78a553d855
80f8eacb099c42c4f83885f884eeccb1e755599a
1051958 F20101124_AABDOP rogers_c_Page_038.jp2
8da4174df5b4ff78149f22d092ffb4f4
72a89ae2af530a523fd06c01d4d5eb0a7d433bf4
17580 F20101124_AABESG rogers_c_Page_039.QC.jpg
f5983bf66004855ebff819eaf90d6647
55c84a65bdabb0374e9a7a9132b7286029304955
6555 F20101124_AABERS rogers_c_Page_027thm.jpg
6566f7e9668f9a410b58faeafa20a5b0
badfde22943e393bec092023aca2d8599f3f419d
17097 F20101124_AABDPE rogers_c_Page_124.pro
4c8f0813c82710baffac4a5ad06620f0
46b7bc74363059a4bb6d9509af2b714b4e31a110
1649 F20101124_AABDOQ rogers_c_Page_115.txt
3be9c0c9c23fa4e1c8b968ca8b82c710
ebf2edc8b22a3c3d0d9552dc299e7e9494e9dea4
4638 F20101124_AABESH rogers_c_Page_039thm.jpg
3c7dbcf6c809553396ad9253afbd3d7e
0bcd22c3b5ce737a14b9b3e3e28588be48ab35f6
27364 F20101124_AABERT rogers_c_Page_028.QC.jpg
abaf47c7315483c3f5d145634fb2f2f8
66379924963c2ecf860d969539597150239931b8
50330 F20101124_AABDPF rogers_c_Page_131.jpg
9c243dec6fdc42fc21012ef95f7035a2
d30878e529b7a53b47e1a1f02e3ffe63125d3825
11261 F20101124_AABDOR rogers_c_Page_070.jpg
31cf7bc0f00c06687e482e268d53df9b
66524fd54ce789b2bd39f95f7d1549a1406a6c05
27252 F20101124_AABESI rogers_c_Page_040.QC.jpg
732c9197d67b837b93a95c2b40e9603a
9dd71f0511e34d2f99ea2183de6d8f5e4f8e32eb
2162 F20101124_AABDPG rogers_c_Page_125thm.jpg
486012814833dc86c902b1f391e5071e
c08e3b72be9d3f07e922cd89e82b56ce644439ab
6881 F20101124_AABESJ rogers_c_Page_040thm.jpg
5e28d634bb6b9086e52e70fa3d2e2ef8
3cadbd2a2e5f78a875bb2602429256873edda8f5
6801 F20101124_AABERU rogers_c_Page_028thm.jpg
ff58cd89614a0668db9c5142c0bfdc4b
348fffa8ad4ed71c699a761f7c3507588ea36fb2
83351 F20101124_AABDOS rogers_c_Page_127.jpg
c7df285b47b17b236ef370324c6d0053
1db6333daab6dcf47fb641fe02663a9314c9c6f2
F20101124_AABDPH rogers_c_Page_058.txt
90309dc899c451a3687ab6b092258731
00a2f2e926e22352e3034070a58199e1c65f6933
6431 F20101124_AABESK rogers_c_Page_041thm.jpg
65ca65c08ce9534f78a0f5a3f29f4bf2
0914694bebc3759836a05ea5ee2892b8e00c54b4
26998 F20101124_AABERV rogers_c_Page_029.QC.jpg
caee378ceabc2721877cbd4171cf1c8d
1cfc8f4ac78e29d508e1ee5e1e42dd1f1f7346ac
F20101124_AABDOT rogers_c_Page_064.tif
905551b7ada51310063cbb37a7c6daee
cf2577a674a25bb386102ddf4f1271293bdc11b3
2321 F20101124_AABDPI rogers_c_Page_104.txt
b38c93e023b3f64a47bef951633138f7
e36a278444fe1ba7e1b35f807765f9ab1708763b
6127 F20101124_AABESL rogers_c_Page_042thm.jpg
08fd22b8150496cd1373b7fe97bbaf08
a3f08fe71bc09d2ce803c02cb606e28911730b12
6703 F20101124_AABERW rogers_c_Page_031thm.jpg
cc1fe223f6db0fdded51da1089203155
818b363e7eebe4c1c1acb656ccc59bcc673ca025
24703 F20101124_AABDOU rogers_c_Page_004.QC.jpg
f388482ff33ec588cb50ec532101a6ec
1db875a353f2edf672c1c89e90645dbe4ba856c7
3725 F20101124_AABDPJ rogers_c_Page_070.QC.jpg
c72903e0b0aa8ff9ba7e049218c5fa71
ac327ee6815d1ecb1326934f0ddaa9979132ff1c
8216 F20101124_AABETA rogers_c_Page_054.QC.jpg
c4c246933e29792539fdc3a4386ff799
936f37e0cedc633245e00c02d07d7eb4134d2e6f
F20101124_AABESM rogers_c_Page_043.QC.jpg
c2effc8b03d877d26d761d4b5f4b4c1c
64c190dc9439dc98b8b1c1efce002ae4529bb96b
7035 F20101124_AABERX rogers_c_Page_032thm.jpg
09949c6b18fd4b785b2246ee6c495006
6a1506d7138fff0814c5a7c677c2a15ba25d9685
4544 F20101124_AABDOV rogers_c_Page_111thm.jpg
1456dc0320f5ab697b8890b0c5caef76
6228389ce9e5411c2e93c0484c8f1603281edf06
F20101124_AABDPK rogers_c_Page_022.tif
920ee6e3379480020d8ffe3e69477560
6f27d7dd87a889ada59f29a096a4321278ed9892
2870 F20101124_AABETB rogers_c_Page_054thm.jpg
dbbb97221d8a5e36b6561df9337da8ed
df32b760235cf2294fa5794a09f48a9254f0e9e9
16577 F20101124_AABESN rogers_c_Page_044.QC.jpg
846eaa436fabcf4ebf87af805c273058
ed3d2ce95df5950c6f62575f8403497d3c0954fb
13030 F20101124_AABERY rogers_c_Page_033.QC.jpg
7c10deaa09441bec37c544f9736156cb
4f5a9b8ecb8899c3fea93cc28039d142140f7842
F20101124_AABDOW rogers_c_Page_119.tif
1532551dca766f187f76548643bf7982
1b232244aebe6fc470fe324574bd548b6679d874
6069 F20101124_AABDPL rogers_c_Page_118thm.jpg
9261e38054412775d8090399b8721fa7
05e5e01bb209d1ad1f9cf57ac11d8b336ed61dc5
21193 F20101124_AABETC rogers_c_Page_055.QC.jpg
568193300cab983ee792a26b3c7ff37f
822ab233550b3304b9b41ba465089a8a98e68f0d
13875 F20101124_AABESO rogers_c_Page_045.QC.jpg
14ff6faa2807b204a46f3185c083cae5
f1cbf3a10e8a2aaed8e3ebd8d1fbbcaaadff14b8
3961 F20101124_AABERZ rogers_c_Page_033thm.jpg
d6d11e9af550d83cbd5e4fdc52ecbde8
423518f73c39949c942b0cf3d4ad247f172e044f
24321 F20101124_AABDOX rogers_c_Page_096.QC.jpg
d8558d02791ec2ffbeb96cf729531161
9d9f4cd29e57556c88666abd3749e93a4285ea44
18586 F20101124_AABDQA rogers_c_Page_079.QC.jpg
67864f677724f15c67d872952db4777b
b1ed9c94c4b5e8b18f1bbf7f92d5ab666f83703e
F20101124_AABDPM rogers_c_Page_029.tif
07b1c921dccc0c3c01f0750ec7f9f997
f1d775a0617317da45242fb31d6f8d896b0d246b
6102 F20101124_AABETD rogers_c_Page_055thm.jpg
a8a62a7abe55297a40377f9e64476aac
64b3129f1dcfd9dbb60f719af2c7010d30e69188
4650 F20101124_AABESP rogers_c_Page_045thm.jpg
157e1b01bb0f661d3d433370125f8b26
1c1950411ec73a98c7b835952c0bbb36e1e714ad
244 F20101124_AABDOY rogers_c_Page_054.txt
ac77dc2b94ada52b6a497961d4e06ed7
1d084c1f2f6ba65b32d37fc4e18dba7489fe9c78
81433 F20101124_AABDQB rogers_c_Page_006.jpg
3ff3c06220f118d224f04eae43fd1acb
17fae8e4c4d210613e1aebe9d21b133c3297b287
440963 F20101124_AABDPN rogers_c_Page_054.jp2
b0cfe60a8db307c123a93308c269fad1
9ff9d327b9d62dc4ec101ae448805abc18adc0df
14591 F20101124_AABETE rogers_c_Page_056.QC.jpg
4c7ada8fc570728a151ce031422c065d
4408c2168bab2c3a634e528bee8ba92527a0b3ee
5415 F20101124_AABESQ rogers_c_Page_046thm.jpg
eb55e18add3b24f25ea845ad7a2f881c
44acb7d2b513694a0fcb3eaaefcac9e06b98cbbf
66131 F20101124_AABDOZ rogers_c_Page_036.jpg
d44bff8d7301aef0f5bac8ccfd1dd42a
d3eb0afa57b4c1da9dc9bff6957835a6269bf50d
858311 F20101124_AABDQC rogers_c_Page_036.jp2
2de9ace77cd856632c1b45cbe44d21a0
db6e9b1cdaa55aebe6e9f8b876ae0986623c8efb
84114 F20101124_AABDPO rogers_c_Page_126.jpg
2ea8b737901dab42586cf1b7da205ab3
ba8ecd00f9b6eb3c15f06280f4796a8c6414df33
8839 F20101124_AABETF rogers_c_Page_057.QC.jpg
f68f2bd72fd944f8c5668ece71c3baa3
985970650f247fb9a8d2f45b9ac19f9605c0bff5
6752 F20101124_AABESR rogers_c_Page_047thm.jpg
048312d6056f1574eef07c4298b2bffd
6266da7417d2cc0607308ce597f099df4d7026fa
2578 F20101124_AABDQD rogers_c_Page_116.pro
81b7424a1e22ea69b051d54d816c2993
148c7230d27c80a06e40244fa8920f72639bc7a5
7027 F20101124_AABDPP rogers_c_Page_058thm.jpg
a39956e5e56733d130aa3eea6f4eb189
f1c7ebf25e12955b49bd9bf73e74225dbf4b7d2a
3084 F20101124_AABETG rogers_c_Page_057thm.jpg
7d3e7b8afaf6d92fda0fa0dafdef498b
65395da31597eeac1e45b2d8ff589041bcced780
13008 F20101124_AABESS rogers_c_Page_048.QC.jpg
5ce65de0cb3304b236745d28699a4cf4
828ba2a67b296ecc332e4707de808da72c9245cf
328 F20101124_AABDQE rogers_c_Page_103.txt
bb4d29fa584394a17d53baf066ba5095
d70b644e7d325927b03374b19f73a08cdbf34abe
82105 F20101124_AABDPQ rogers_c_Page_088.jpg
baae849184cdb707253669da16119420
0ddd5fbe73921b5c190c3170dec260dbb3adc6e3
28394 F20101124_AABETH rogers_c_Page_058.QC.jpg
36c0c12ad00c61fd448e0a3fa4e033ac
3f6d83a5089f59a99efd791cf39387198247f4cc
26180 F20101124_AABEST rogers_c_Page_049.QC.jpg
698c5299d47a07e900c94c6765982154
684556f42e67a43cf08a725cca8e673bfc32f47c
21103 F20101124_AABDQF rogers_c_Page_123.jpg
e730f3ba4c5db2ef2ab926910aa89e5e
5c8aec34751f4c01dc3a2c190ab7f033981584a3
F20101124_AABDPR rogers_c_Page_001thm.jpg
ae4bcaec308128e3de8305431c13924c
6b4579c1adf310dfdb75a1fc02bb534db3887394
21083 F20101124_AABETI rogers_c_Page_059.QC.jpg
b14a940b411bb0469e4597f7b929b0b3
26c2d42cd4913d1f85e06a962d43088d8846d98b
6544 F20101124_AABESU rogers_c_Page_049thm.jpg
d8d5db30a5b82b25e21ecaa11336d8d9
f3dbb847d0ec066ea2659ea7694a30c62a849590
85655 F20101124_AABDQG rogers_c_Page_024.jpg
c24f6c472ba1606b9947dd22adc2d3d8
ef19816a291cd16f02c4981e39b337336bf5b58d
5692 F20101124_AABDPS rogers_c_Page_082thm.jpg
d076c60e52088a8e39e85903b082b2c5
cb8102a8cfef40e2457dab039498076e061f1bfe
3249 F20101124_AABETJ rogers_c_Page_060thm.jpg
5d22f58bcd11fe01c7ebcf79e284b5c3
bae8255bcf15c464322042e570b862ec8e88f2c6
F20101124_AABDQH rogers_c_Page_092.tif
140e9547404592106427adec6033f958
c079f5f5bbc36ea58de43d7540c5cfe6afcf5200
25390 F20101124_AABETK rogers_c_Page_061.QC.jpg
4bbfb6171050ad71fa5bebe7cf8d142c
24b856ad289fba6d2149a06e32202bad4d635b44
22268 F20101124_AABESV rogers_c_Page_050.QC.jpg
414339d1ccd7295a0d59a56150a2a444
dff5706751d6e51e43cdd00ca56b13f87c24ccc2
27698 F20101124_AABDQI rogers_c_Page_030.QC.jpg
6bdcc3da811ebb217ab0b23a6cbe0589
9374c548dc918c2c14035dc9a89e24829bf15f64
5099 F20101124_AABDPT rogers_c_Page_107thm.jpg
021dff5e0e36616e6bd344b4aaa02f6d
5ca8970ec9fbeb2d09593ed191f3dffb9b53aa86
14726 F20101124_AABETL rogers_c_Page_062.QC.jpg
a7b043b44678ef836b6bf9be7ebe7872
d8a17c29d590d991cee39efed0bd95aefe9035cf
24582 F20101124_AABESW rogers_c_Page_051.QC.jpg
b091f4496c45d7208358e987877a548b
0c2d723b9f5246106148b79d505a85e715915867
23610 F20101124_AABDQJ rogers_c_Page_064.QC.jpg
ea4b8e57e879655765001a7067a61f46
c00999e9eb2c5ef7fc9a3fcced199d98c3616d42
F20101124_AABDPU rogers_c_Page_048.tif
47794bc0e5d25b3f058868d52b8f9baf
a61e5aa68a75fde6ce294caf1e5cb866f75baf95
6456 F20101124_AABEUA rogers_c_Page_076thm.jpg
187008932d3b5ddbb7830bc1df4a726e
66baa156e6a354649b6e05baf68501fb85a06986
3379 F20101124_AABETM rogers_c_Page_063thm.jpg
539ffb8f2d6b0eb4b85aa502500309bd
74ef6cbe348f16bfc5f4f6fcd605fed6d140fe33
6746 F20101124_AABESX rogers_c_Page_051thm.jpg
ec0045db3b3ffdcb56e68d5119a242de
459d38b2328884300deb232f71d5f7d08fecb7fa
871 F20101124_AABDQK rogers_c_Page_081.txt
8f762e280f5a6bf3bd3587e6c537eeb6
fcf8d2606c48e774bdf88c84dd68d95f4614b996
29763 F20101124_AABDPV rogers_c_Page_059.pro
093fe0d81d6d271b2692dc6ce94a02cd
4a24c68a295a400cc40bc1261e73ee0db6ff69a7
5269 F20101124_AABEUB rogers_c_Page_079thm.jpg
29e521a993c327a7863df15a15864d5f
becada30bc485fbaf81f68bc3d6145e9cf49ae65
6276 F20101124_AABETN rogers_c_Page_064thm.jpg
697aa3901bb777de968865dcf7e654f0
db57042ffef419a73ad265ed05d3d5f93c48acdb
25736 F20101124_AABESY rogers_c_Page_052.QC.jpg
2d75f880bbbd7621e1330da609554dd6
2ef4c55d0623ea7eb32459f1e829fe8e0486db31
F20101124_AABDQL rogers_c_Page_035.tif
05a3254e596e7933c3de295f4582deae
32679639f2d0ba49f44588e90465047c050c3479
F20101124_AABDPW rogers_c_Page_049.tif
206e5b668df0c4cd93b18f692222d3d4
b52637c57d30147b4682971e2cc4ac7f99a4c02a
21057 F20101124_AABEUC rogers_c_Page_080.QC.jpg
4fe35bbd415ea573bad696df7f6f214d
685ec3261efa152d14f0e8f30a1f45108ce9260e
4862 F20101124_AABETO rogers_c_Page_066thm.jpg
d4f5b3fef8340d9d8219d4e9846c540d
4629bc6d28db11da98fbdbf1bcbdd986d77965fc
5563 F20101124_AABESZ rogers_c_Page_053thm.jpg
9725c89e93941b99c6b74e07f8f98ab7
78878e8184a8bc28bb4f826a28265f679ce9bda8
75852 F20101124_AABDRA rogers_c_Page_012.jpg
d13ca044dd8e9bdabc8250d6dfce2487
2bd4bd05b00bd953e5b8b6ca831f63d22021ca61
15592 F20101124_AABDQM rogers_c_Page_053.pro
6da90364f317806586cdbda839e224ed
dcd2720ffc463aec681d6ebc7099327542377084
5992 F20101124_AABDPX rogers_c_Page_122thm.jpg
fcb72e0ab4d8702f48b2960f289b0756
a31cb758a7824039d112fdaeb162ff4a76ace169
5679 F20101124_AABEUD rogers_c_Page_080thm.jpg
0b0df653434df2d18d08bed642aeebe0
0b5caa1c738ec0184b8ab11740ba830825e5c60b
14863 F20101124_AABETP rogers_c_Page_067.QC.jpg
247b87807ea5f2eda4d29c4eed316f2e
f819020b75e3b9a44b2006318f43667ba57530d2
70758 F20101124_AABDRB rogers_c_Page_013.jpg
6ef3b89816959b33743327cff409dd9d
9866f4c1d3a982880dddaeaae47a9a46595352a9
54657 F20101124_AABDQN rogers_c_Page_027.pro
86d26444765cac288ceb865baa68eca1
bf11e1c1742cb795cfce35e4f7bb8f2bf4b6eb8e
29494 F20101124_AABDPY rogers_c_Page_128.QC.jpg
3587e5fc110a613d9aa00970560e04de
c9a62bc4156bacf9dfb15d2f88f254bcfc34a9e0
13825 F20101124_AABEUE rogers_c_Page_081.QC.jpg
ced856a6765566da5688985df97be715
71d69d445ffb7ed6b0349e7b4fc6ce1d9c2bdf58
26019 F20101124_AABETQ rogers_c_Page_068.QC.jpg
2428a467e521a9adac590b6e4c077917
eaf9c275d9fdae38e25d5bda96bef18a477bb95b
84925 F20101124_AABDRC rogers_c_Page_014.jpg
514511f28931c135a1ab0e87ec040b74
b805e4e89769982374b31ac249b601488eb05065
F20101124_AABDQO rogers_c_Page_094.tif
fb0215f81c7fcdd787f18d647cb88d57
96e1899c973e96b441ef78ded052d9d48652c77e
6625 F20101124_AABDPZ rogers_c_Page_038thm.jpg
db6f6dcd227a1db725b0de23e5a5362d
969f79a7e5c31b4a4719a624ed7ca54bbe43d494
4414 F20101124_AABEUF rogers_c_Page_081thm.jpg
e5589649a7bb0b397c22e4b86ee3dd4e
e6a120aec2d74987579423b8e41085ed13d67041
13141 F20101124_AABETR rogers_c_Page_069.QC.jpg
b350c3a80fd57f4ebcfa5efb591f4a33
b3df9e92c5af0f4d28ae034c96fb99c059ae2027
89668 F20101124_AABDRD rogers_c_Page_015.jpg
d0b3d561d8ac8ea7c3772b61212311e4
de80fb0b58cd3258d2a77cc5de7946e0f1bbda7d
151621 F20101124_AABDQP UFE0019615_00001.mets
d91bfc9cce287c16d924ca1dceb2ca75
7b00743bd44805a5210c28e03eec8199efa5e8d1
19558 F20101124_AABEUG rogers_c_Page_082.QC.jpg
c719a14192bce4db95c45789a6aa80b0
21b6ba754fa7873da7647969eb98f37b8fd6a191
4818 F20101124_AABETS rogers_c_Page_069thm.jpg
499074e4964d9734277ca030bd582218
de2e6d1f3bafa75ce9a23eb5c694358146f6e9c4
51005 F20101124_AABDRE rogers_c_Page_016.jpg
80c0e2e28df97f65388183fc402d2919
ae951cfd2198b8d572473e174011fd513668d683



PAGE 1

1

PAGE 2

2

PAGE 3

3

PAGE 4

TheguidancefrommyundergraduateassociateDeanDr.MerckelandProfessorDr.KhoiehelpedmepursueaPh.D.intherstplace.EncouragementfrommyanceXiao,myadvisorDr.Harris,andmyparents,andfriendsKwansunandDuhavehelpedmestaythecourseevenwhenIdoubtedwantingtonishmyPhD.ForthatIamforevergrateful.IthankmyadvisorDr.JohnG.Harrisfortheopportunitytoworkunderhimandlearnsomuchaboutresearchandacademiaingeneralandspecicallybioinspiredcircuitry.IamespeciallygratefulforhisencouragementtotaketheinternshipatTexasInstrumentswhereImeetmyanceandhisguidanceinmyresearchaswellaslifesuchasencouragementtonishmydegreebeforegettingmarried.EventhoughIdidn'talwayslistensowell,Iatleastgotpartofthemessage.IappreciateDr.Harris'understandingand,whileattimespushingtoohard,notgivinguponmeevenwhenIwasdistractedfrommyresearchandmyproductivitywasverylacking.Dr.Harris'experienceandknowledgefrommanyyearsasaprofessorfacilitatedmystudiesduringmyPhD.MycommitteeDr.Harris,Dr.Principe,Dr.Bashiruulah,Dr.Sanchez,andDr.DingaswellasDr.Fox,helpedmethoughoutmyresearchbyprovidingvaluablefeedback.Theirinsights,comments,andquestionstoponderwereveryvaluabletoshapingthisdissertationcontent.Theircombinedexpertiseacrossareasallowedeachareaofmyresearchtobecarefullyexamined.IwanttothankDr.SanchezandhisNeuroprostheticsResearchGroupforalloftheirhardworkinobtainingtheneuralrecordingsfromtrainingtheratsandperformingtheimplantationsurgery,tosettinguptheentireexperimentalrecordingapparatus.Dr.Sanchezalsomarkedthegroundtruthsusedtotestourspikedetectionandfeatureextractionmethods.JackDiGiovanna,Dr.Sanchez'sPh.D.student,assistedwiththeTucker-DavisTechnologies(TDT)codingforthechiptestsetup.Manylabmateshaveparticipatedindiscussionaboutmyresearchandprovidedhelpinginsightand/orquestionsbutmostparticularDr.DuChenwhograduatedlast 4

PAGE 5

5

PAGE 6

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 9 LISTOFFIGURES .................................... 10 ABSTRACT ........................................ 13 CHAPTER 1INTRODUCTION .................................. 14 1.1NeuralSignalProperties ............................ 15 1.2ExtracellularNeuralRecordingSystemOverviewandConstraints ..... 19 1.2.1Electrodes ................................ 21 1.2.2Amplier ................................. 22 1.2.3WirelessDataTransmission ...................... 24 1.2.4SpikeSorting .............................. 24 1.3NeuralDataReduction ............................. 27 1.3.1DataReductionforSpikeSorting ................... 28 1.3.2DataReductionwithSpikeDetection ................. 29 1.4UniversityofFlorida'sNeuralRecordingBandwidthReductionStrategies 32 1.4.1BiphasicSignalCodingwithReconstruction ............. 32 1.4.2Pulse-BasedFeatureExtraction .................... 36 1.4.3SpikeDetection ............................. 38 1.5DissertationStructure ............................. 39 2PULSE-BASEDFEATUREEXTRACTIONANDSPIKESORTING:IMPLEMENTATION1BACK-ENDSOFTWARE ............................. 40 2.1Pulse-BasedFeatureExtraction ........................ 40 2.2DataReductionwithPulse-BasedFeatureExtractor ............ 41 2.3SpikeSortingwithPulseTrains ........................ 42 2.4MatlabSimulationsResults .......................... 47 2.4.1Data ................................... 47 2.4.1.1Neurosimulatordata ..................... 47 2.4.1.2Caltechsimulateddata .................... 49 2.4.1.3Ratdata ............................ 49 2.4.2Spike2 .................................. 51 2.4.3SpikeSortingResults:NeurosimulatorData ............. 52 2.4.4SpikeSortingResults:CaltechSimulatedData ............ 52 2.4.5RatDataSpikeSortingResults .................... 58 2.4.6FutureWork ............................... 68 6

PAGE 7

................. 71 3.1BandwidthParameters ............................. 71 3.2MatlabSimulationsResults .......................... 72 3.2.1SpikeSortingResults:NeuralSimulatorData ............ 72 3.2.2FutureWork ............................... 76 4PULSE-BASEDFEATUREEXTRACTIONANDSPIKESORTING:IMPLEMENTATION3HYBRID ...................................... 78 4.1CircuitDesign .................................. 78 4.1.1Circuitry ................................. 78 4.1.1.1Voltagetocurrentconvertercircuit ............. 79 4.1.1.2Comparatorcircuit ...................... 79 4.1.1.3Resetandrefractoryperiodcircuit ............. 80 4.1.1.4Leakycircuit ......................... 81 4.1.1.5Chipspecics ......................... 82 4.1.2TestSetup ................................ 82 4.2ChipResults ................................... 85 4.2.1NeuralSimulator ............................ 86 4.2.2InVivowithRat ............................ 89 4.2.3FutureWork ............................... 89 5SINGLE-SCALESPIKEDETECTOR ....................... 91 5.1Algorithm .................................... 91 5.2MatlabSimulations ............................... 91 5.2.1Data ................................... 91 5.2.2ReceiverOperatingCharacteristics(ROC)Curves .......... 92 5.3CircuitDesign .................................. 96 5.4ChipResults ................................... 97 6MULTI-SCALESPIKEDETECTOR ........................ 104 6.1OptimalThreshold ............................... 104 6.2Algorithm .................................... 104 6.3MatlabSimulations ............................... 105 6.3.1ScaleCombinationMethod ....................... 105 6.3.2ThresholdScalingfromOneScaletoOthers ............. 107 6.3.3Data ................................... 110 6.3.4ReceiverOperatingCharacteristics(ROC)Curves .......... 112 6.4CircuitDesign .................................. 114 6.5ChipResults ................................... 117 7

PAGE 8

................................... 118 7.1OverallConclusions ............................... 118 7.2ContributionSummary ............................. 119 APPENDIX APROOF:THEDIFFERENCEOFTHEMULTI-SCALEGAMMAFILTER'SADJACENTTAPSIMPLEMENTACONTINUOUSWAVELETDECOMPOSITION 121 BLEAKYINTEGRATE-AND-FIRE(LIF)T69K-ASCHIPPINOUT ....... 124 REFERENCES ....................................... 126 BIOGRAPHICALSKETCH ................................ 131 8

PAGE 9

Table page 2-1Featureextractorsmisclassications. ........................ 57 2-2FeatureextractorsmisclassicationscomparedtoSpike2. ............. 64 3-1Spikesortingperformancepercenterror. ...................... 75 3-2Bandwidthreductionsortingerrorcomparison ................... 76 4-1Spikesortingperformance(percenterror)fromleakyintegrate-and-re(LIF)featureextractionchip. ................................ 89 4-2Bandwidth(pulses/s)fromLIFfeatureextractionchip. .............. 89 4-3Bandwidthreduction,powerconsumption,andsortingerrorcomparison .... 90 9

PAGE 10

Figure page 1-1Typicalextracellularspikewaveformwithhighsignaltonoiseration(SNR). .. 16 1-2Sketchofaneuronwiththepartslabelled. ..................... 16 1-3Waveformsrecordedfromalinearsiliconhexatrodefromapyramidalcellwithhypothesizedpositionofthehexatrodealongthesomatodendriticaxis. ..... 17 1-4Blockdiagramofwirelessfront-endneuralrecordingsystem. ........... 19 1-5Blockdiagramforfourdegreesofdatareduction. ................. 21 1-6UFoverallneuraldatareductionapproaches. ................... 33 1-7Anexampleofaninputsignalandit'sbiphasicrepresentation. .......... 35 1-8Blockdiagramofbiphasicencodingwithintegrate-and-re(IF)neuron. ..... 36 1-9Blockdiagramofbiphasicencodingwithleakyintegrate-and-re(LIF)neuron. 37 1-10Featureextractionandsubsequentsortingimplementationschemes. ....... 38 2-1Blockdiagramofbiphasicencoding. ........................ 42 2-2Neuralsimulatorspikesignaturesfromsixdierentneuronsattwodierenttimeperiods. ........................................ 44 2-3NeuralsimulatorspikesignaturesconvolvedwithaGaussiantodeterminethedistancebetweenitandthetemplates. ....................... 45 2-4SpikesortingerrorasafunctionofthepulsetraindistanceGaussian. ..... 46 2-5Neuralsimulatorsignalwithallsixneuronsandthebiphasicpulsetrainoutputfromtheleakyintegrate-and-re(LIF)circuitforabandwidthof455pulses/s 48 2-6AsegmentoftheCaltechdataset. .......................... 50 2-7Neuralwaveformrecordedfromrat003.Columntwoiszoomedinfromcolumnone. .......................................... 51 2-8Spike2'stemplatesforneurosimulatordata. .................... 53 2-9Spike2'sprincipalcomponentanalysis(PCA)forneurosimulatordata. ...... 54 2-10ActualclassiedspikesfortheCaltechsimulateddata. .............. 55 2-11Spike2'stemplatesfortheCaltechsimulateddata. ................. 56 2-12Spike2'sprincipalcomponentanalysisfortheCaltechsimulateddata. ...... 56 10

PAGE 11

........... 57 2-14Spike2exampletemplateforratdata. ........................ 59 2-15Spike2'sPCAanalysisforratdata. ......................... 60 2-16Spike2'shistogramanalysisforratdata. ...................... 61 2-17Spike2'stemplates. .................................. 62 2-18Featureextractortemplates. ............................. 63 2-19PileplotsofSpike2'ssortedspikes. ......................... 65 2-20Pileplotoffeatureextractor'scorrectlysortedspikesreferencedtoSpike2'sresults. 66 2-21Pileplotoffeatureextractor'scorrectlysortedspikeswiththosemisclassiedasthatneuronoverlaidwithadashedblackline. ................... 67 2-22Pileplotoffeatureextractor'scorrectlyclassiedspikesoverlaidwiththespikesfromthatclassbutmisclassiedasanotherclassinthatclassescolor ...... 69 3-1Bandwidth(pulses/s)changesforthresholdandleakagevaluesparametersandintegrationcapacitorat10pF. ............................ 72 3-2Spikesortingerrorchangesforthresholdandleakagevaluesparameters ..... 73 3-3SpikesortingerrorasafunctionofbandwidthforthreeSNRs. .......... 74 4-1Leakyintegrate-and-re(LIF)circuit. ....................... 79 4-2VoltagetocurrentconvertorcircuitforLIF. .................... 80 4-3OperationaltransconductanceamplierOTAforvoltagetocurrentconvertorcircuitforLIF. .................................... 81 4-4ComparatorcircuitforLIF. ............................. 82 4-5ResetandrefractoryperiodcircuitforLIF. ..................... 83 4-6LeakycircuitforLIFimplementedasaGmcurrentsource. ............ 84 4-7LayoutfortheLIFcircuit. .............................. 85 4-8Initialchiptestsetup:CompareUFsfeatureextractorwithDr.SanchezsSpike2resultsusingneuralsimulator(487NEB). ..................... 86 4-9Intermediatechiptestsetup:SetUFsfeatureextractorchipparametersusingprerecordeddatafromratthatwillbeusedininvivoexperiments(487NEB). 87 11

PAGE 12

............................. 88 5-1Single-scaledetectorblockdiagramwithsnapshotsofthewaveformaftereachblock. ......................................... 92 5-2Datausedforsimulations. .............................. 93 5-3Plotofreceiveroperatingcharacteristic(ROC)curves,0dBSNR. ........ 94 5-4Single-scalespikedetectorcircuitdiagram. ..................... 97 5-5Layoutforonsetspikedetectorchip. ........................ 98 5-6Singe-scalespikedetectorchipresultswithsignalgeneratorpulsewaveformasinput. ......................................... 100 5-7Single-scalespikedetectorchipresults. ....................... 102 5-8Single-scalespikedetectorchipresults. ....................... 103 6-1SNRversusspikewidth. ............................... 105 6-2Blockdiagramofthewaveletalgorithm. ...................... 106 6-3Stepresponseofbandpasslters. .......................... 107 6-4Filteredscalesanddetectedspikesoneachscaleandcombinedoutput. ..... 108 6-5IllustrationofBayes'detectorprinciples. ...................... 110 6-6Figure 6-4 withthethreshold,shownwithadashedhorizontalline,setforplotB)andautomaticallysetfortheremainingbands. ................. 111 6-7Datasetusedforsimulations. ............................ 113 6-8PlotofROCcurves,0dBSNR. ........................... 114 6-9Multi-scaleGammaltercircuit. .......................... 115 6-10Multi-scalespikedetectorchiplayout. ....................... 116 A-1Multi-scaleGammaltercircuit. .......................... 121 A-2Cascadedlterstructureforcontinuouswavelettransform(CWT)decomposition(TypeII). ....................................... 122 12

PAGE 13

Thepurposeofthisdissertationistoinvestigateanultra-lowpowerimplantforspikedetectionandspikefeatureextractioninneuralrecordingapplicationstodramaticallyreducetherequiredcommunicationbandwidththroughtheskin.Implantedsystemsimposefourmajorconstraints:lowpowerconsumption,smallsize,robustness,andlimitedbandwidth.Thedevelopedsolutionistwofold.Forapplicationswhichdonotrequirespikesortingalowerpowerandlowerbandwidthsolutionexists.Anovelmulti-scalecontinuouswaveletapproachisusedtodecomposethesignalintoseveralfrequencybandstoallowforindividualthresholdsateachbandtomoreaccuratelydetectthepresenceofaspike.Forapplicationswhichrequirespikesorting,aspikefeatureextractionalgorithmwasdevelopedtoextractinformationaboutthespikessobandwidthisnotwastedtransmittinginformationnotrelevanttospikesorting.Thefeatureextractor'sbandwidthreductionwasdesignedwithasystemlevelviewtooptimizetheback-endspikesortingwhileusingminimalbandwidth.Analogverylargescaleintegration(VLSI)circuitrywaschosentoimplementboththespikedetectionalgorithmandfeatureextractionalgorithmtoallowforanultra-lowpowerandcompactsolutionfortheintegrationofmanychannelsinanimplanteddevice.Preliminarytheoreticalanalysisandchipmeasurementresultsshowsuitabilityforinvivoneuralrecordingapplicationsforbothalgorithms. 13

PAGE 14

Theneuronisthebasicinformationprocessingunitinthebrain.Neuronsuseelectricalpulses,calledactionpotentialsorspikes,totransmitinformation.Theirextracellularelectricalpulsescanberecordedusingmicroelectrodesthatareimplantedintothebrain.Theserecordingsarecalledneuralrecordingssincetheyrecordthevoltagepotentialcausedbyneurons.Tobeststudyneuralinformationprocessing,manyneuronsmustbesimultaneouslyrecordedinawakebehavingsubjects.Sate-of-the-artrecordingsystemsrequiremicroelectrodearrayswithhundredsofelectrodesimplantedintothebrain. Whilemuchisstillunknownaboutthebrain,researchershavenowlearnedenoughtointegrateneuralprostheseswiththebrain[ 1 ].OneexampleofaneuroprostheticisaBrain-MachineInterface(BMI)[ 2 ].Motor-basedBMIsextractinformationfromneuralrecordingscollectedinthemotor,premotorandparietalcorticeswiththegoalofcreatingpredictivemodelsforthesubject'sintentofmotormovementtodirectlycontrolaroboticdevice.Eventually,thesedevicescouldallowparaplegicstocontrolaroboticarmtofeedthemselvesorturnthepagesofabook.Neuralprostheticsrequirelong-termneuralrecordingswhichnecessitatewirelesslytransmittingthedatafromtheelectrodesthroughtheskin.Ifawirefromtheelectrodepassedthroughtheskintosendthedata,infectionisriskedandthesubjecttethered.Also,havingmanywirescomingoutoftheheadandbeingtetheredrestrictsmovementofthesubject. Currentinstrumentationtechnologyandsurgicalproceduresallowforthesimultaneousrecordingofhundredsofelectrodes.Thebottleneckishowtotransferthelargebandwidthrawdatastreamswirelessly.Transmittingtherawvoltagesignalsfromhundredsofchannelsisnotpossiblewiththecurrentwirelessbandwidthlimits.Furthermore,evenifthesehighdataratescouldbemet,thepowerdissipationoftheelectroniccircuitrywouldseverelydraintheimplantedpowersupplyandexceedthepowerdissipationlimitsforpreventingtissuedamage. 14

PAGE 15

3 ].Theamplitudevariesinverselywiththedistancebetweentheelectrodeandtheneuron.Atypicalextracellularspikewaveformfromahighsignaltonoiseratio(SNR)neuralrecordingfromaratisshowninFigure 1-1 .Thefrequencycontentofspikesismainlybetween100Hzand6KHz[ 4 ]withwidthsvaryingbetween0.4msand3ms[ 5 6 ]dependingonhowfartheelectrodeisfromtheneuronandwhatpartoftheneuronisclosesttotheelectrode.AlabelleddrawingoftheneuronisshowninFigure 1-2 .Asthespikepropagatesfromthesoma(cellbody)alongtheaxon,thespikeamplitudebecomesattenuatedandthewidthincreases[ 7 ]asshowninFigure 1-3 15

PAGE 16

Typicalextracellularspikewaveformwithhighsignaltonoiseration(SNR). Figure1-2. Sketchofaneuronwiththepartslabelled. 16

PAGE 17

Waveformsrecordedfromalinearsiliconhexatrodefromapyramidalcellwithhypothesizedpositionofthehexatrodealongthesomatodendriticaxistotheright.NotevariationinwaveshapealongthesomatodenriticaxisadaptedfromHarriset.al.[ 7 ]. 17

PAGE 18

8 ].ItisdiculttodeneSNRforaneuralwaveformbecausethesignal(thespike)hasvaryingamplitudesandwidths.SpikeswithalargeramplitudeandwidthwillhaveahigherSNRthansmalleramplitudeandnarrowerspikesfromthesamechannelwiththesamelevelofnoise.ThisalonewouldsuggestSNRshouldbedenedforeachneuronwithinarecordedwaveform.However,thesignalisalsotransientandnon-stationarysotheamplitudeandwidthofspikesfromasingleneuronalsochangeovertime.Therefore,SNRisoftendenedusingtheaverageofallthespikesinthewaveformwithtypicalSNRrangesbetween0dBto12dB[ 9 ]. Singleneuronsdepletetheirchemicalreservoirsafterproducingaspike,inactivatingtheNachannelswhichcannotreopenuntilthemembranepotentialreturnstoanegativevaluenearthethreshold.Thissetsaminimumtimerequiredtoreplenishtheirreservoirsbeforetheneuroncanspikeagain.Thisperiodiscalledtherefractoryperiodandistypicallyaround1ms[ 10 ].Becauseextracellularneuralrecordingsmaycontaintheresponseofmorethanoneneuron,spikesfromdierentneuronscouldreveryclosetogether,withintherefractoryperiodofasingleneuron.Spikesfromdierentneuronscanevenoverlapandcreateasuperimposedwaveform. Manyelectrodesareoftenusedtorecordneuronsfrommultiplesiteswithinthebrain.Acollectivegroupofimplantedelectrodesarereferredtoasanelectrodearrayandisencasedincranioplastthatcoatsascrewanchoredintotheskull.Thisphysicallykeepstheelectrodesfrommovinginreferencetotheskull.Thebrainhoweverisitselfoatingincerebrospinaluid(CSF)andcanshiftbymanymicronsovertime.Asthebrainmoves,thedistancebetweentheelectrodeandtheneuronsitisrecordingfromalsochangesothespikeshape(amplitudeandwidth)canchangeovertimeasthebrainshifts.ThisresultsinSNRuctuationsinthesignal.Tomakematterworse,unavoidableelectrochemical 18

PAGE 19

Blockdiagramofwirelessfront-endneuralrecordingsystem. eectsattheelectrode-tissueinterfaceintroduceDCosetsrangingfrom1{2Vacrosstherecordingsites[ 11 ]. 1-4 19

PAGE 20

Inadditiontotheelectrodesseveralpiecesofneuralinstrumentationelectronics,suchasanamplier,datareducer,andcircuitrytotransmitthedata,shouldalsobeimplantedtoavoidanywirespassingthroughtheskin.Lowpowercircuitryisnecessaryduetothedicultyofchargingorchangingimplantedbatteriesaswellastopreventtissuedamage.Powercanbebroadcastintothedevice,butstudieshaveshownthatifthebraintissueincreasesintemperatureanymorethanabout1Cthebraintissuewillbedamaged[ 12 ].Powerdissipationover80mW=cm2hasbeenreportedtocausegeneraltissuedamage[ 13 ].Circuitareamustalsobesmallduetothelimitedareaofimplantationbetweenthesubjects'skullandskinespeciallyonsmallanimals. BMIsystemsplacestrongconstraintsonthewirelesstransmissionbecausehundredsofchannelsarecurrentlyrecordedwiththedesiretoreachthousandsinthefuture.Transmittingrawvoltagesfrom100channelsata25KHzsamplingrateand8-bitsofresolutionwillgeneratedataratesaround20Mbps.Furthermore,evenifthesehighdataratescouldbemet,thepowerdissipationoftheelectroniccircuitrywouldseverelydraintheimplantedpowersupplyandexceedthepowerdissipationlimitsforpreventingtissuedamage.Therefore,datareductionisrequired.Becausespikesaresparsewithinneuralrecordings,onlytransmittinginformationaboutthespikesprovidesfurthersignicantreductionintherequiredtransmissionbandwidthcomparedtosamplingandquantizingtheentiresignal.Afterthespikesaredetected,therearethreemajor 20

PAGE 21

Blockdiagramforfourdegreesofdatareduction. optionsforbandwidthreductionthatdependontheapplication.Thisresultsinatotaloffourdatareductiontechniques.Intheorderoftheirdatareduction,withallbuttherstrequiringspikedetection,theapproachesare:Approach1:sampleandquantizetherawdatausingconventionaltechniques,Approach2:onlytransmitaclipofthewaveformaroundthespike,Approach3:sendfeaturesneededforspikesorting,Approach4:onlytransmitthespiketimes.ThefourdatareductionschemesaredepictedinFigure 1-5 1 ].Electrodematerial,size,tipsharpness,andpliabilityshouldbecarefullyselectedtominimizenoiseandtissuedamagewhilemaintainingtheabilityforpreciseimplantation.Thedetailsonspecictradeosinelectrodedesignisnotpresentedherebutonecanrefertotheliterature[ 8 14 15 ]. 21

PAGE 22

9 ].Theelectrodecontributiontothetotalnoiseofthesystemismainlycomprisedofneuralbackgroundnoiseandthermalnoise.Neuralbackgroundnoiseisthesummationofallthedistantneurons'electricalpotentials.Thethermalnoiseoccursatthemetal-electrolyteinterfaceandisrelatedtoelectrodeimpedanceandtherecordingbandwidth,whichhasa1/ffrequencydependence[ 15 ]. Therearetwomajorcategoriesofelectrodes:passiveandactive.Passiveelectrodesdonotcontainanyinterfacingelectroniccircuitryontheelectrodesubstrate[ 16 ]andareusuallymadeofmetal[ 8 17 18 ]orglass[ 16 ].Activeelectrodesincludeelectroniccircuitryonthesamesubstrateastherecordingelectrode[ 16 ].Theon-chipcircuitrycanminimizethenumberofleadsonthechipaswellasminimizetheleakageandnoiseassociatedwithsendingaverysmallsignal(V)overwires.In1975,WiseattheUniversityofMichiganwasthersttoproduceanactiveelectrode[ 19 ]andadvancesarestillbeingmade[ 16 20 ]. 4 ],shouldberemovedfromthesignaltoreducethenoise.Ifacleverschemeisnotusedtolteroutthelowfrequencynoise,alargeo-chipcapacitorwillberequiredmakingthecircuitrytoolargetoimplant. Jietal.havereportedonetypeofComplementarymetaloxidesemiconductor(CMOS)amplierwhichsharesthesamesubstratewithsiliconelectrodes[ 20 ].Theamplierprovidesamid-bandgainof51dBwithoutamplifyingtheDCfrequencycomponents.However,onemajorissuewiththisdesignisthegainvariabilityfromprobetoprobeor 22

PAGE 23

NajaandWisewereoneoftheearliestteamstoreducetherandomDCcomponentattheelectrode-electrolyteinterfacewithareverse-biaseddiodetoclamptheinputwiththehighresistanceofthejunctiondepletionregion[ 4 ].Jietal.andAkinetal.employedaninternalbandlimitingmethodbyusingdiode-capacitorlterstoformthelowcut-ofrequency[ 20 21 ].Thisschemesuersfromseveralissues:opticaldriftthatreducesreliability,limiteddynamicrange,andhighvariabilityofthelowercut-ofrequency[ 22 ].Dagtekinetal.reportedamulti-channelchopper-modulatedneuralrecordingamplierthatusesthechoppermodulationtechniqueandanunbiasedlocationinthesystemasareferencetominimizetheeectsoftheDCdriftoftheneuralsignals[ 23 ].Nochipperformancemetricswereeverpublished,onlysimulateddata. Chandranetal.employasub-thresholdNMOStransistorasahighvalueshuntresistortoattenuatetheDCosettostabilizetheDCbaseline[ 24 ].Thisresistorcreatesthelowercut-ofrequencyoftheamplier.Upto400mVofDCinputcanbehandledwithoutsacricingtheACperformance.However,theamplierisunabletorejectnegativeDCinputvalues.Mohesenietal.modiedChandran'sdesignbyemployingasubthresholdPMOStransistortorejectbothnegativeDCandpositiveDCinput[ 22 ].Alasertrimmedresistorisusedtoaccuratelysetthelowercut-ofrequency.ThisdesigncannottolerateaDCshifthigherthan400mV.Thedrawbackistheextraprocesssteprequiredforlasertrimming. ChenandHarrisfromtheUniversityofFloridahaveusedclevercircuitrybasedonHarrison'sdesign[ 11 ]todevelopalow-noiseultra-lowpowerfullyintegratedneuralamplier(bioamplier)tomeettherequirementsofanimplantabledevice.TheUFbioamplierprovidesagainofalmost40dB,inputreferrednoiseofonly9:56Vrms,aCMRRofabout59dB,andapowersupplyrejectionration(PSRR)ofabout45dB[ 25 ].Ithasalowcutofrequencyof0.3Hzandahighcutofrequencyof5.4KHz.These 23

PAGE 24

21 ]reportasystemwherethedigitizedsignalisencodedintoan8-bitpulse-positionmodulation(PPM),pulse-codemodulation(PCM)andtransmittedwithAM.Thesystemdissipates2.8mWofpoweranditsdieareais0:7mm2perchannel.Huangetal.[ 26 ]usea9-bitPPMtoencodethedataandtransmititusingFM.Thecoreareais3:6mm4:350mmforonechannel.Bothofthesesystemsutilizeinductivelycoupledradiofrequency(RF)telemetryforboththepowerandthedatatransfer.Theanalog-to-digitalconvertors(ADCs)andmodulatorsexpandthedieareaandpowerconsumptionconsiderablysotheymustbecarefullydesignedtoallowsmallareaandlowpowerconsumptionnecessaryforimplantation. Neuroscientistsrelyonavarietyofspikesortingmethodsutilizingdierentfeaturesofthespikeswithnocommunitywideagreementastowhichspikesorteristhebest.Mostresearcherssimplyusethedefaultspikesortingsoftwarethatcomeswiththeneuralinstrumentationhardware.Tradeosexistbetweenperformanceandtheabilitytorun 24

PAGE 25

27 ]. Thereareseveralmajorcategoriesofspikesorters:templatematching,clusteringapproaches,independentcomponentanalysis(ICA),neuralnetworks,andsimplethresholdbasedmethods.Anoverviewofeachgroupofspikesorterswithproperreferencestotheliteraturefollow.Also,somegroupsusecombinationsoftheabovemethodstorenetheclassesthroughtheuseofdierentfeaturesinaneorttoimprovespikesortingperformance. Templatematchingdoesnotrequireanyfeatureextractionasanaveragewaveshapeforthespikesfromeachneuronisused.Aneuroscientistdeterminethenumberofdistinctneuronsaswellasthetemplates.Toolssuchasprincipalcomponentanalysis(PCA)andhistogramscanbeusedtoseeexaminethetemplates.Plottingtherst2-3principlecomponentsshowshowmuchseparationexistsbetweenclassesandgivestheuseravisualaidtoseehowmanydistinctclassesexist.Histogramsshowiftheringrateofeachneuronviolatesfundamentallimitsmeaningthatclasscontainspikesfrommorethanoneneuronandthetemplatesneedtobereformed.Oftentemplatematchingisperformedbyusingamatchedltertondwhichtemplatemostcloselymatcheseachspiketoclassifyit.Athresholdisusedsoawaveshapeisclassiedasnoiseifitdoesnotcloselymatchanyofthetemplates. Whenfeatureextractionisuseditisoftenfollowedbyclusterbasedspikesortingtodeterminetheneuronforeachspike.Byplottingeachfeatureonitsownaxisamulti-dimensionalgraphisformedwhereanyclusteringalgorithmcanbeusedtoclassifythedata.Popularfeaturesarethespikeamplitudeandwidth,principalcomponents,waveletcoecients[ 28 ],andslopeofriseandfalltime.Themostpopularfeature 25

PAGE 26

29 ]becauseitcanbeusedtoextractacompactsetoforthogonalfeatures.Somepopularoptionsforclusteringarethesimplek-meansornearestneighborwhereeachclusterlocationismarkedasthemeanofthedatawithinthecluster.AspikeisthenclassiedtotheclusterwiththeclosestmeanEuclideandistance.Moreelaboratemethods,suchasBayesianclustering,usestatisticalinformationabouttheneuronsandtheirspikeshapesandarebestsuitedwhentheclustershavesignicantoverlapordierfromasphericaldistribution.ManymoreclusteringalgorithmsexistandaresurveyedinbookssuchastheclassiconebyDuda,HartandStork[ 30 ]. ICAisaspecialcaseofblindsourceseparation.Itseparatesamultivariatesignalintoadditivesubcomponents.Itassumesthenumberofelectrodesequalsthenumberofsourceswhichonlyapproachesthetruthforalargenumberofelectrodes.Italsoassumessourcesaremixedlinearly.ICAisusedintetroderecordingasusuallythenumberofsignalsisclosertothenumberofneuronscomparedtosinglesiteelectroderecordings[ 31 ]. Somegroupshaveappliedneuralnetworkstosolvethefeatureextractionproblembutanothermethodmustbeusedtosortthedatatoprovidethegroundtruthstotraintheneuralnetwork[ 32 ].Thus,theneuralnetworkbasedspikesortercanonlybeasgoodasthespikesorterusedtoprovidegroundtruthstotraintheneuralnetwork. Threshold-basedspikesortersarethesimplestspikesortersandoftenprecedemorecomplicatedspikesortersasaspikedetector.Thethresholdscanbesimplevoltagethresholds(positiveandnegative)alongwithsomerulessuchasthewaveformmustpassthroughtwothresholdswithinacertaintime(ahoop).Thisthenformsavoltage-timethreshold.Onemightalsoimposethatthespikemustrstgopositiveandthennegative.Todierentiatebetweenspikes,dierentthresholdrulesareappliedtodierentiatespikesbytheiramplitude,width,and/orriseorfalltime[ 27 ].Someon-linespikesorters,suchasTucker-DavisTechnolgies'(TDT's)SortSpike2[ 33 ],usethisthreshold-basedspikesorterasaspikedetectionsteptoseparatespikewaveformsfromtherawdata. 26

PAGE 27

27 ]. Allofthementionedspikesortersrequirehuman-tunedparameterswhichaecttheaccuracyofspikesorting.Generallylessthanthreefeaturesareusedtospikesortsoneuroscientistscanviewthefeatureclustersandproperlysettheparameters.Afewattemptsatfullyautomaticspikesortinghavebeenmadebutneuroscientistshavenotembracedthemastheyincreasesortingerror.Thehumantunedparametersintroduceavariabilityinspikesortingacrossdierentneuroscientists. Astudyonthevariabilityofmanualspikesortingusinghuman-conguredon-linesortingalgorithmsbyWood,Black,Vargas-Irwin,Fellows,andDonoghue[ 34 ]showedawidevariabilityinthenumberofneuronsandspikedetectedinrealdata.Thenumberofspikesvariedfourfoldandthenumberofneuronswasonlycorrect25%ofthetime.Toobtainspecicerrorvalues,syntheticdatawasusedsothegroundtruthswereknown.Averageerrorratesof23%falsepositivesand30%falsenegativeswereobtainedwiththesyntheticdata.Thisvarianceanderrorincurrentspikesortingmethodsmakesitdiculttocomparespikesorterswithrealdataasaccurategroundtruthsarenotknown.Thebestapproachcurrentlyavailableisformanyexpertspikesorterstouseo-linespikesorterstomarkadatasetandanaveragetaken. 27

PAGE 28

BMIsystemscurrentlyuseahumantunedandcomputationallyintensivespikesortingprocess,whichrecoversseveralindividualneuralsignalsfromeachelectrodeatthecostofadditionalpowerconsumptionandincreasedsystemsizewhichnecessitatesperformingspikesortingoutsidethebody.Thisresultsinalargercommunicationbandwidthbecausewindowsofdataaroundpossiblespikesorfeaturesfromthosewindowsmustbetransmittedcomparedtoonlysendingoutspiketimes.RecentresultssuggestthatthespikesortingstepmaypossiblybeeliminatedwithoutseveredegradationofBMIperformance[ 2 35 ]thuslendingcredibilitytosolelytransmittingspiketimefordatareductionforsomeapplicationssuchalowprecisionBMIs(approach4inFigure 1-5 ).DetailsaregiveninSection 1.3.1 .Howeverthisclaimhasnotbeenexhaustivelytestedandneuroscientistswillcontinuetorequirespikesortingforstudyingthebrain.Forcaseswherespikesortingisnecessarytherearedatareductionschemesthatretainmoreinformationthansolelytransmittingspiketimesbutattheexpenseofhigherdatabandwidths(approaches1,2,and3inFigure 1-5 andSection 1.3.1 ). 1-5 )orwindowsaroundthespikes(approach3inFigure 1-5 )canbetransmittedwhilestillobtainingsignicantdatareductioncomparedtotransmittingthesampledandquantizedrawwaveforms(approach1inFigure 1-5 ).Ofcoursethereisatradeobetweendata 28

PAGE 29

Oneoptionistowirelesslytransmitasampledanddigitizedclipoftherawwaveformsurroundingthespikeforspikesortingoutsidethesubjectwherepowerandsizeconstraintsarelessstringent[ 9 ].Thisallowsfordatareductionwhileretainingtheuseoftraditionalspikesortingmethodsontheback-end.ThemaindrawbackiscurrentlycomplexdigitalVLSIcircuitryisusedtostorethewaveformuntilspikedetectionisperformedandperformspikedetection.ThepowerrequirementoftheVLSIcircuitryiscurrentlyprohibitiveforimplantation. Anotheroptionistoextractandsendthefeaturesthemselvesforspikesorting[ 36 ],buthowtogetthefeaturesoutwirelesslyatlow-powerisproblematicastheyneedtobequantizedandsentinagroupwithallthefeaturesfromonespikealongwiththespiketime.Currentlynogrouphasyettosolvethisissue. 37 ]approach4inFigure 1-5 ).Thisgreatlyreducesthebandwidthrequiredtotransmittheneuralsignalsbecausespikeoccurrencesaresparsewithinneuraldata.SpikedetectionisalsotherststepinmostofthedatareductionmethodswhichpreserveenoughinformationforspikesortingasshowninFigure 1-5 Spikedetectionmustbeasaccurateaspossiblebecausemisseddetectionerrorspropagatethroughthesystemasmissedinformation(missedneuralspike).Falsedetectionsalsopropagatethroughthesystemasincorrectadditionalinformation(falseneuralspikes)unlesswindowsofdataaroundthespikesaretransmitted.Then,aspikesortingprocesscanallowsomeofthewindowstobeclassiedasnoisetoreducethefalse 29

PAGE 30

Spikedetectionisalongstandingissueinneuroscience.Popularspikedetectionmethodsincludeamplitudethresholding,absolutevalue,energybased,wavelets,matchedlters,andtemplatematching.Currently,thereisnoconsensusinthecommunityastothebestapproachtospikedetection,particularlyforrobust,unsupervised,andcomputationallysimplemethods.Eachoftheproposeddetectiontechniqueshaveshortcomingsforimplantedapplications. Amplitudethresholdingisthesimplestandlowestinpowerspikedetectorsinceitisasubsetoftheothermethodsforthisbinarydetectionproblem.Itistheeasiesttousewithonlyoneparametertoset,thethresholdlevel,anditisthemostcommonspikedetectiontoolusedthoughitisoftenpairedwithadditionalprocessingtoachieveacceptabledetectionperformance.Forinstance,itcanbepairedwithrequirementsthatthesignalpassthroughtwothresholds(sometimescalledahoop)withinacertainamountoftimetoincreaseitsperformance.Thiscaninsuretheslopeissucientasitpassesthethresholdsorthatthethresholdcrossingisspikelikeinthatthesignalrisestocrossthethresholdandthenfallstocrossthethresholdagainwithinacertainperiodoftime.Theamplitudethresholddetector'sperformancequicklybeginstofailasSNRdropsthoughanditisnotrobusttoDCdrift[ 27 ]. Theabsolutevalueofthesignalcanbeusedtoallowforanequaldetectionrateofspikeswithlargerpositiveornegativeamplitudes.Thetradeoisthedetectormustbeblindedafteradetectionsothatspikeswithmorethanonephasearenotdetected 30

PAGE 31

9 ]. Severalgroupshavetriedtousenon-linearenergyoperators(NEO)todetectspikes[ 38 { 40 ]buttheyaretoosensitivetonoise,sotheyonlyperformwellforextremelyhighSNRs.RecentlyanewclassofmultiresolutionTEOs(TeagerEnergyOperator,atypeofNEO)waspresentedwithnoticeableimprovementoverpreviousNEO'sforlowSNRdata[ 41 ].Themultiresolutionapproachallowsthedetectortoimposeadditionalconstraintsontheenergyfunctiontoconsideritaspike,whichkeepsthedetectorfrombeingassensitivetonoise.Comparedtothepreviousenergydetectorsthismethodrequiresmorecomputationsince,formultiresolution,itrequiresseveralTEO'stobeevaluatedinparallelandthencombinedtomakeanaldecision. Recentlywaveletshavebecomepopularbecausetheyallowforlocalizationinboththefrequencyandtimedomainwhichisimportantfortransientandnon-stationarysignalssuchasneuraldata.Severalgroupshavedevelopedo-linespikedetectorsbasedonwavelets[ 42 43 ].WiththeincreaseinPCcomputationalpower,algorithmshavebeendesignedforwaveletdecompositiontoruninalmostreal-time[ 44 ].Applyingnearreal-timewaveletalgorithmstospikedetectionhasresultedinbetterperformancethanexistingsinglescalemethodsbutcurrentwaveletcircuitsconsumetoomuchpowerforimplantation. Themostcomplexandpossiblyoneofthemoreaccuratespikedetectionmethodsismatchedltersortemplatematching.Ifthesignalwasembeddedinwhitenoise,thematchedlterwouldmaximizetheSNRandbethemostaccuratespikedetector,butdistantneurons,correlatedwiththesignal,creatingnon-whitenoise[ 45 46 ].Matched 31

PAGE 32

Theproblemwithmatchedltersisthatthelteringcircuitryrequirestoomuchpowertobeimplantedandthespiketemplatestendtobeunstable.Anoptionistouseasimplespikedetector(suchasamplitudethreshold),withthethresholdsetlowastonotmissmanyspikes,asapreprocessortoparseoutwindowsofdataaroundpossiblespikestotransmit.Then,outsidethebody,wheremorepowerisavailable,thewindoweddatacouldbecomparedtothetemplateandifthewindowdieredtoomuchthewaveformcouldbedisregardedasnoise.Whilealowerthresholdincreasesthenaldetectionperformance,thedrawbacksareanincreaseintransmissionbandwidthfromsendingoutmorenoisewaveformclipstoreducemissedspikes. 1-6 .Neuralsignalshavearathersparsenumberofneuralspikeswiththerestofthesignalnoise.Asthenoiseisnotimportant,itwouldbebesttousethebandwidthonthespikeportionsofthesignalandallthreemethodstakeadvantageofthis. 32

PAGE 33

UFoverallneuraldatareductionapproaches. 33

PAGE 34

47 ]andconferencepaper[ 48 ]foradditionalinformation. Thisencodingisdoneusingabiphasic-pulserepresentationbecausepulsesaredigitalwhicharemorerobusttonoisethananaloginwirelesstransmissionyetitislowerpower(100W)thandigitalbecauseitdoesnotrequireanADCs.ThebandwidthcanbereducedbymorethanfourtimesoveratraditionalADCsampledsystemat25KHzwith12-bitsofresolution. Thebiphasicsignalencodingusespulsestorepresentwhentheintegralofthewaveform(itsarea)surpassesapositiveornegativethreshold.Anexampleofaninputsignalandit'sbiphasicrepresentationisshowninFigure 1-7 .Thisautomaticallyincreasesthebandwidthduringspikesastheareaislargerandreducesitduringnoiseastheareaissmaller.Bysettingtheareathresholdappropriatelyyoucantheoreticallyobtainperfectreconstruction[ 48 ].Inpractice,onlythespikeportionsneedtobereconstructedclosetoperfectsotheareathresholdcouldbesetonlywithconsiderationtoperfectlyreconstructthespikeportionsofthesignal. AblockdiagramforthebiphasicencodingsystemisshowninFigure 1-8 .Iftheoutputoftheintegrator,y(t)reachesthepositivethresholdofthecomparator,,theoutputofthecomparatorraisesandresetstheintegratorafterashortdelay,,inthefeedbackloop.Similarly,iftheoutputoftheintegratory(t),reachesthenegativethreshold,,theoutputofthecomparatordropsandalsoresetstheintegrator.Thedelay,notshowninthesimpliedblockdiagram,setsamaximumbandwidthandwithmorestrictconstraintsitstillallowsfortheoreticalperfectreconstruction[ 47 ].Thetimingoftwoconsecutivepulsesmustsatisfythefollowingequation: 1 34

PAGE 35

Anexampleofaninputsignalandit'sbiphasicrepresentation.A)Showstheinputsignal.B)Showsthebiphasicrepresentation. 35

PAGE 36

Blockdiagramofbiphasicencodingwithintegrate-and-re(IF)neuron. wherei2f;g: Thisspikefeatureextractionmethodalsousesbiphasicpulsestoencodethedata,butitonlypreservesfeaturesaboutthespikeandlittleinformationaboutthenoisebynotfollowingthestrictconstraintsforreconstruction.Also,aleakyterm,shownasaresistorinFigure 1-9 ,isaddedtoallowagreaterreductioninbandwidthbysubtractingoutthenoiseaswellasprovidingsynchronization(rstpulsesdonotdependontheprevioussamples(noise))forthepulse-trainoutputatthetimeofthespike. Itallowsspikesortingtobedirectlyperformedondatathatiswirelesslytransmittedreducingthecomplexityonthebackend.Thespikesortingusesatraditionalmethodoftemplatematchingbutisuntraditionalbecausethewaveformispulsetrains.Thefeature 36

PAGE 37

Blockdiagramofbiphasicencodingwithleakyintegrate-and-re(LIF)neuron. extractorcanreducethebandwidthmorethanone-orderofmagnitudelowerthantheUFbiphasicencodingandmorethantwo-ordersofmagnitudelowerthantraditionalADCsampleddataat25KHzwith12-bitsofresolutionwhilemaintainingasimilarspikesortingerror. ThisdatareductionapproachisdividedintofourschemesbasedonimplementationasillustratedinFigure 1-10 .Alloftheschemesuseapulse-trainbasedspikesorterinsoftwareonaPCthatwasdesignedspecicallyforthefeatureextractionalgorithm.TherstimplementationschemeusesexistingfrontenddatareductiontechniquessuchasusingandADCandreplacesthebackendspikesortingsoftwarewithitsfeatureextractionalgorithmandspikesorteralgorithmimplementedonaPCplatform.Implementationschemetwotakesadvantageofthefeatureextractorsbandwidthreductionbyplacingthealgorithmonadigitalsignalprocessor(DSP)ininthefront-endandusingthesamesorterasinapproachone.Implementationschemethreeisahybridapproachanditimplementsthefront-endfeatureextractorinanalogtouseit'slowpoweradvantagewiththesameback-endspikesorter.Implementationschemefourispurelyanalogbothinthefront-endwiththefeatureextractorandtheback-endwiththespikesorter. 37

PAGE 38

Featureextractionandsubsequentsortingimplementationschemes. Implementationscheme1{3willbepresentedindetailinChapters 2 3 and 4 .Implementationscheme4willnotbepresentedindetailaspowersavingswithananalogback-endisnotcurrentlynecessarybecausethepowerlimitsaremuchgreaterthaninthefront-end. 1-6 )isthemostdramaticinbandwidthreductionsinceonlythespiketimeistransmitted,butitdoesnotallowforspikesorting[ 49 50 ].Aspreviouslymentionedthismaybeappropriateforapplicationssuchasinlow-precisionBMIs[ 2 35 ].Thedramaticreductioninbandwidthallowsmoreelectrodestoberecordedwhichishelpfulinmanyapplications.TheUFapproachforspikedetectionoriginatedwithabandpasslterandevolvedtoamulti-scalespikedetectioncircuitbasedonwavelets.Bothapproachesareultra-lowpowerandrobustwhilethemulti-scalespikedetectorallowsforbetterperformancethanothersimplespike 38

PAGE 39

5 and 6 2 3 and 4 .Chapter 2 introducesthenovelpulse-basedfeatureextractorfollowedbyChapter 3 wherethebandwidthreductionisexaminedandChapter 4 providesthedetailsofthefeatureextractorcircuitandchipresults.DatareductionwithFeatureextractionisfollowedbydatareductionwithspikedetectionwherechapter 5 introducesthenovelsingle-scalespikedetectorandChapter 6 extendsthesinglescalespikedetectortomultiplescalesincreasingtheperformancewithminimalpowerincrease.Chapter 7 concludesthedissertationwithanoverviewofpreviouschaptersandasummaryofcontributions. 39

PAGE 40

51 ].Thepulse-basedfeatureextractorcanbeimplementedentirelyinsoftware,entirelyinanalogcircuitry,orahybridwithanalogcircuitryforthefront-endfeatureextractorandsoftwarefortheback-endspikesorter.Pulsebasedspikefeatureextractionhasbeenusedinthisworkforlow-powerandlow-bandwidthdatatransmission.Thecircuitimplementationentailsmodifyingthecurrentspikedetector'scurrentthresholdtoanareathresholdusingaleakyintegrate-and-reneuron.Insteadoftransmittingtherawwaveform,thepulse-basedfeatureextractionmethodencodesinformationaboutthespikeinabiphasicpulsetrain.Thisgreatlyreducesthebandwidthrequiredtotransmitthespiketrainsespeciallybecausespikeoccurrencesaresparsewithintheneuraldata,whilethepulsecommunicationoerslowerpowertransmissionoptionssuchasultra-wideband(UWB).Theencodingschemeusespulsesbasedonareapertimethresholdstorepresentthespikewhilethenoiseismostlydiscarded.Onlythespikesandtheirtimewithinthespiketraincontaininformationsonottransmittinginformationaboutthenoisesavespowerwithoutanydropinsystemperformance. Thenoiseisdiscardedmoreseverelythanwhenreconstructionisneededbyusingaleakytermwithbiphasicencoding.Thisallowsthepulsetrainsforspikeswithdierentprecedingnoisetostillsynchronizewhichaidsinspikesorting.ThesystemdiagramisshowninFigure 2-1 withtheleakytermtosubtractoutnoiseinthesignal.Theleakyvaluesetsthecutofrequencyforthelow-passlterformedwiththeintegrator.Thisleakinessalongwiththeproperthresholdsettingsallowsforveryfewpulsestorepresentthenoiseandthemajorityofpulsestocontaininformationaboutthespikes.Theleakycomponentchangestheequationfortheconstraintoftwoconsecutivepulsesto 40

PAGE 41

wherei2f;gandCisrelatedtotheintegrationcapacitorandtheRisrelatedtotheleakvalue. Oncethepulsetrainshavebeentransmitted,aclassiercanthenperformspikesortingoutsidethebodywherepowerissuesarenotsocritical.Theencodedpulsesforeachspikeserveasaspikesignature,whereapulse-basedspikesortingalgorithmisusedtoclassifythespikes.Theclassierwouldbetrainedonceintheinitialsetupandthencouldbeperiodicallyretrainedifnecessarybysendingshortsegmentsoftherawwaveformsfromoneelectrodeatatime.Oneofthemoredicultcasesforthistypeofspikesorterwouldbetwospikesfromdierentneuronsbutwiththesamearea.However,inthiscase,thetallerandnarrowerspikewouldhavemorespikesinashortertimeperiodsothetwowouldhavedierentspikesignaturesandcouldstillbedistinguished. 2-1 .Iftheoutputoftheintegrator,y(t)reachesthepositivethresholdofthecomparator,,theoutputofthecomparatorraisesandresetstheintegratorafterashortdelay,,inthefeedbackloop.Similarly,iftheoutputoftheintegratory(t),reachesthenegativethreshold,,theoutputofthecomparatordropsandalsoresetstheintegrator.Theleaktermisaxedvaluetolteroutnoise.Theleakvaluesetsthecutofrequencyforthelow-passlterformedwiththeintegrator.Thisleakvaluealongwiththeproperthresholdsettingsallowsforveryfewpulsestorepresentthenoiseandthemajorityofpulsestocontaininformationaboutthespikes.Thetimingoftwoconsecutivepulsesmustsatisfythefollowingequation: 1 41

PAGE 42

Blockdiagramofbiphasicencoding. wherei2f;gandCisrelatedtotheintegrationcapacitorandtheRisrelatedtotheleakvalue. 52 ].Anotherideaistolow-passlterthepulsetrainswithafunction,suchasanexponential,somoretraditionalsignalprocessingcanbeapplied[ 53 ].Insteadoftryingtoreconstructthesignal,thespikesorterusedforthepulse-basedfeatureextractorsimilarlyconvolvesthepulsetrainwithaGaussianfunction,wherethedeterminesifthedetectorismoreofacoincidencedetector(muchsmallerthantheinterpulseinterval)orapulsecountdetector(large).AGaussianfunctionwaschosenasitismoreconcentratedaroundthepeakallowingthetobettercontrolthedetectortype.OncethepulsetrainisconvolvedwiththeGaussian,itisthencomparedtoeachuserdenedneurontemplate.ThetemplatewiththelowestMSEisamatchunlessitexceedsthemaximumallowedMSEandthenitisconsiderednoise. 42

PAGE 43

2.4.1.1 )areshowninFigure 2-2 .Aspreviouslymentioned,comparingtwopulsetrainsiscomputationallyexpensivesothesignatureisconvolvedwithaGaussiantoallowtraditionaltemplatematchingsignalprocessingtechniquestobeapplied.AnexampleofthesignaturesoncetheyareconvolvedwithaGaussianfortheneuralsimulatordataareshowninFigure 2-3 ToshowtheimportanceofselectingtheGaussian,Figure 2-4 showstheerrorversusvalueontheneuralsimulatordataset.Ifthebestdetectorwassomewherebetweenacoincidencedetectorandspikecountdetection,thecurvewouldhaveaU-shapewithasweetspotfortowherethedistancebetweenthepulsetrainsissomewherebetweenacoincidencedetectorandapulsecountdetector.Thesixneuronsintheneuralsimulatorareallverydistinctandthenoiseislowsoitdonotrequireanycoincidencedetectingtoclassifythespikes.Thus,thecurveismorelikehalfaUwiththerightendatteningoutbecauseascontinuestoincrease(lessandlesscoincidencedetector)thereislittlechangeinerror. Thereisaproblemwithusingasinglevalueof.Spikeswithdierentamplitudeshavedierentinterspikeintervalsandinfactwithinasinglespiketheinterspikeintervalchangessinceitispartofthesignalencoding.Theproblemisthatthevalueofwhichcorrespondstoacoincidencedetectororapulsecountdetectordependsontheinterspikeintervalofthepulsetrain.Thus,forasinglespikeonevaluesmeanspartofpulsetraindistancewillbecomputedusingonetypeofdetectorwhileotherpartsofthespikewillbemoretowardtheothertypeofdetector.Thisisaproblembecausethedetectortypechangesbasedoninter-pulseintervalwhichisusefulasafeature.Ausefulvarianceofintimemightbeforthedetectortobecomelessofacoincidencedetectortowardstheendofthespikebecausetheleakvalueonlysynchronizesthebeginningofthespikeandbytheendofthespiketheaccumulatednoisewillcausethelaterpulsetimestodeviatemove.Threedierentwaystosetwillbediscussed. 43

PAGE 44

Neuralsimulatorspikesignaturesfromsixdierentneuronsattwodierenttimeperiods.Thersttworowsrepresentonetimeperiodandthenexttworowsrepresentanothertimeperiods,eachwiththerawsignalontopandthesignatureforeachspikebelow.

PAGE 45

NeuralsimulatorspikesignaturesconvolvedwithaGaussiantodeterminethedistancebetweenitandthetemplates.A)Pulsespikesignature.B)PulsespikesignatureconvolvedwithaGaussian.

PAGE 46

Figure2-4. SpikesortingerrorasafunctionofthepulsetraindistanceGaussian. Onesolutionisanadaptiveleakyintegrate-and-re(LIF)thresholdtocreateamoreuniformpulserateandisinspiredfromthebiologicalneuron'sadaptivethresholdmechanismtokeeptheringratefromsaturatingandinformationbeinglost[ 54 ].AnothersolutionistheadditionofarefractoryperiodwhichdoesnotallowtheLIFtoreanotherpulseuntilafteracertainperiodoftimewhichsetsamaximumringrate.Inthiscasethoughthesignalisignoredduringtherefractoryperiodsoinformationislostandpresumablytheadaptivethresholdmethodpreservesmoreinformationandisthusmoredesirable.Thethirdoptionistochangethevalueaccordingtothespiketemplateinterspikeintervaltoproduceamoreconstantdetectoracrossthespike. Thefocusofthisresearchisfeatureextractionnotspikesorting,butinordertoanalyzetheperformanceofthefeatureextractorspikesortingmostbeperformed.Thus,thespikesortingprocedurewaskeptsimpletopurelyshowthefeatureextractionhaspotential.Resultsinthispaperwereobtainedbysimplyusingtherstspikefromeachneuronasatemplate.Thisisaworsecasetemplateformationbecauseoftenanaverageoverseveralspikesistakentoeliminatenoise.Asecondorthirdspiketemplatecouldbe 46

PAGE 47

2.4.1Data TheUFbioamplier[ 25 ],withagainof100andalowcutofrequencyof0.3Hzandahighcutofrequencyof5.4KHz,wasusedtoamplifytheneuralsimulatoroutput.Theampliedsignalwasthendigitizedat24:4KHzand34.6swerecapturedwithadigitallogicanalyzer.Theaveragespikeringrateforthedatasetis19Hz.Thesignal'sSNRisabout30dB.AportionofthesignalduringburstingwithallsixneuralspikesisshowninFigure 2-5 (A). 47

PAGE 48

NeuralsimulatorsignalwithallsixneuronsandthebiphasicpulsetrainoutputfromtheLIFcircuitforabandwidthof455pulses/sA)Neuralsimulatorsignalwithallsixneuralspikes,onethroughsixfromlefttoright.ThesecondrowisthebiphasicpulsetrainoutputfromtheLIFcircuitwiththebandwidthat455pulses/s.B)Zoomedinspikethree.C)Zoomedinspikefour.

PAGE 49

28 ]andavailableonlineat Thesyntheticsignalswereconstructedusingadatabaseof594dierentaveragespikeshapescompiledfromrecordingsintheneocortexandbasalganglia.Tomimicthebackgroundnoisegeneratedbytheactivityofdistantneurons,spikesrandomlyselectedfromthedatabasesuperimposedatrandomtimesandamplitudesforhalfthetimesofthesamples.Then,threedistinctspikeshapes(alsopreselectedfromthesamedatabaseofspikes)weresuperimposedonthenoisesignalatrandomtimes.Theamplitudeofthethreespikeclasseswasnormalizedtohaveapeakvalueof1.Thenoiselevelwasdeterminedfromitsstandarddeviation,whichwasequalto0.05,0.1,0.15,and0.2relativetotheamplitudeofthespikeclasses.Spiketimesandidentitiesweresavedforsubsequentevaluationoftheclusteringalgorithm. Thesamplerateis24KHz.Inallsimulations,thethreedistinctspikeshadaPoissondistributionofinterspikeintervalswithameanringrateof20Hzwitha2msrefractoryperiodbetweenspikesofthesameclass.Thereareatotalofvedatasetstwoeasy,twohard,andonebursting.ThemorediculteasydatasetEasy2wasusedwithanoiselevelof0.15.ThedatasetisshowninFigure 2-6 andwillbereferredtoastheCaltechdatahereafter. 49

PAGE 50

AsegmentoftheCaltechdataset. 0.4ms{1.2mswithamplitudesashighas137V.HighSNRrecordingswerechosentoincreasethecondenceofthegroundtruthspikestimes. GroundtruthswerelabelledbyahumanexpertusingSpike2[ 55 ].MoredetailedanalysisofspikesortingwithSpike2followsinSection 2.4.2 whileabriefexplanationofthespikedetectionusedinSpike2isexplainedhere.Themethodofmarkingspikeswastorstparseoutdatasegmentswithapossiblespike.Thiswasdoneusingaconservative(low)thresholdandextractingsegmentsaroundthethresholdcrossingfromthewaveform.Then,thesesegmentswereexaminedandonlythosewhichactuallycontainedaspikewerekeptandlabelledinthespiketimele.Thismeanstherewerefewfalsealarmsbutspikeswithlargenegativepeakscouldhavebeenomittediftheirsecondphasedidnotcrossthepositivethreshold.Wedeterminedthatinoneofthedatasetsovertherstvesecondsofdata18spikestthiscriteria.Thus,whencharacterizingthefeatureextractor'ssortingperformance,theerrorcomponents,falsealarmsandmisseddetections,willbeanalyzedseparately.ThiswillallowfalsealarmsfrommisseddetectionsinSpike2(ourgroundtruth)nottocountasanerror. 50

PAGE 51

Figure 2-7 showstheoriginalneuraldatawaveform.SNRwascalculatedforthesignalintermsofpowerusinganaveragedspikeshape. Figure2-7. Neuralwaveformrecordedfromrat003.Columntwoiszoomedinfromcolumnone. 55 ],apopularcommercialprogramrstwrittenin1988,whichcanspikesortoine,isusedasacomparisontothefeatureextractor'sspikesortingperformance.Spike2rstperformscrudespikedetectionbycapturingwindowsaroundeventsthatcrossauserdenedthreshold(s).Then,spikesortingisperformedwithacombinationoftemplatematchingandaPCAbasedclustercutting.templatesfromtheneuralsimulatordataareshowninFigure 2-8 andPCAanalysisshowsthewellseparatedsixclassesinFigure 2-9 .Thisprocessrequirestheusertoselectmanyparametersduringthetemplatesetupsuchasthenumberoftemplatesandallowablevariationwithinthetemplate.Spike2provides 51

PAGE 52

3{1 .Figure 2-5 (B)showsexamplesofthebiphasicoutputforspikesfromtwodierentneurons.Theregionsbetweenspikesdidnothaveanypulses. %error=missedspikes+falsepositives totalnumberspikes+falsepositives(2{3) Thereareseveralreasonsforthefeatureextractor'soutstandingperformance.First,thedatasethasahighSNRwithdistinctspikeshapesthusitisarelativelyeasyspikesortingdataset,butthisisalsotrueforSpike2.Second,thefeatureextractorhasfewerparameterstosetthanSpike2sowithlimiteddataiscouldbebetteroptimized.Thisiswhythisisnottheonlydatasetusedtoanalyzethefeatureextractor'sperformance.Third,thefeatureextractorutilizestheleakyparametertoreducenoisewhichleadstoincreasedsortingerroranditsfeaturesarerobusttonoise. 34 ]. 52

PAGE 53

Spike2'stemplatesforneurosimulatordata. 53

PAGE 54

Spike2'sprincipalcomponentanalysis(PCA)forneurosimulatordata. 54

PAGE 55

2-10 .Spike2'stemplatesandPCAareshownininFigure 2-11 andFigure 2-12 respectively.AndthetemplatesfromthefeatureextractorareshowninFigure 2-13 Figure2-10. ActualclassiedspikesfortheCaltechsimulateddata. Thefeatureextractor'serrorincreasedmuchmoresignicantlythantheneuroscientists'.Thisispartiallyduetoasinglevalueacrossthetemplateandtheelementarytemplateselectionmethod.ThethirdspikeonlyhasapositivepeakandastheGaussianisthesameacrossthetemplates,itistoolargeatthepeakstocompensatefornottheshorteramplitudes.Soatthelargepeaks,thenedetailsarelostanditbecomesdiculttodierentiatebetweenclassesonlyusingthatportionofspike.Analyzingthemisclassications,showninTable 2-1 ,conrmsthesinglesigmaissueasoftenclass1or2aremisclassiedasclass3. Thesecondproblemistemplateselection.Ifthevariabilityinthespikewaveshapescouldbeaccountedforinthepulsedomainwithanaveragetemplate,thiscouldlowerthe 55

PAGE 56

Spike2'stemplatesfortheCaltechsimulateddata. Figure2-12. Spike2'sprincipalcomponentanalysisfortheCaltechsimulateddata. 56

PAGE 57

Featureextractor'stemplatesfortheCaltechsimulateddata. Table2-1: Featureextractorsmisclassications. classiedas neuron 1 2 3 actual 1 5 33 neuron 2 5 56 3 20 1 57

PAGE 58

The\groundtruths"areusedasarstpassandthenthefalsealarmsandmisclassicationsareexaminedtoseeiftheywerecorrectandthegroundtruthswerewrongoriftheyareindeederror.Statisticalanalysiscanbeusedtoseeifthe\errors"areindeederrors(forexamplePCAcanbeperformedtoseehowwellthegroundtruthclassesareclusteredetc.)butforthemostpartitisacombinationofstatisticalanalysisaswellasanexpertsgutinstinctfromyearsofexperience.Thetermserror,misclassication,andfalsealarmwillbeusedratherlooselyasthe\groundtruths"arenotabsoluteinthemselvesandcontainerror. Theratdataspikesorting\groundtruths"obtainedfromanexpertinthearea,usingSpike2.Somemisseddetectionsoccurredbecauseonlyasinglepositivethresholdwasusedsotherearemissedspikeswhichcanbepickedoutjustbylookingatthedata,butthepositiveportionofthosespikeswasattenuatedduetonoise.ThetemplateswereformedandrenedusingPCA.TheresultstemplatesareshowninFigure 2-14 andthePCAclustersareshowninFigure 2-15 .Eachcolorcorrespondstoadierentclasswithblackbeingspikesthatwereclassiedasnoise.Itisdiculttotellhowwellisolated 58

PAGE 59

Figure2-14. Spike2exampletemplateforratdata. Histogramsoftimesbetweenspikescanalsobeplottedtoseeifmorethanoneneuronwasclassiedintothesameclass.Becauseabsoluterefractoryperiodsareknown 59

PAGE 60

Spike2'sPCAanalysisforratdata. 60

PAGE 61

2-16 showsthehistogramforthesortedratdata. Figure2-16. Spike2'shistogramanalysisforratdata. ThedierencebetweenSpike2'sclassicationandthefeatureextractoris33%error.Becausethegroundtruthshaveerrorthemselvesthisdierenceisnotnecessarilyerrorandmustbeexaminedinadierentmanner.ArstcomparisonistoseeifthetemplateslooksimilarbutasSpike2usesthetimedomainandthefeatureextractorencodesthesignalinapulsetrainandthenconvolvesitwithaGaussiansotheycannotbedirectlycompared.ThetemplatesforSpike2areshowninFigure 2-17 andthetemplatesforthefeatureextractorareshowninFigure 2-18 Next,thecorrectlyclassiedspikeswillbeexamined.PileplotsofSpike2'ssortedspikesareshowninFigure 2-19 .TheapileplotofthespikessortedcorrectlyreferencedtoSpike2'sresultsfromthefeatureextractorareshowninFigure 2-20 .Noticeneurons 61

PAGE 62

Spike2'stemplates. 62

PAGE 63

Featureextractortemplates. 63

PAGE 64

Therearethreetypesoferrorstoexamine:misclassication,falsealarms,andmisseddetections.Table 2-2 showsthedierenceinclassicationbetweenSpike2andthefeatureextractor(misclassications).Thetableshowsthepulse-basedfeatureextractorandsortermisclassiescertainneuronsmorethanothers.Forinstanceneuron2and4neverreceivemisclassicationssotheirtemplatesandrulesmustbedistinctenoughfromtheotherneurons.Howeverneuron3hadmanyofitsspikesmisclassiedatneuron1andneuron5hadmanyofitsspikesmisclassiedasneuron3sotheseclassesneedbetterseparability. Table2-2: FeatureextractorsmisclassicationscomparedtoSpike2. classiedas neuron 1 2 3 4 5 6 tot 1 0 4 2 48 65 119 2 0 0 1 0 1 2 actual 3 197 0 0 5 6 208 neuron 4 3 2 0 2 62 69 5 42 0 241 2 11 296 6 8 3 1 3 28 43 tot 250 5 246 8 83 145 Totryanddiscernwhichmisclassicationmayhavebeenlegitimate(meaningwereactuallyanerrorwithSpike2orjustdistortedbynoiseenoughthespikeresembledtwotemplates)pileplotsofthetrueneuronspikeswithallofthespikesmisclassiedasfromthatneuronareexamined.Figure 2-21 showspileplotsofthefeatureextractor'scorrectlyclassiedspikesoverlaidwiththespikesmisclassiedasfromthatneuroninthecolorneurontheyactuallybelongto.Mostofthespikesmisclassiedas2and5resideclosetothosedetectedas2sotheymaybeerrorsfromSpike2.However,mostofthespikesmisclassiedas1and6dierenoughfromthosedetectedtoshowtherulesforthoseclassesaretooloose. 64

PAGE 65

PileplotsofSpike2'ssortedspikes. 65

PAGE 66

Pileplotoffeatureextractor'scorrectlysortedspikesreferencedtoSpike2'sresults. 66

PAGE 67

Pileplotoffeatureextractor'scorrectlysortedspikeswiththosemisclassiedasthatneuronoverlaidwithadashedblackline. 67

PAGE 68

2-22 showspileplotsofthefeatureextractor'scorrectlyclassiedspikesoverlaidwiththespikesfromthatclassbutmisclassiedasanotherclassinthatclassescolor.Thesepileplotsshowshowthefeatureextractor'serrorsrelatetoSpike2'sclassication.Forexampleforneuron3thespikesthefeatureextractormisclassieddierentenoughtheymayhavecomefromanotherneuron.Thosespikesfromneuron2misclassiedbythefeatureextractorwereattheedgesofthepileplotsolikelyifthefeatureextractorrulesforclass2wereloosenedortherulesforclass4weretightenedthosemisclassicationwoulddisappear. Thisanalysisshowsthefeatureextractorandsorter'sparameterscouldbetweakedtoimproveperformancebutasthedesiredresultistooutperformSpike2(aswasshownpossiblewiththeneuralsimulatordata)onlyexpertscanobjectivelystatewhichsortingisacceptable. 68

PAGE 69

Pileplotoffeatureextractor'scorrectlyclassiedspikesoverlaidwiththespikesfromthatclassbutmisclassiedasanotherclassinthatclassescolor 69

PAGE 70

70

PAGE 71

Movingthefeatureextractortothefront-endallowsforadramaticdecreaseinbandwidththoughitisnotwithouteectonspikesortingperformance.ThatrelationshipisexaminedinthischapterwhiletheprinciplesofoperationforthefeatureextractorandsorterarenotrepeatedfromChapter 2 Twoextremesaretosettheleakagesohighthatnoneofthesignalpassesthroughortosetthethresholdssohighthatthesignalneversurpassesittosendanypulses.Ineitherofthesetwosituationsnoinformationispreserved.Anotherextremeistosettheleakagetozeroandthethresholdsverylowandthispreservesalloftheinformation,morethanenoughtoallowforperfectreconstructionontheback-endsobandwidthiswasted. IncreasesintheleakageandthresholdvaluesbothdecreasethebandwidthasshowninFigure 3-1 .Whilemorethanoneleakageandthresholdvaluewillgivethesamebandwidththeydonotnecessarilyprovidethesamesortingerror,describedinSection 3.2.1 andisdesirabletominimize,astheydonotpreservethesameinformationasshownintheplotofFigure 3-2 .Figure 3-3 showssortingerrorversusminimumbandwidthfortheleakageandthresholdcombinationsatthreedierentSNRs.Thisplotsshowtheinverserelationshipbetweenbandwidthandsortingerror.Italsoshows,thatasSNRdecreasesthesortingerrorincreasesasexpected.ThelowertheSNRthelesspointwise 71

PAGE 72

Bandwidth(pulses/s)changesforthresholdandleakagevaluesparametersandintegrationcapacitorat10pF. separationthereisbetweenpulsetrainsofdierentneuronsmeaningitiseasiertomisclassifyoneastheotherincreasingtheerrors.However,theperformanceofallspikesortersdropsasSNRdecreases. 2.4.1.1 2.4 butnotalsoinregardstobandwidth.TheLIFwassetwithathresholdandleakagevaluesuchthatitsspikesortingerrorwassimilartoSpike2'swhichresultedinabandwidthof455pulses/s.Figure 2-5 (B)showsexamplesofthebiphasicoutputforspikesfromtwodierentneurons.Theregionsbetweenspikesdidnothaveanypulses.ThebiphasicoutputwasthenspikesortedwiththeresultsshowninTable 3-1 along 72

PAGE 73

Spikesortingerrorchangesforthresholdandleakagevaluesparametersandintegrationcapacitorat10pF.Thecolorrepresentslog(bandwidth)(pulses/s) withtheresultsfromSpike2.Theresultsaredividedintoeachneuronclassandthepercentcorrectlyclassied(truepositives,tp)andthefalsepositives(fp)whicharespikesincorrectlyclassiedasfromthatneuron.Thebestcaseis100%tpand0%fp.Thepercenterroriscalculatedwithequation 3{1 %error=missedspikes+falsepositives totalnumberspikes+falsepositives(3{1) AsTable 3-1 shows,neuron2(thesmallestspike)isoneofthehardesttoclassifyforbothsorters.Neuron2ismorepoorlyclassiedwiththefeatureextractoratthelowerbandwidth,455pulses/sbecauseitdidnothaveenoughpulsestorepresentitandsomeinformationwaslost.Theadditionofanadaptivethreshold[ 56 ]wouldhelptoevenoutthenumberofpulsesfordierentamplitudespikestokeepmoresimilaramountsofinformationforallspikeswithouthavingtoincreasethenumberofspikesforallneuronsasintheresultwith680pulses/s.Theadaptivethresholdwillcreatea 73

PAGE 74

SpikesortingerrorasafunctionofbandwidthforthreeSNRs. 74

PAGE 75

54 ].AnothersolutionistheadditionofarefractoryperiodwhichdoesnotallowtheLIFtoreanotherpulseuntilafteracertainperiodoftime.Inthiscasethoughthesignalisignoredduringtherefractoryperiodsoinformationislostandpresumablytheadaptivethresholdmethodpreservesmoreinformationandisthusmoredesirable. Table3-1: Spikesortingperformancepercenterror. Featureextractor Spike2 455pulses/s 680pulses/s 300Kbps neuron %tp %fp %error %tp %fp %error %tp %fp %error 1 100 0 0 96.4 0 3.6 100 0 0 2 69.7 2.6 31.5 93.6 0.9 7.3 89.9 0 10.1 3 100 0 0 99.1 0 0.9 90.8 5.7 13.9 4 97.3 0 2.8 96.4 0.9 4.6 97.3 0 2.8 5 96.4 1.9 5.4 99.1 1.80 2.7 99.1 6.8 7.8 6 100 0.9 0.9 100 3.5 3.5 98.2 0 1.8 avg 93.9 0.8 6.8 97.4 1.2 3.8 95.9 2.1 6.1 Overallat455pulses/sthefeatureextractorhad6.8%errorcomparedwithSpike2whichhad6.1%error.WhilemaintainingasimilarclassicationerrortotraditionalsortingwithSpike2,thefeatureextractorrequiresmuchlessbandwidthwithonly455pulses/scomparedto300Kbpsforatraditional25KHzsampledsignalatonly12-bits.1pulse/sisequivalentto1bps.UF'sbiphasicoutputforreconstructionontheback-endwouldrequire71.9Kpulses/s.Thepulse-basedfeatureextractorcanreduceitsbandwidthevenfurtherifmoresortingerrorcanbetoleratedorincreaseitsbandwidthtolessensortingerrors.Thetwoareinverselyrelated. At680pulses/sthefeatureextractoractuallyhaslesserrorthanSpike2forthisdatasetshowingitiscompetitive.MoredatasimulationsneedtobeperformedacrossdierentSNRsanddatasetstoseeifthetrendcontinues,butonepossibleexplanationisthatthefeatureextractorpreservestheimportantinformationindistinguishingbetweenspikes 75

PAGE 76

AsummarytablecomparingdierentdatareductiontechniquesandtheiraectonsortingerrorareshowninTable 3-2 Table3-2: Bandwidthreductionsortingerrorcomparison Biphasicwith Biphasic Biphasic Spike reconstruction feature feature detection extraction extraction Front-end 72Kpulses/s 455pulses/s 680plulses/s 20bps+ spikesorting 6.8% 3.8% N/A 56 ]wouldhelptoevenoutthenumberofpulsesfordierentamplitudespikestokeepmoresimilaramountsofinformationforallspikeswithouthavingtoincreasethenumberofspikesforallneuronsasintheresultwith680pulses/s.Theadaptivethresholdwillcreateamoreuniformsortingperformanceacrossneuronswithouthavingtoincreasebandwidthassignicantly.Theadaptivethresholdisinspiredfromthebiologicalneuron'sadaptivethresholdmechanismtokeeptheringratefromsaturatingandinformationbeinglost[ 54 ].AnothersolutionistheadditionofarefractoryperiodwhichdoesnotallowtheLIFtoreanotherpulseuntilafteracertainperiodoftime.Inthiscasethoughthesignalisignoredduringtherefractoryperiodsoinformationislostandpresumablytheadaptivethresholdmethodpreservesmoreinformationandisthusmoredesirable. 2 ,severalareasofthefeatureextractionalgorithmneedtobefurtherstudiedintoimprovesortingperformancewhiledecreasingbandwidth.Theyarelistedbelow: 76

PAGE 77

77

PAGE 78

WhileapurelysoftwareimplementationshowspromiseincomparisontoSpike2,amajoradvantageofthepulse-basedfeatureextractoristhatitcanbeecientlyimplementedusingcompactlow-poweranaloghardware.Thisisdonewithapproachthree,ahybridsolution,anananalogfeatureextractorinthefront-endandasoftwarespikesorterattheback-end. 2 and 3 .Nowthedetailsofthecircuitwillbepresented.ThechipwasbuiltusingtheAMI0:5mCMOStechnology. Thefeatureextractorcircuit,showninFigure 4-1 ,takesacurrentinputandencodestheneuralsignal'sshapeinabiphasicpulsetrainusingaleakyintegrateandre(LIF)neuron,asimpleextension(suchasaddingaGmcurrentsourceoraresistorinparalleltotheintegratorcapacitorfortheleakiness)ofthebiphasicIFneuron[ 48 ].TheLIFneuronintegratesthesignalandthenproducesapositivepulsewhentheintegratedsignalrisesaboveonethresholdandanegativepulsewhenitfallsbelowasecondthreshold.TheleakinessoftheLIFsetsanareapertimethresholdtolteroutnoisewhilepreservingthespikes.Thisallowsthenoiseinthesignaltoonlytriggeranoccasionalstraypulse,andthuskeepsthebandwidthandpowerconsumptionevenlower. 78

PAGE 79

Leakyintegrate-and-re(LIF)circuit. EachblockoftheLIFwillbeexplainedthoughonlytheleakycomponentwillbediscussedindetailastheotherIFpartsfollowpreviousworkasexplainedindetailinDr.Chen'spublications[ 47 48 ]andDr.Li'sdissertation[ 57 ]. 25 ]whichprecedestheLIFprovidesavoltage.Thus,avoltagetocurrentcircuitmustbeusedwhichisshowninFigure 4-2 .C1rejectsDCwhichiscrucialtolimitthefeatureextractor'sbandwidth.AcommondierentialPMOSinputandcascodedoutputoperationaltransconductanceamplier(OTA)isusedinthevoltagetocurrentblockandit'sschematicisFigure 4-2 .AmoredetailedanalysisofthiscircuitcanbefoundinDr.Li'sdissertation[ 57 ]. 4-4 showsthecomparatorcircuitwhichconsistsofthreestages:theinputpreamplier,adecisionblock,andandoutputbuer.Theinputsignalissensedbythe 79

PAGE 80

VoltagetocurrentconvertorcircuitforLIF. inputpairM1/M2anddierentialcurrentsarecopiedtothenextstage,thedecisionblock.Inthedecisionblockpositivefeedbackisusedtoisolatetheinputpairtohelpdecreasethekickbacknoise.Thecross-connectedpairM7/M8increasethegainofthecomparator.Thediode-connectedpairM10/M11provideshysteresistorejectthenoiseontheinputsignal.M13providesaDCshifttoguaranteetheswingofthedecisioncircuitoutputisinthecommon-moderangeoftheoutputbuer.M14-M18formanoutputbuertoconvertthenaloutputofthecomparatorintoalogic-levelsignal.Aninverterwasaddedtotheoutputofthebuerasanadditionalgainstage.AmoredetailedanalysisofthecomparatorcanbefoundinDr.Chen'swork[ 25 47 ]. 4-5 andDr.Chen'sdissertation[ 47 ]includesamorethoroughanalysis.ThecircuitiscomposedoftheinputCMOSpairM1/M2withanadditional 80

PAGE 81

OperationaltransconductanceamplierOTAforvoltagetocurrentconvertorcircuitforLIF. series-connectedPMOStransistorM3inthepull-up.APMOStransistorM4isemployedfortestingpurpose.ThecontrolvoltageVbiasrefractoryadjuststheoutputrisetimeandthustherefractoryperiod.ForfeatureextractiontheVbiasrefractoryissetforminimalrefractoryperiod. 81

PAGE 82

ComparatorcircuitforLIF. gain-follower.TheOTA'sschematicisdepictedinFigure 4-6 .TheOTA'sbiasvoltageadjuststheamountofleakage.Forasmallinputsignalrangesuchasneuralsignalshavethecurrentislinear.Apositiveinputvaluewillsinkcurrentwhileanegativeinputvaluesourcescurrent. 4-7 Thepinoutfortheleakyintegrate-and-refeatureextractorchipusedcanbefoundinAPPENDIX B .Thechiphasdierentialinputanduses+/-2.5Vpowersupplies. 4-8 4-9 ,and 4-10 .The 82

PAGE 83

ResetandrefractoryperiodcircuitforLIF. testsetupsweredevelopedtolimittheamountoftimerequiredtotestthechipintheratlabasthoseresourcesarelimitedandsharedbymanypeople. TheinitialchiptestsetupshowninFigure 4-8 allowsthebasicparallelrecordingsetuptobetestedinNEB487withoutrequiringarat.ThetwothingsrecordedinparallelaretheoutputoftheUFbio-amplier[ 25 ]andtheoutputofthecurrentTDTneuralrecordingsystemusedintheNeuroprostheticsResearchGroup(NPG)lab.TDTisusedasthemasterclockandtriggersthelogicanalyzertorecordtheLIF'sbi-phasicoutputatthedesiredtime.Thepulsetimesarelaterusedforthepulse-basedsortingalgorithmtocomparethefeatureextractorandpulse-basedsortersortingperformancetoSpike2'sperformancebasedonthebio-amplier'soutput.Usingtheneuralsimulatorastheinputsignaltothetwosystemsallowsthetestingtobedoneoutsideoftheratlabandprovidesaknowngroundtruth. TheintermediatetestsetupshowninFigure 4-9 replacestheneuralsimulatorwithaleoutputusingTDT.Theleoutshouldbe+=10Vtominimizeaddednoisefromthe 83

PAGE 84

LeakycircuitforLIFimplementedasaGmcurrentsource. RX5digital-to-analogconvertor(DAC)andthePA5thatisusedtoattenuatethesignaltothedesiredneuralsignallevel.Thisallowsanyneuraldataletobeusedasinputforthechip.Thisisespeciallyusefulifsimulateddatawithknowngroundtruthsistobeused. ThenaltestsetupusesinvivorecordingsandisshowninFigure 4-10 .ItisthesameastheinitialsetupexcepttheinputsignalisnowfromaliveratandthecompleteTDTsystemusedintheNPGlabisrecordedinparallelaswell.Thisallowsacomparison 84

PAGE 85

LayoutfortheLIFcircuit. betweentwocompletesystemsaswellasanyoinespikesortersastheampliedrawneuraldataisrecorded. 4.1.2 andaredescribedindetailbelow.FrombenchtoptestingtheLIFchipwasfoundtohavemorenoisethanthepreviousIFchipthoughbothsuerfromfeedbackfromthedigitalpulseoutputtotheanaloginputevenwiththeadditionofochipdecouplingcapacitorsbetweenthedigitalpowerandgroundinputpinsandseparatedecouplingcapacitorsbetweentheanalogpowerandgroundinputpins.AnotherstudentlaidouttheLIFchipbutdidnotfollowtheIFchiplayoutwithjustaddingtheleakycomponent 85

PAGE 86

Initialchiptestsetup:CompareUFsfeatureextractorwithDr.SanchezsSpike2resultsusingneuralsimulator(487NEB). soapparentlythelayouttominimizethedigitaltoanalogfeedbackwasnotascarefulaswiththeIFchip.ThisnoticeablydegradestheperformanceoftheLIFwhenusedforreconstructionbutisnotaslargeafactorwhentheLIFchipisusedforfeatureextractionasthefeaturesarepurposelyrobusttonoiseaswellasthesortingalgorithm. 4-9 wasused.However,tomimicthecompletesystem,neuralsimulatordatarecordedfromtheUFbio-amplier[ 25 ]wasused.Thedatawasrescaledto+/-10VtoreducenoiseasmentionedinSection 4.1.2 andthenattenuatedwiththePA5backtoit'soriginalamplitudelevel.Theprocessofdigital-to-analogconversionintheRX5andattenuationinthePA5introduceadditionalnoiseontheorder 86

PAGE 87

Intermediatechiptestsetup:SetUFsfeatureextractorchipparametersusingprerecordeddatafromratthatwillbeusedininvivoexperiments(487NEB). ofVthanifthedirectoutputoftheamplierwenttotheLIFchiphowevertheabilitytoreplaytheexactsamedataforeachleakageandthresholdparametercombinationoutweighedtheadditionalnoisepresent. AleakageandthresholdvaluewerechosentoobtainthedesiredbandwidthfromthesimulationresultspresentedinSection 3.2.1 .Then,avalueaboveandbelowthatwaschosentotesttheLIFchipassimulationresultsdonotalwaysmapexactlytochipresultsbecauseofnon-idealfactors.Thisresultedinatotalofninedatasetrecordings. ThesortingerrorisshowninTable 4-1 andthebandwidthisshowninTable 4-2 .ThespikesortingerrorissimilartothesimulationresultsinSection 3.2.1 forthelowerbandwidthvaluesbutasthebandwidthincreasesthenoisefromthedigitaloutputfeedbacktotheanaloginputisincreased.Thus,thelargestbandwidthdatahasthemosteect.Also,atthispointdetection(beingabletoseparateindividualspikes)becomesanissuebecauseofthecurrentimplementationofthesoftware.Thiscombinationofproblemsresultsinaninabilitytospikesortatallwith99%error.Theleakagelevelof747.4nAisnotlargeenoughtolosethenoiseandthusproducedpoorsortingresults.Evenwhentheleakagevalueisincreasedto847.3nAtheerrorisstilllargerthanexpectedforthebandwidthrequired.Thisislikelyduetothelargefeedbacknoisefromthedigital 87

PAGE 88

Finaltestsetup:CompareUFsanalogampandfeatureextractorwithTDTsampandDr.SanchezsSpike2results(ratlab).Firstsetupwithneuralsimulatorthenwhenworkinguserat. outputtotheanaloginputwhichaddsmanypulsestothesignal.Thiswasconcludedafterobservingindividualspikesignaturevaryingmorethanexpectedforeventhelargestofleakagecurrent.Thefeedbacknoiseisshortbutlargeinamplitudesoitwillperturbthetimingofthepulsesandoftenaddextrapulseswhichcouldaccountfortheincreasedvariationinspikesignaturescomparedtosimulationresults.Anothernoisesourceisthequantizationofpulsetimesbythelogicstateanalyzer(LSA)being5nsbutthiswasaccountedforinthesimulationsresults. Thebestperformancewasobtainedwithaleakagevaluesof946.1nAandthresholdvalueof130mVwhichproducedabandwidthof1.31Kpulses/sandanerrorof4.7%.ThisshowstheLIFchipiscapableofobtaininggoodsortingperformancebutrequiresmorebandwidththanexpected.Morecarefullayouttoseparatethesensitiveanalogsignalsformthedigitalsignalsonthechipshouldallowthebandwidthtodecreaseforthesameperformance. Asummarytablecomparingthedierentbandwidthreductiontechniques'bandwidth,powerconsumption,andsortingperformanceareshowninTable 4-3 .Thefeatureextractoroersthelowestbandwidthandlowestpoweroptionwhilestillbeingcompetitiveinspikesortingsoitappearsverypromising. 88

PAGE 89

Spikesortingperformance(percenterror)fromleakyintegrate-and-re(LIF)featureextractionchip. threshold(mV) 130 150 170 leakage 747.4 99.3 25.0 21.8 nA 847.3 27.3 18.7 15.2 946.1 4.7 13.2 21.3 Table4-2: Bandwidth(pulses/s)fromLIFfeatureextractionchip. threshold(mV) 130 150 170 leakage 747.4 1.65k 1.19k 988 nA 847.3 1.42k 1.05k 885 946.1 1.31k 948 804 4-9 .Thisallowsawidevarietyofrealisticdatatobetestedbutbecausethegroundtruthsareknownabetteranalysisoftheresultscanbedone. 2 and 3 ,thecircuitdesignandchipdesignaddseveralitemsthatneedfurtherwork. 56 ]toimplementtheadaptivethresholdasmentionedinChapter 3 asanimprovementtothefeatureextractorsothepreviousworkcanbebuiltupon. 58 ]. 89

PAGE 90

Bandwidthreduction,powerconsumption,andsortingerrorcomparison Biphasicwith Biphasic Biphasic Spike reconstruction feature feature detection extraction extraction Front-end 72Kpulses/s 455pulses/s 680plulses/s 20bps+ spikesorting 6.8% 3.8% N/A power 100W 30W 30W 3W 4.1.2 .ThetestplanhasbeenfullydemonstratedupuntiltheinvivotestingandtheTDTinvivotestingcodehasbeenwrittenandtested. 90

PAGE 91

UF'sthirdapproachtoneuraldatabandwidthreductionisthemostdramaticandrequiresanimplantablespikedetector.Forimplantedcircuitryananalogimplementationisadvantageousoveradigitalimplementationbecauseithasamuchlowerpowerconsumptionanditcanbemorecompactinsize.Thus,thespikedetectionalgorithmchosenwaslimitedtoonethatwasamenabletoananalogcircuitimplementation.TakinginspirationfromSmith'sauditoryonsetdetectionscheme[ 59 ]asingle-scalespikedetectionalgorithmbasedonlteringwasdeveloped.Someofthesingle-scalespikedetectorworkpresentedinthischapterhasbeenpreviouslypublished[ 49 ]. 5-1 alongwithexamplesofthesignalateverystage.Lowpassltersareknowntohaveasimple,low-power,andsmallareaimplementationinanalogusingthesubthresholdregionofoperation[ 60 61 ].MoredetailsonthecircuitryareprovidedlaterinSection 5.3 91

PAGE 92

Single-scaledetectorblockdiagramwithsnapshotsofthewaveformaftereachblock. noisewasaddedtogivethedetectionproblemamorerealisticSNRlevel.Aslowingvarying1Hz,10mVamplitudesinusoidwasalsoaddedtothesignaltosimulatetheslowlyvaryingDCoset.Figure 5-2 A)showstheoriginalneuraldatawaveformandB)showsthe0dBSNRwaveformwithanoset.SNRwascalculatedforthesignalintermsofpowerusinganaveragedspikeshape. 62 ].ROCcurvesplottheprobabilityofacorrectdetection(alsoknownasahit)versustheprobabilityofafalsedetection(alsoknownasafalsealarm).Thereisalwaysatrade-obetweentheoptimaldetectionofallthespikesandtheerroneousdetectionofnoiseasaspike.Thisspikedetectionproblemalsorequiresspiketimeestimation,meaningthattheperformancecurvecouldliebelowthechancelineforthedetectionproblem.Adetectionwasconsideredcorrectifitoccurredwithin300softheactualspiketime.TheROC 92

PAGE 93

Datausedforsimulations.A)Originalwaveform.B)0dBsignaltonoiseration(SNR)waveformwithoset.Columntwoiszoomedinfromcolumnone. 93

PAGE 94

Plotofreceiveroperatingcharacteristic(ROC)curves,0dBSNR. curvesinFigure 5-3 showsthatasalargerpercentageofspikesaredetectedmorenoisewillbefalselydetectedasaspike(alsoknownasafalsealarm).Theoptimalcurvewouldstartwithnodetectionsandnofalsealarmsandgostraightto100%correctdetectionwithnofalsealarms.TheratioofcorrectdetectionstoincorrectdetectionscanbesettothedesiredoperatingpointontheROCcurvebychoosingthecorrespondingthresholdlevel. Todeterminethedesiredcircuitcut-ofrequenciesforthespikedetector,ROCcurveswereconstructedfromnestedcut-ofrequencyiterationsaroundtypicalspikefrequencies100Hz{6KHz[ 4 ].Thecircuit'scut-ofrequencieswerechosenwiththeminimumnumberoffalsealarmsat90%correctdetectiontobe1.4KHzand5.3KHz. Oncethecutofrequencieswereselected,thealgorithmwastestedusingtheinvivorecordingsdescribedinSection 5.2.1 .Thesingle-scaledetectionmethodwascomparedtothethresholdmethodwithoutanylteringat0dBSNRwiththeresultsshowninFigure 5-3 .Forcomparisonpurposesthetwomethodswereexaminedattheir90%correctdetectionoperatingpoint.Thesingle-scalemethodoutperformedtheamplitudethresholdwithoutlteringmethodbyover30dBintermsoffalsealarmrate.Becausespikesare 94

PAGE 95

Thedatausedhasanaveragespikingrateof76Hzsoduringonesecondofdataat90%correctdetectionsthereshouldbe68correctdetectionsoutof76.At0dBthesingle-scalemethodhada2104probabilityofafalsealarm,4falsedetectionspersecond.Theincorrectdetectionprobabilityforthesingle-scaledetectorwas6%.Fortheamplitudethresholdmethodtherewere230falsedetectionspersecond,soitsincorrectdetectionpercentageismuchgreaterat76%. Thesingle-scaledetectionmethodconsistentlyoutperformedtheamplitudethresholdmethodovervaryingSNRvalues.OncetheSNRbecametoolow,2dB,neithermethodperformedwell.Herethesingle-scalemethoddegradedto25%incorrectdetectionsandtheamplitudethresholdmethodwasextremelypoorat82%incorrectdetections. Second-orderltersweresimulatedforthesingle-scalespikedetectioncircuitbuttheirperformanceoverrst-orderlterswasnegligible.Sincesecond-orderltersrequireadditionalchipareaandpowerwithoutnoticeableperformanceimprovement,theywerenotinvestigatedfurther. AnalysisoftheMatlabsimulationresultsshowedthatat90%correctdetectionalmostallofthefalsealarmscamefromnoiseridingonthesecondpeakoftheactionpotential.Spike-likenoiseoverotherpartsofthesignalwasonlydetected2%ofthetimeafalsealarmoccurred.Themajorportionoffalsealarmscouldbereducedbyblindingthedetectorforashortperiodafteritdetectsaspike.Thetrade-otothiswouldbethedetectorlosingresolutionbetweenspikes.Theamountoftimethedetectorisblindedequalstheminimumtimerequiredbetweenspikesfordetection. Sincetheamplitudethresholdisasubsetofthesingle-scalemethod'scircuitry,itwillconsumelesspower.However,itslackofrobustnesstolowSNRandslowlyvaryingDCosetshindersitsperformanceforBMIdevices.Matlabsimulationswithrealneuralrecordingsshowedthatwith90%correctdetectionsat5dBSNRthesingle-scalemethod 95

PAGE 96

5-4 showstheoverallcircuitblockdiagramforthesingle-scalespikedetector.Eachoftheblockswillbebrieydiscussed.Anoperationaltransconductanceamplier(OTA)isconguredasafollowerintegratorfortherst-orderlow-passlters.TheOTAsareruninthesubthresholdregiontoreducepower[ 60 61 ]andtoallowthecapacitorstobesmallenoughtotonchip.Thebiasvoltagesaresetochiptoenableadjustmentofthecutofrequenciesafterfabrication. Thedesiredcut-ofrequenciesforthetwolterswerefoundtobe1.4KHzand5.3KHzfromMatlabsimulationsdescribedinSection 5.2.2 .WiththeOTAs'biasvoltagessetforatransconductance,gm,of150nA/Vanddesiredcutofrequenciesof1.4KHzand5.3KHzforthelow-passltersinFigure 5-4 thecorrespondingcapacitancevaluesareC1=22:5pFandC2=4:9pFfromEquation. 5{1 96

PAGE 97

Single-scalespikedetectorcircuitdiagram. Neuralspikescanvaryinwidthfrom0.4msto3ms[ 5 6 ]dependingonthespeciesandbrainareasothecut-ofrequenciesaresettoremoveallofthenoiseoutsidethespikefrequencyrangesfortheparticularapplication. Afterthesignalhasbeenltered,thedierenceofthetwolteredsignalsistakenusinganOTA.Theoutputisthenthresholdedwithcurrent,whichissetwithVthresh.Forcompleteunsupervisedoperationanautomaticmethodforsettingthethresholdwouldbeneeded.Thisthresholdedsignalissentthroughtwoinverterstoensureabinaryoutputdecision. 5-5 97

PAGE 98

Layoutforonsetspikedetectorchip. 98

PAGE 99

Becausethecircuitdetectstheonsetofthespike,theeectivespikewidthisthewidthoftherstriseintheactionpotential.WithinniteSNR,thiswouldmeanapproximatelyhalftheactionpotentialtime,butasSNRdegradesitreduces.Thechipsfunctionalitywastestedwithpulsewidthsof100{400s: Amplitudeisthethirdcharacteristicoftheinputsignal.Extracellularneuralsignalshavepeak-to-peakamplitudesof50V{500V[ 3 ].Thissmallsignalmustrstbeampliedtogivealargervoltageswingfortheanalogspikedetectioncircuittobemoreaccurate.Today,lownoise,lowpowerneuralampliers,suchastheUFbioamplier[ 25 ],canachieveagainofupto100,sotheinputsignalamplituderangesbetween5and50mV. Theresultofa35mVsquarewavewitha125spulsewidthat25%dutycyclecombinedwitha15mVhighfrequencysinewave(tomimicneuralnoise)isshowninFigure 5-6 asthebottomwaveform.Itshowsthat10saftertheinputspikestheoutputgoeshighforashortperiod. Thechipwastestedoverawiderangeofinputsignalcharacteristicslooselypatternedafterneuraldata.Thethresholdvoltageallowsthechiptobeadjustedtochangethefalsealarmpenalty,andcorrespondinglyitsprobabilityofcorrectdetection,inaccordancewith 99

PAGE 100

Singe-scalespikedetectorchipresultswithsignalgeneratorpulsewaveformasinput.Ch2,topwaveform,istheinputsignalfromasignalgeneratorandCh3,bottomwaveform,istheoutputsignal. 100

PAGE 101

Theseinitialchiptestingresultsareinnowayexhaustiveandthecrudesignalapproximationsusedforthesetestsisnotanadequateperformancemeasure.Bionic's128-ChannelNeuralSignalSimulatorwasusedasamorerealistictestinputforthespikedetectorchip.Theneuralsimulatoroutputsarepeated11secondpatternofspikesfromthreedierentactionpotentialswithamplitudesof100V{150Vandawidthof1ms.Theinterspikeintervalis1sfor10sandthentheinterspikeintervalreducesto10msforonesecondofburstring.Noneofthespikesweresuperimposedintheoutput.TheUFbioamplier[ 25 ],withagainof100alowcutofrequencyof0.3Hzandahighcutofrequencyof5.4KHz,wasusedtoamplifytheneuralsimulator'soutput,aswouldbeusedinaBMIsystem. Thespikedetectorwasabletodetect99%ofthespikeswithoutanyfalsealarm.Thisisapproximatelyonemisseddetectionpersecondduringpeakneuralrings.Bydecreasingthethresholdslightlythedetectorreacheda100%detectionratebutafewspikesweredetectedtwicecreatingfalsealarms.Blindingthedetectorforashortperiodaftereverydetectionwouldeliminatethisproblem,butwouldalsokeepthedetectorfromdetectingtwospikescloserthantheblindingperiod.Anotheroptionwouldbetoonlyusethepositivehalfofthesignalwithahalf-waverectier.Thedisadvantageofthismethodisiftheelectroderecordingisreferencedtoanotherelectrodethepeak-to-peakamplitudeofaspikethatrstgoesnegativeandthensharplyrisesisdecreasedbyabouthalf.Two50msexamplesofthespikedetector'soutputfromtheampliedneuralsimulatorsignalareshowninFigs. 5-7 and 5-8 101

PAGE 102

63 ]parallelrecordingtestsetupasdescribedinSection4.5.2,thechipcouldnowbetestedwithmoreextensivestatisticaltestsofthesystemperformanceanddirectcomparisontopanexistingsystem.Thiswasnotdoneasatthetimetherecordingsystemwasinplacethefeatureextractoralgorithmwasbeingworkedonanddeemsmoreimportanttofocusourresourceson. Figure5-7. Single-scalespikedetectorchipresults.A)Ampliedneurosimulatorwaveforminputtothechip.B)Chip'soutput. 102

PAGE 103

Single-scalespikedetectorchipresults.A)Ampliedneurosimulatorwaveforminputtothechip.B)Chip'soutput. 103

PAGE 104

Themulti-scalespikedetectorextendsthesingle-scalespikedetectionmethod,presentedinthepreviouschapter,tomultiplescalestoallowthedetectionofspikeswithawiderrangeofwidthswithoutsacricingperformance.Thekeyideaistoimplementwaveletdecompositionandimprovespikedetectionbyindependentlycontrollingthresholdsforeachscale.Eachthresholdedscalecanthenbecombinedtoprovideasingleoutputindicatingaspikeoccurrence.Anotheroptionistouseeachthresholdedscale,whichcorrespondstoasmallrangeofspikewidths,asafeatureforspikesorting.Widthisoneofthemostimportantspikesortingfeatureswiththeotherbeingamplitude,whichcouldeasilybeaddedtothedetectorusingasimplepeakdetectorasattemptedinHoriuchi'spaper[ 36 ].Someofthemulti-scalespikedetectorworkpresentedinthischapterhasbeenpreviouslypublished[ 50 ]. 6-1 showshowtheSNRforasignalwiththesamenoiseisaectedwhenthespikewidthchanges.Sincetypicalneuralspikesvaryinwidthfrom0.4msto3ms[ 5 6 ],themultiresolutionapproachisessentialtothespikedetector'sperformancebecauseitenablesaseparatethresholdtobesetforeachfrequencyband,whichallowsforaequalratiooftypeIerrors(falsealarms)totypeIIerrors(misseddetections)acrossawidevarietyofspikewidths. 104

PAGE 105

SNRversusspikewidth. (alsoknowasconstantQ)wheretheltersspanthefrequencyspace.Thiswaveletmethodwaschosenbecauseanultra-lowpoweranalogimplementationalreadyexits,themulti-scalegammalter.ItisillustratedinFigure 6-2 andthecircuitdetailswillbeexplainedinSection 6.4 Afterthesignalisdecomposedintofrequencybandseachbandisthresholdedtodeterminethepresenceofaspikeatthatscale.TheoutputsofallthescalesareORedtogetherafterappropriatecompensationfortheirvaryingdelays.Ifthisspikedetectorwastobeusedinanapplicationrequiringspikesorting,theindividualscaleoutputcouldbetransmittedtosendthespikewidthfeatureandasimplepeakdetectorcircuit[ 36 ]couldbeimplementedtosendtheamplitudefeatureofthespike.Thiswouldallowforspikesortingoutsidethesubjectsbodywheretheconstraintsoncircuitsizeandpowerarenotasstringent. 6.3.1ScaleCombinationMethod 105

PAGE 106

Blockdiagramofthewaveletalgorithm. eachscalewasdeterminedfromthestepresponsedelayforitsbandpasslterasshowninFigure 6-3 .Thecombinationstartswiththelowestfrequencyscalebeingshiftedintimetoaccountfortheadditionaldelayfromthesecondlowestfrequencyscale.Then,thetwoscalesareORed.Itisimportanttoensurethataspikewillnotbedetectedtwiceduetothenoiseamplitudevaryingbetweenfrequencyscales.Inordertoeliminatethespuriousdetections,asmallminimumdistancecriterionisenforcedinthealgorithmateachscale.Ifthisminimumdistanceisnotmet,thelatterspikeisremoved.Then,thenewcombinedscaleiscombinedwiththenextlowestfrequencyscale.Thiscontinuesuntilthereisonlyoneoutputscaleleft. Figure 6-4 illustratestheoutputfromeachscaleforseveralspikewidthsandtheirnalcombinedoutput.Figure 6-4 A)showstheinputwhichcontainsspikewidthswithintherangeofthetypical0.3ms{3msvalues.Thebeginningofthewaveformstartswiththewidestspike,3ms,andgoestothenarrowest,0.3ms.Thewidthsare:0.3,0.4,0.5,0.6,0.7,0.8,0.9,1,1.5,2,2.5and3mswide.Thenextfourplots,Figure 6-4 B),C),D)andE)showthebandpasslteroutputscorrespondingto6K{3.4KHz,3.4K{1.9KHz,1.9K{1.1KHz,and1.1K{600Hzrespectivelyandthedetectedspikesforeachscale 106

PAGE 107

Stepresponseofbandpasslters. withcircles.Itcanbeseenthatthenarrowspikesarebetterdetectedwiththehighfrequencybandsandthewiderspikesarebetterdetectedwiththelowfrequencybandsasisexpected.NotethatwhileforthisveryhighSNRinputthemiddlefrequencybandcoulddetectallofthespikes,formoretypicalSNRsignalsitwouldnotbepossibletodetectallofthespikesonasinglescalewithouthavingmanyfalsealarms.Figure 6-4 F)showsthenalchipoutput,thecombinationofeachscalesoutput. 107

PAGE 108

Filteredscalesanddetectedspikesoneachscaleandcombinedoutput.A)Theconcatenatedinputsignalwithspikeswidthsstartingat3msandgoingassmallas0.3ms.B)6K{3.4KHzbandpasslteredsignalwiththedetectedspikesonthisscaleshownwithcircles.C)3.4K{1.9KHzbandpasslteredsignalwiththedetectedspikesonthisscaleshownwithcircles.D)1.9K{1.1KHzbandpasslteredsignalwiththedetectedspikesonthisscaleshownwithcircles.E)1.1K{600Hzbandpasslteredsignalwiththedetectedspikesonthisscaleshownwithcircles.F)Combinedoutputfromeachscale. 108

PAGE 109

6{1 providestheoptimalthresholdforthedesiredcostsoffalsealarms(C10),misseddetection(C01),andcorrectdetections(C11forthesignalandnoiseandC00fornoise). Ifallofthecostsareweightedequally,thecosttermbecomesoneanddoesn'taecttheequation.Inthisequation2representsthevariance,themean,andPiistheprobabilityofnoise(P0)ortheprobabilityofspikeplusnoise(P1).AvisualrepresentationofBayes'detectorequationisshowninFigure 6-5 .Thedistributionontheleftisthenoiseandthedistributionontherightisthesignal(spikeplusnoise).ytisthethresholdorfromEquation. 6{1 .Thedarkgreycoloredregionisthefalsealarms(typeIerrors)andthelightgreyshadedregionisthemisseddetections(typeIIerrors).Ifthecostsofallerrorsandcorrectdetectionswereweightedequallythenforthisexamplethethresholdwouldbesetat0.5. UsingBayes'equationtheindividualthresholdsofeachscalearerelatedtooneanothersothatsettingonethresholdautomaticallysetstherest.Weareworkingtomakethethresholdrelationshipsfromonescaletoanothermorerobustsincesomeparameters,suchastheprobabilitiesandvariance,arepresetbasedonpastperformancebutinrealitytheyvaryovertimeandthisisincludedinthefutureworkssection.Oneoptionforsettingthethresholdontheonescaleistoallowrawsegmentsofeachchanneltobeperiodicallytransmittedoutsothatahumancanadjusttheparametersforthecircuitperiodicallyandsendthenewparametersbacktothecircuit.Thiscouldmeanthateachchanneltakesturnshavingitsrawdatatransmittedwhiletherestofthechannelsonlysendtheirdetectedspikessotheoveralltransmissionbandwidthlimitationsarestillmet. 109

PAGE 110

IllustrationofBayes'detectorprinciples. Groundtruthswerelabelledbyahumanexpert.Themethodofmarkingspikeswastorstparseoutdatasegmentswithapossiblespike.Thiswasdoneusingaconservative(low)thresholdandextractingsegmentsaroundthethresholdcrossingfromthewaveform.Then,thesesegmentswereexaminedandonlythosewhichactuallycontainedaspikewerekeptandlabelledinthespiketimele.Thismeanstherewerefewfalsealarmsbutspikeswithlargenegativepeakscouldhavebeenomittediftheirsecondphasedidnotcrossthe 110

PAGE 111

Figure 6-4 withthethreshold,shownwithadashedhorizontalline,setforplotB)andautomaticallysetfortheremainingbands. 111

PAGE 112

Toincreasethegroundtruthaccuracyandremoveanybiasofanindividualexpertseveralexpertsshouldmarkthegroundtruthswiththeiraveragedresultsbecomingthegroundtruth.Markingspiketimesisatedioustaskthough,soitisdiculttoenlisthumanexpertstomarkalotofdata.Fortheparsingoutofpossiblespikesasecondnegativethresholdcouldbeusedinadditiontothepositiveonetodecreasethenumberofmisseddetectionsinthegroundtruths. TomakethishighSNRrecordingmoresimilartothetypicalrecordings,whiteGaussiannoisewasadded.Aslowlyvarying0.1Hz,10VamplitudesinusoidwasalsoaddedtothesignaltosimulateDCosets.Figure 6-7 A)showstheoriginalneuraldatawaveformandB)showsthe0dBSNRwaveformwithanoset.SNRwascalculatedforthesignalintermsofpowerusinganaveragedspikeshape. 62 ].Thereisalwaysatrade-obetweentheoptimaldetectionofallthespikesandtheerroneousdetectionofnoiseasaspike.Thisdetectionproblemalsorequiresspiketimeestimation.Adetectionwasconsideredcorrectifitoccurredwithin500softheactualspiketime.TheratioofcorrectdetectionstoincorrectdetectionscanbesettothedesiredoperatingpointontheROCbychoosingthecorrespondingthresholdlevel. Themulti-scaledetectionmethodwascomparedtothesingle-scalemethodandthesimpleamplitudethresholdat0dBSNRover120sofneuraldatawiththeresultsshowninFigure 6-8 .ThesinglescalemethodwasshowntooutperformtheamplitudethresholdinChapter2,sothisdiscussionwillonlycomparethemulti-scaledetector 112

PAGE 113

Datasetusedforsimulations.A)OriginalwaveformB)0dBSNRwaveformwithoset.Columntwoiszoomedinfromcolumnone. tothesingle-scaledetector.Forcomparisonpurposesthemethodswereexaminedattheir90%correctdetectionoperatingpoint.Themulti-scalemethodonlyhad15falsealarmspersecond,whilethesingle-scaledetectorhad112.Thus,themulti-scaledetectoroutperformedthesingle-scaledetectorbyover15dBintermsoffalsealarmsfora90%correctdetectionrate.Becausespikesaresparseinneuraldatatheprobabilityofafalsealarmneedstobeafractionofapercentnottoswampthenumberofcorrectdetections. Thedatausedhasanaveragespikingrateof63Hzsoduringonesecondofdataat90%correctdetectionsthereshouldbeabout57correctdetectionsoutof63.Themulti-scalealgorithmhasasimilarperformanceboostovervariousSNRsdownto-5dBbutbelowthisSNRlevelnoneofthemethodstestedperformedwellenoughtobeusedinaBMIsystem.ForlargerSNRvaluesthemulti-scalealgorithmcontinuedtooutperformthesingle-scalemethod.Ananalysisofthefalsealarmsdetectedbythemulti-scalemethodat90%correctdetectionsshowedthatalmostone-thirdofthefalsealarmswere 113

PAGE 114

PlotofROCcurves,0dBSNR. fromspikeswhosenegativesidewaslargerthanthethresholdbutwhosepositivesidewassmallerthanthethreshold.Thismeansthatgroundtruths,suchasourown,donebyexaminingeventsthatsurpassasinglethresholdproduceabiasandmaybemisleadingifnotexaminedintheperformanceanalysis. 64 ](cascadeoflow-passlters)asshowninFigure 6-9 bytakingthedierenceofadjacenttaps,XkXk1.Toachieveawide-rangeofcut-ofrequencies,aresistivelineisconnectedalongthebiascontrolsofeachlow-passlter.WiththeOTAsoperatedinthesubthresholdregion,thislinearvoltagedropacrosstheresistivelineprovidesanexponentialchangeinthebiascurrents,whichinturnproportionallyvariesthecutofrequencies.ThisallowseachltertobeconstantQ,meaningtheratiobetweenthecenterfrequencyofthelterandthebandwidthremainsconstant,whichprovideslocalizationin 114

PAGE 115

Multi-scaleGammaltercircuit. boththetimeandfrequencydomains.Theformalproofthatthedierenceofneighboringtapsofthemulti-scalegammalterisawaveletisincludedinAPPENDIX A Thetransferfunctionofthekthstageofthemulti-scalegammalterisgivenby whereaisapresetattenuationfactorgivenbyEquation 6{3 Currently,ave-tapgammalter,Figure 6-9 ,with10pFcapacitorsisusedtoprovidefourfrequencyscales.Thedierenceofeachsetofneighboringtaps,XkXk1,formbandpassltersandarethresholdedtodeterminethepresenceofaspikeateachscale.IftheoutputsofallthescalesshouldbecombinedtheycanbeORedtogetherafterappropriatecompensationfortheirvaryingdelays;however,circuitryforthishasnotbeendesigned.Ifthisspikedetectorwastobeusedinanapplicationrequiringspikesorting,eachscale'soutputwouldbetransmittedtosendthespikewidthfeatureandasimplepeakdetectorcircuit[ 36 ]couldbeimplementedtosendtheamplitudefeatureofthespike. 6-10 115

PAGE 116

Multi-scalespikedetectorchiplayout. 116

PAGE 117

2 4 )wasbeingdevelopedandthedecisionwasmadetofocusalloftheeortsonthefeatureextractor. 117

PAGE 118

ThefeatureextractorcircuitwasfabricatedusingAMI0:5mCMOStechnology.Thechipisa1:5mm1:5mm40-pinDIPwith504m356mofcircuitarea.Itconsumes30Wofpowerandthechipsperformedclosetothesimulationresultsevenwiththenoisefromthefeedbackfromthedigitaloutputtotheanaloginputofthecircuit,showingthefeaturesarerobusttonoise.Thechipshowspromisingresultstowardssuitabilityforinvivoneuralrecordings. Afully-integratedultralow-powermulti-scaleneuralspikedetectorhasalsobeendemonstrated.Itimplementsacontinuouswavelettransformusingultralow-powercircuitrytoallowforimplantation.Itonlyconsumes3WofpowerandwhenitisusedinconjunctionwiththeUFbioamplieritwillconsumeevenlesspowersincetheamplierlowpassltersthesignalat5.4KHz.Thisallowsforonelesscascadedlterinthemulti-scalegammalter.Themulti-scalespikedetectorcircuitwasfabricatedusingAMI0:5mCMOStechnology.Thechipisa1:5mm1:5mm40-pinDIPwith461m221mofcircuitarea.Themulti-scalespikedetector'sperformancewascharacterizedagainstsimplerdetectionmethodsthroughROCcurvesasshowninFigure 118

PAGE 119

.Thegureshowsthatthemulti-scalespikedetectoroutperformstheamplitudethresholdingmethod,thesimplestspikedetector. Thepowerrequiredforamplitudethresholdingislessthanforthemulti-scaledetectorsinceitisasubsetofthemulti-scaledetector.Thepowerrequiredforthemulti-scalespikedetector,3W,isstillnegligiblethoughcomparedtothepowerrequiredfortherestoftheneuralrecordingsystemsincetoday'sneuralampliersaloneconsumearound80Wofpower.Therefore,thepowersavedbyusinganamplitudethresholdforspikedetectionisnegatedbyitslargernumberoffalsealarmstoobtainthesamedetectionrate.Similarly,themulti-scalespikedetectoroutperformsthesingle-scalemethodandthoughituses2Wmorepower,atthesystemleveltheextrapowerrequiredisnegligible. Theoreticalanalysis,simulations,andchipmeasurementresultsshowthatthemulti-scalespikedetectorisagoodcompromisebetweenpower,transmissionbandwidth,area,andperformanceforanimplantabledevice. Anovelpulsetrainbasedsortingalgorithmwasalsodeveloped,analyzed,andtestedtoenableasystemperformancemetricforthefeatureextractorwithspikesortingerror. 119

PAGE 120

120

PAGE 121

Themulti-scalegammaltercircuitwaspreviouslystatedinChapter3tobeacontinuouswaveletbutwithoutaformalproof.Theformalprooffollows:Waveletslocalizewellinboththetimeandfrequencydomain.Onewaytodeneamotherwaveletiswithabandpasslter[ 65 ]whichcanbeimplementedasthedierenceoftwolowpassltersasinEquation A{1 withEquation A{2 beingthetransferfunctionofthekthstageofacascadeoflowpasslterssincethedierenceofadjacentlterswillbeabandpasslter. k(!)=k+1(!)k(!)(A{1) k(!)=NYk=i1 Inourcasethiscascadeoflowpassltersisthemulti-scalegammalterwiththecutofrequenciesofeachltervaryingonalogscalewhentheoperationaltransconductanceamplier's(OTA)areruninthesubthresholdregionandtheirbiasvoltagesvariedlinearlythroughtheresistiveline(seeFigure A-1 ).Thislogvariationofcut-ofrequenciesallowseachltertobeconstantQ,meaningtheratiobetweenthecenterfrequencyofthelterandthespectrumwidthremainsconstant.Therefore,thedesiredlocalizationinboththetimeandfrequencydomainsisobtained.Theequationforthecontinuouswavelet FigureA-1. Multi-scaleGammaltercircuit. 121

PAGE 122

Cascadedlterstructureforcontinuouswavelettransform(CWT)decomposition(TypeII). transform(CWT)atscalekisEquation A{3 andthecorrespondingcascadedlterstructureisshowninFigure A-2 [ 66 ]. Toprovethatthemulti-scalegammaltercircuitisawaveletthemotherwaveletmustmeettheadmissionscondition,Equation A{4 ,whichguaranteeslocalizationinboththetimeandfrequencydomainaswellastheexistenceofaninversionformulaforthecontinuouswavelettransform[ 65 67 ].Ifthenumberoffrequencybandsissucientlyhigh,thegammakernelconstitutesacompletesetinL2space[ 68 ]andiscontinuoussoEquation A{4 simpliesto(0)=0.Forourcasek(!)j!=0=0,sotheadmissionsconditionholdsandguaranteesthatthemulti-scalegammalterisawavelet. Whilewaveletbasesarenotrequiredtobeorthogonaltheyoftenaretosimplifythereconstructionofthesignal.Thoughthegammabasesarelinearlyindependenttheyarenotorthogonal.Thegammalteriseasytoimplementinanalogwithlowpowerconsumptionandsmalldieareawhicharebothveryimportanttoourprojectsincethiscircuitwillbeimplantedundertheskin.Sincesignalreconstructionisnotneededinspike 122

PAGE 123

66 ]couldbeused. Thiscompletestheproofthatthedierenceofmulti-scaleGammalter'sadjacenttapsimplementacontinuouswaveletdecomposition. 123

PAGE 124

1.VssofGmanalogpower(-2.5V) 2.VccofGmanalogpower(-2.5V) 3.Sameas1 4.N/A 5.Sameas2 6.Ground(0V) 7.N/A 8.Referenceinput(Theoretically0V) 9.N/A 10.BiascurrentofGm(8uA) 11.Leakycurrent 12.N/A 13.BiascurrentofoutputbuerOTA(100nA) 14.N/A 15.OutputofGm,alsotheinputofcomparators 16.N/A 17.Input 18.N/A 19.VddofPad(2.5V) 20.N/A 21.Vssofcomparatorsdigitalpower(-2.5V) 22.Vddofcomparatorsdigitalpower(2.5V) 23.N/A 24.Biascurrentofcomparatorrststage(450nA) 25.Negativethresholdvoltage 124

PAGE 125

29.N/A 30.VssofPad(-2.5V) 31.XORofpulses 32.ComplimentaryXORoutput 33.Refractorycurrent 34.Negativedirectionpulse 35.N/A 36.Resetvoltage(0V) 37.N/A 38.Positivedirectionpulse 39.Vdddigitalpower(2.5V) 40.Vssdigitalpower(-2.5V) 125

PAGE 126

[1] JohnK.ChapinandKarenA.Moxon,NeuralProsthesesforRestorationofSensoryandMotorFunction,CRCPress,2001. [2] J.Carmena,M.Lebedev,R.Crist,J.O'Doherty,D.Santucci,D.Dimitrov,P.Patil,C.Henriquez,andM.Nicolelis,\Learningtocontrolabrain-machineinterfaceforreachingandgraspingbyprimates,"PLoSBiology,vol.1,no.2,pp.193{208,Nov.2003. [3] K.D.Wise,J.B.Angell,andA.Starr,\Anintegrated-circuitapproachtoextracellularmicroelectrodes,"IEEETransactionsonBiomedicalEngineering,vol.BME-22,pp.283{247,1970. [4] K.NajaandK.D.Wise,\Animplantablemicroelectrodearrayswithon-chipsignalprocessing,"IEEEJ.Solid-StateCircuits,vol.SC-21,pp.1035{1044,1986. [5] D.R.Humphrey,\Electrophysiologicaltechniques,"inSoc.Neurosci.,Atlanta,GA,1979. [6] R.Lemon,MethodsforNeuronalRecordinginConsciousAnimals,Wiley,1984. [7] KennethD.Harris,DarrellA.Henze,JozsefCsicsvari,HajimeHirase,andGyorgyBuzsaki,\Accuracyoftetrodespikeseparationasdeterminedbysimultaneousintracellularandextracellularmeasurements,"JournalofNeurophysiology,vol.84,pp.401{414. [8] K.Wise,D.Anderson,J.Hetke,D.Kipke,andK.Naja,\Wirelessimplantablemicrosystems:High-densityelectronicinterfacestothenervoussystem,"ProceedingsoftheIEEE,vol.92,no.1,pp.76{97,2004. [9] IyadObeid,AWirelessmultichannelneuralrecordingplatformforreal-timebrainmachineinterface,Ph.D.dissertation,DukeUniversity,Durham,NC,2004. [10] P.DayanandL.F.Abbot,TheoreticalNeuroscience:ComputationalandMathemati-calModellingofNeuralSystems,MITPress,2001. [11] R.R.HarrisonandC.Charles,\Alow-powerlow-noisecmosamplierforneuralrecordingapplications,"IEEEJournalofSolid-StateCircuits,vol.38(6),pp.958{965,2003. [12] R.Clark,P.Kochanek,D.Marion,J.Schiding,M.White,A.Palmer,andS.DeKosky,\Mildposttraumatichypothermiareducesmortalityafterseverecontrolledcorticalimpactinrats,"JournalofCerebralBloodFlowandMetabolism,vol.16,pp.253{261,1996. [13] T.Seese,H.Harasaki,G.Saide,andC.Davies,\Characterizationoftissuemorphology,angiogenesis,andtemperatureintheadaptiveresponseofmuscletissuetochronicheating,"Lab.Invest.,vol.78,no.12,pp.1553{1562,1998. 126

PAGE 127

L.A.GeddesandR.Roeder,\Criteriaforselectionofmaterialforimplantedelectrodes,"AnnalsofBiomedicalEngineering,vol.31,pp.837{890,2003. [15] M.M.SteckerandT.Patterson,\Electrodeimpedanceinneurophysiologicalrecordings:Theoryandintrinsiccontributiontonoise,"Am.J.ENDTechnol.,vol.38,pp.174{198,1998. [16] K.Naja,\Solid-statemicrosensorsforcorticalnerverecording,"IEEEEngineeringinMedicineandBiology,vol.38(6),pp.375{387,1994. [17] K.NajaandK.D.Wise,\Animplantablemultielectroderecordingarraywithon-chipsignalprocessing,"IEEEJournalofSolid-Statecircuits,vol.SC-21(6),pp.1035{1044,1990. [18] J.J.K.NajaandK.D.Wise,\Scalinglimitationsofsiliconmultichannelrecordingprobes,"IEEETransactiononBiomedicalengineering,vol.BME-37(1),pp.1{11,1990. [19] K.D.WiseandJ.B.Angell,\Alow-capacitancemultielectrodeprobeforuseinextracellularneurophysiology,"IEEETransactiononBiomedicalengineering,vol.BME-22,pp.212{219,1975. [20] J.JiandK.D.Wise,\Animplantablecmoscircuitinterfaceformultiplexedmicroelectroderecordingarrays,"IEEEJournalofSolid-StateCircuits,vol.20(6),pp.433{443,1992. [21] T.Akin,K.Naja,andR.M.Bradley,\Awirelessimplantablemultichanneldigitalneuralrecordingsystemforamicromachinedsieveelectrode,"IEEEJournalofSolid-StateCircuits,vol.33(1),pp.109{118,1998. [22] P.MohseniandK.Naja,\Afullyintegratedneuralrecordingamplierwithdcinputstabilization,"IEEETransactionsonBiomedicalEgnineering,vol.51(5),pp.832{837,2004. [23] M.Dagteckin,W.Liu,andR.Bashirullah,\Amulti-channelchopper-modulatedneuralrecordingsystem,"inIEEE/EMBSConference,Istanbul,Turkey,2001,pp.757{760. [24] A.P.Chandran,K.Naja,andK.D.Wise,\Anewdcbaselinestabilizationschemeforneuralrecordingmicroprobes,"inJointIEEEEMBS/BMESConference,Atlanta,GA,1999,pp.386{390. [25] D.Chen,J.G.Harris,andJ.C.Principe,\Abio-amplierwithpulseoutput,"inInt'l.Conf.IEEEEngineeringinMedicineandBiologySociety,SanFrancisco,California,2004. [26] Q.HuangandM.Oberle,\A0.5-mwpassivetelemetryicforbiomedicalappli-cations,"IEEEJournalofSolid-StateCircuits,vol.33(7),pp.937{946,1998. 127

PAGE 128

M.Lewicki,\Areviewofmethodsforspikesorting:thedetectionandclassicationofneuralactionpotentials,"ComputationandNeuralSystems,vol.9,pp.R53{R78,1998. [28] R.QuianQuiroga,Z.Nadasdy,andY.Ben-Shaul,\Unsupervisedspikedetectionandsortingwithwaveletsandsuperparamagneticclustering,"NeuralComputation,vol.16,pp.1661{1687,2004. [29] JingHu,JennieSi,ByronP.Olson,andJipingHe,\Featuredetectioninmotorcorticalspikesbyprincipalcomponentanalysis,"IEEETransactionsonNeuralSystemsandRehabilitationEngineering,vol.13,no.3,pp.256{262,2005. [30] RichardO.Duda,PeterE.Hart,andDavidG.Stork,PatternClassication,JohnWileyandSons,20001. [31] SusumuTakahashi,YuichiroAnzaia,andYoshioSakuraib,\Anewapproachtospikesortingformulti-neuronalactivitiesrecordedwithatetrodehowicacanbepractical,"NeuroscienceReseach,vol.46,no.3,pp.265{272,2003. [32] J.S.Oghalai,W.N.Street,andW.S.Rhode,\Aneuralnetwork-basedspikediscriminator,"JournalofNeuroscienceMethods,vol.54,pp.9{22,1994. [33] Tucker-DavisTechnologies,\Tdt(tucker-davistechnologies),"(updated2007;cited2007,Jan.)[Online].Available: [34] FrankWood,MichaelJ.Black,MatthewFellows,andJohnP.Donoghue,\Onthevariabilityofmanualspikesorting,"IEEETransactionsonBiomedicalEngineering,vol.51,no.6,pp.912{918,2004. [35] J.C.Sanchez,J.C.Principe,andP.R.Carney,\Isneurondiscriminationpreprocessingnecessaryforlinearandnonlinearbrainmachineinterfacemodels?,"in11thInternationalConferenceonHuman-ComputerInteraction,2005. [36] T.Horiuchi,T.Swindell,D.Sander,andP.Abshier,\Alow-powercmosneuralamplierwithamplitudemeasurementsforspikesorting,"inInternationalSympo-siumonCircuitsandSystems,May2004,pp.IV29{32. [37] PaulWatkins,RyanKier,RobertLovejoy,DanielBlack,andReidHarrison,\Signalamplication,detectionandtransmissioninawireless100-electrodeenuralrecordignsystem.,"inIEEEISCAS,Kos,Greece,May2006. [38] S.MukhopadhyayandG.Ray,\Anewinterpretationofnonlinearenergyoperatoranditsecacyinspikedetection,"IEEETransactionsonBiomedicalEngineering,vol.45,no.2,pp.180{187,Feb.1998. [39] KyungKimandSungKim,\Neuralspikesortingundernearly0-dbsignal-to-noiseratiousingnonlinearenergyoperatorandarticialneural-networkclassier,"IEEETransactionsonBiomedicalEngineering,vol.47,no.10,pp.1406{1411,2000. 128

PAGE 129

J.H.ChoiandT.Kim,\Neuralactionpotentialdetectorusingmulti-resolutionteo,"ElectronicLetters,vol.68,no.12,pp.541{543,2002. [41] J.H.Choi,H.Jung,andT.Kim,\Anewactionpotentialdetectorusingthemteoanditseectsonspikesortingsystemsatlowsignal-to-noiseratios,"IEEETransactionsonBiomedicalEngineering,acceptedforpublication. [42] XiaoweiYangandShihabA.Shamma,\Atottalyautomatedsystemforthedetectionandclassicationofneuralspikes,"IEEETransactionsonBiomedicalEngineering,vol.35,no.10,pp.806{816,Oct.1988. [43] KyungHwanKim,\Awavelet-basedmethodforactionpotentialdetectionfromextracellularneuralsignalrecordignswithlowsingal-to-noiseratio,"IEEETransac-tionsonBiomedicalEngineering,vol.50,no.8,pp.999{1011,2003. [44] Z.NenadicandJ.Burdick,\Spikedetectionusingthecontinuouswavelettransform,"IEEETransactionsonBiomedicalEngineering,vol.52,no.1,pp.74{87,Jan.2005. [45] IsaacN.Bankman,KennethO.Johnson,andWolfgerSchneider,\Optimaldetection,classication,andsuperpostiionresolutioninneuralwaveformrecordings,"IEEETransactionsonBiomedicalEngineering,vol.40,no.8,pp.836{841,1993. [46] IsaacN.Bankman,KennethO.Johnson,andWolfgerSchneider,\Optimalrecognitionofneuralwavefroms,"inInt'l.Conf.ontheIEEEEngineeringinMedicineandBiologySociety,1991. [47] DuChen,AnUltra-lowPowerNeuralRecordingSystemUsingPulseRepresentations,Ph.D.dissertation,UniversityofFlorida,Gainesville,FL,2006. [48] D.Chen,Y.Li,D.Xu,J.Harris,andJ.Principe,\Asynchronousbiphasicpulsesignalcodinganditscmosrealization,"inIEEEISCAS,Kos,Greece,May2006. [49] ChristyL.RogersandJohnG.Harris,\Alow-poweranalogspikedetectorforextracellularneuralrecordings,"inInt'l.Conf.IEEEElectronics,CircuitsandSystems,Tel-Aviv,Israel,Dec.2004,pp.290{293. [50] ChristyL.Rogers,JohnG.Harris,JoseC.Principe,andJustinC.Sanchez,\Ananalogvlsiimplementationofamulti-scalespikedetectionalgorithmforextracellularneuralrecordings,"inInt'lIEEEEMBSConferenceonNeuralEngineering,Arlington,VA,Mar.2005,pp.213{216. [51] ChristyL.Rogers,JohnG.Harris,JoseC.Principe,andJustinC.Sanchez,\Apulse-basedfeatureextractorforspikesortingneuralsignals,"inInt'lIEEEEMBSConferenceonNeuralEngineering,KohalaCoast,HI,May2007. [52] JohnathanD.Victor,\Spiketrainmetrics,"CurrentOpinioninNeurobiology,vol.15,pp.585{592,2005. 129

PAGE 130

M.C.W.VanRossum,\Anovelspikedistance,"NeuralComputation,vol.12,pp.751{763,2001. [54] MishaMahowaldandRodneyDouglas,\Asiliconneuron,"Nature,vol.354,no.19/26,pp.515{518,1991. [55] CambridegeElectronicDesign(CED),\Spike2,"(updated2007;cited2007,Jan.)[Online].Available: [56] GiacomoIndiveri,\Alow-poweradaptiveintegrate-and-reneuroncircuit.,"inIEEEISCAS,May2003,pp.820{823. [57] YuanLi,ANINTEGRATEDMULTICHANNELNEURALRECORDINGSYSTEMWITHSPIKEOUTPUTS,Ph.D.dissertation,UniversityofFlorida,Gainesville,FL,2007. [58] AlanHastings,TheArtofAnalogLayout,PrenticeHall,2005. [59] L.Smith,\Usinganonset-basedrepresentationforsoundsegmentation,"inInt'l.Conf.onNeuralNetworksandtheirApplications,Marseilles,France,Dec.1995. [60] C.Mead,AnalogVLSIandNeuralSystems,Addison-Wesley,1989. [61] Liu,Krameria,Indiviso,Debauch,andDouglas,AnalogVLSI:CircuitsandPrinci-ples,MITPress,2002. [62] R.Hippenstiel,DetectionTheoryApplicationsandDigitalSignalProcessing,CRCPress,2002. [63] Tucker-DavisTechnologies,\Tdt(tucker-davistechnologies),"(updated2007;cited2007,Jan.)[Online].Available: [64] J.G.Harris,J.Juan,andJ.C.Principe,\Analoghardwareimplementationofcontinuous-timeadaptivelterstructures,"AnalogIntegratedCircuitsandSignalProcessing,vol.18,pp.209{227,Feb.1999. [65] LokenathoDebnath,WaveletTransformsAndTheirApplications,Birkhauser,2002. [66] D.ChenandJ.G.Harris,\Ananalogvlsicircuitimplementinganorthogonalcontinuouswavelettransform,"inIEEEInt.Conf.onElectronics,Circuits,andSystems,1998,vol.2,pp.290{293. [67] GeraldKaiser,AFriendlyGuidetoWavelets,Birkhauser,1994. [68] SamelCelebi,Representationoflocallystationarysignalsusinglowpassmoments,Ph.D.dissertation,UniversityofFlorida,Gainesville,FL,1995. 130

PAGE 131

ChristyL.RogerswasborninOrangePark,FL,onJanuary7,1980tothelovingparentsReginaandGregRogers.Shehasoneyoungerbrother,ShawnRogers.Christywillmarryherdreamguy,XiaoShe,inJune2007.ChristyreceivedtheB.S.degreeinelectricalengineeringSummaCumLaudeandwithUniversityHonorsfromtheUniversityofNorthFlorida(UNF),Jacksonville,FL,in2002.Since2002,ChristyhasbeenaresearchassistantintheComputationalNeuroEngineeringLaboratory(CNEL)attheUniversityofFloridaworkingunderDr.JohnHarrisontheBrainMachineInterfaceProject.Christyisa2003recipientofaNationalScienceFoundationGraduateResearchFellowshipandalsoreceivedtheUFPresidentialFellowship.Herresearchinterestsarebiologicallyinspiredanalogsignalprocessingandmixed-signalintegratedcircuitdesign.Specically,herinterestslieindevelopinganultra-lowpowerimplantforspikefeatureextractioninneuralrecordingapplications.ChristyreceivedtheM.S.degreeinelectricalengineeringfromtheUniversityofFlorida(UF),Gainesville,FLin2004andherPh.D.degreeinelectricalengineeringinMay2007.ShewillworkatTexasInstrumentsinDallas,TXwiththeirmedicaldevicegroup. 131


Permanent Link: http://ufdc.ufl.edu/UFE0019615/00001

Material Information

Title: Ultra-low power analog circuits for spike feature extraction and detection from extracellular neural recordings
Physical Description: 131 p.
Language: English
Creator: Rogers, Christy Leigh ( Dissertant )
Harris, John G. ( Thesis advisor )
Principe, Dr. ( Reviewer )
Bashiruulah, Dr. ( Reviewer )
Ding, Dr. ( Reviewer )
Fox, Dr. ( Reviewer )
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2007
Copyright Date: 2007

Subjects

Subjects / Keywords: Electrical and Computer Engineering thesis, Ph.D   ( local )
Dissertations, Academic -- UF -- Electrical and Computer Engineering   ( local )
Genre: bibliography   ( marcgt )
non-fiction   ( marcgt )
theses   ( marcgt )

Notes

Abstract: The purpose of this dissertation is to investigate an ultra-low power implant for spike detection and spike feature extraction in neural recording applications to dramatically reduce the required communication bandwidth through the skin. Implanted systems impose four major constraints: low power consumption, small size, robustness, and limited bandwidth. The developed solution is two fold. For applications which do not require spike sorting a lower power and lower bandwidth solution exists. A novel multi-scale continuous wavelet approach is used to decompose the signal into several frequency bands to allow for individual thresholds at each band to more accurately detect the presence of a spike. For applications which require spike sorting, a spike feature extraction algorithm was developed to extract information about the spikes so bandwidth is not wasted transmitting information not relevant to spike sorting. The feature extractor's bandwidth reduction was designed with a system level view to optimize the back-end spike sorting while using minimal bandwidth. Analog very large scale integration (VLSI) circuitry was chosen to implement both the spike detection algorithm and feature extraction algorithm to allow for an ultra-low power and compact solution for the integration of many channels in an implanted device. Preliminary theoretical analysis and chip measurement results show suitability for in vivo neural recording applications for both algorithms.
Subject: detection, extraction, feature, sorting, spike
General Note: Title from title page of source document.
General Note: Document formatted into pages; contains 131 pages.
General Note: Includes vita.
Thesis: Thesis (Ph.D.)--University of Florida, 2007.
Bibliography: Includes bibliographical references.
General Note: Text (Electronic thesis) in PDF format.

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0019615:00001

Permanent Link: http://ufdc.ufl.edu/UFE0019615/00001

Material Information

Title: Ultra-low power analog circuits for spike feature extraction and detection from extracellular neural recordings
Physical Description: 131 p.
Language: English
Creator: Rogers, Christy Leigh ( Dissertant )
Harris, John G. ( Thesis advisor )
Principe, Dr. ( Reviewer )
Bashiruulah, Dr. ( Reviewer )
Ding, Dr. ( Reviewer )
Fox, Dr. ( Reviewer )
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2007
Copyright Date: 2007

Subjects

Subjects / Keywords: Electrical and Computer Engineering thesis, Ph.D   ( local )
Dissertations, Academic -- UF -- Electrical and Computer Engineering   ( local )
Genre: bibliography   ( marcgt )
non-fiction   ( marcgt )
theses   ( marcgt )

Notes

Abstract: The purpose of this dissertation is to investigate an ultra-low power implant for spike detection and spike feature extraction in neural recording applications to dramatically reduce the required communication bandwidth through the skin. Implanted systems impose four major constraints: low power consumption, small size, robustness, and limited bandwidth. The developed solution is two fold. For applications which do not require spike sorting a lower power and lower bandwidth solution exists. A novel multi-scale continuous wavelet approach is used to decompose the signal into several frequency bands to allow for individual thresholds at each band to more accurately detect the presence of a spike. For applications which require spike sorting, a spike feature extraction algorithm was developed to extract information about the spikes so bandwidth is not wasted transmitting information not relevant to spike sorting. The feature extractor's bandwidth reduction was designed with a system level view to optimize the back-end spike sorting while using minimal bandwidth. Analog very large scale integration (VLSI) circuitry was chosen to implement both the spike detection algorithm and feature extraction algorithm to allow for an ultra-low power and compact solution for the integration of many channels in an implanted device. Preliminary theoretical analysis and chip measurement results show suitability for in vivo neural recording applications for both algorithms.
Subject: detection, extraction, feature, sorting, spike
General Note: Title from title page of source document.
General Note: Document formatted into pages; contains 131 pages.
General Note: Includes vita.
Thesis: Thesis (Ph.D.)--University of Florida, 2007.
Bibliography: Includes bibliographical references.
General Note: Text (Electronic thesis) in PDF format.

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0019615:00001


This item has the following downloads:


Full Text





UJLTRA-LOW POWER ANALOG CIRCUITS FOR SPIKE FEATURE EXTRACTION
AND DETECTION FROM EXTRACELLULAR NEURAL RECORDINGS



















By
CHRISTY LEIGH ROGERS


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2007

































S2007 Clal y13 Leigh Rogers




























I dedicate this dissertation to my loving parents Greg and Regina and my loving fianci6,

Xiao for all of their loving support as I worked on my research. My mom's homemade

frozen dinners and my dad's dinner outing when he was near Gainesville for business kept

me from losing too much weight while finishing my dissertation. And Xiao knows all of his

specific encouragements. My entire family and church friends have been praying hard for

me. Only with God's strength and the support of family, friends, and my advisor, Dr.

Harris, could I finish.









ACKENOWLED GMENTS

The guidance from my undergraduate associate Dean Dr. Merckel and Professor Dr.

K~hoie helped me pursue a Ph.D. in the first place. Encouragement from my fiancii Xiao,

my advisor Dr. Harris, and my parents, and friends K~wansun and Du have helped me stay

the course even when I doubted wanting to finish my PhD. For that I am forever grateful.

I thank my advisor Dr. John G. Harris for the opportunity to work under him

and learn so much about research and academia in general and specifically bioinspired

circuitry. I am especially grateful for his encouragement to take the internship at Texas

Instruments where I meet my fiancii and his guidance in my research as well as life such

as encouragement to finish my degree before getting married. Even though I didn't ahr-7- .-

listen so well, I at least got part of the message. I appreciate Dr. Harris' understanding

and, while at times pushing too hard, not giving up on me even when I was distracted

from my research and my productivity was very lacking. Dr. Harris' experience and

knowledge from many years as a professor facilitated my studies during my PhD.

My committee Dr. Harris, Dr. Principe, Dr. Bashiruulah, Dr. Sanchez, and Dr. Ding

as well as Dr. Fox, helped me though out my research by providing valuable feedback.

Their insights, comments, and questions to ponder were very valuable to shaping this

dissertation content. Their combined expertise across areas allowed each area of my

research to be carefully examined.

I want to thank Dr. Sanchez and his Neuroprosthetics Research Group for all of

their hard work in obtaining the neural recordings from training the rats and performing

the implantation surgery, to setting up the entire experimental recording apparatus.

Dr. Sanchez also marked the ground truths used to test our spike detection and feature

extraction methods. Jack DiGiovanna, Dr. Sanchez's Ph.D. student, assisted with the

Tucker-Davis Technologies (TDT) coding for the chip test setup.

M .ny: labmates have participated in discussion about my research and provided

helping insight and/or questions but most particular Dr. Du C'I. is who graduated last










year provided much collaboration and insightfulness. Jie ".I. --!. Xu did the layout for the

leaky integrate-and-fire (LIF) chip and assisted in testing it. Also, as Dr. Harris moved me

from 444 to 489 to distribute the English speakers I was infused with Chinese in 489 which

helped me on the trip to C'I.. I to meet my future in-laws. Specifically Xiaoxiang and her

Chinese lesson recordings and Du have been instrumental in teaching me Chinese.

Last but not least I'd like to acknowledge my funding sources. 1\y research was

supported under a National Science Foundation (NSF) Graduate Research Fellowship, a

University of Florida Presidential Fellowship, and a Defense Advanced Research Projects

Agency (DARPA) sponsor grant #N66001-02-C-8022.










TABLE OF CONTENTS


pagfe


ACKNOWLEDGMENTS

LIST OF TABLES.

LIST OF FIGURES

ABSTRACT

CHAPTER

1 INTRODUCTION


1.1 Neural Signal Properties
1.2 Extracellular Neural Recording System
1.2.1 Electrodes
1.2.2 Amplifier.
1.2.3 Wireless Data Transmission
1.2.4 Spike Sorting
1.3 Neural Data Reduction.


Overview and C





constraints


1.3.1 Data Reduction for Spike Sorting
1.3.2 Data Reduction with Spike Detection.
1.4 University of Florida's Neural Recording Bandwidth
1.4.1 Biphasic Signal Coding with Reconstruction
1.4.2 Pulse-Based Feature Extraction
1.4.3 Spike Detection
1.5 Dissertation Structure


Reduction


Strategies


2 PULSE-BASED FEATURE EXTRACTION AND SPIKE SORTING:
1 BACK(-END SOFTWARE ........ .


IMPLEMENTATION
. 40


Pulse-Based Feature Extraction.
Data Reduction with Pulse-Based Feature Extractor
Spike Sorting with Pulse Trains
Matlab Simulations Results
2.4.1 Data
2.4.1.1 Neurosimulator data
2.4.1.2 Caltech simulated data.
2.4.1.3 Rat data .
2.4.2 Spike2
2.4.3 Spike Sorting Results: Neurosimulator Data
2.4.4 Spike Sorting Results: Caltech Simulated Data
2.4.5 Rat Data Spike Sorting Results
2.4.6 Future Work.










3 PULSE-BASED FEATURE EXTRACTOR AND SPIKE SORTING: IMPLEMENTATION
2 SOFTWARE FRONT-END AND BACK(-END .... .. 71


3.1 Bandwidth Parameters.
3.2 Matlab Simulations Results
3.2.1 Spike Sortingf Results:
3.2.2 Future Work.


Neural Simulator Data


III


4 PULSE-BASED FEATURE EXTRACTION AND SPIKE SORTING: IMPLEMENTATION
3 HYBRID ............ ............ 78


ent converter circuit.
rcuit.
factory period circuit


4.1 Circuit Desigfn ....
4.1.1 Circuitry .....
4.1.1.1 Voltagfe to curr
4.1.1.2 Comparator cir
4.1.1.3 Reset and refr~


4.1.1.4 Leaky circuit
4.1.1.5 Chip specifics
Test Setup.
Results .
Neural Simulator
In Vivo with Rat
Future Work.


4.1.2
4.2 Chip
4.2.1
4.2.2
4.2.3


5 SINGLE-SCALE SPIKE DETECTOR.

5.1 Algorithm .. .......
5.2 M~atlab Simulations.
5.2.1 Data
5.2.2 Receiver Operating C'I I) Il:teristics
5.3 Circuit Design .. ......
5.4 Chip Results. .. ......


(ROC) Curves


6 MULTI-SCALE SPIKE DETECTOR .......... ..


6.1 Optimal Threshold .....
6.2 Algorithm ......
6.3 M~atlab Simulations .......
6.3.1 Scale Combination Method ........
6.3.2 Threshold Scaling from One Scale to Others ...
6.3.3 Data ...... ... ... .
6.3.4 Receiver Operating C'I I) Il:teristics (ROC) Curves
6.4 Circuit Design ......
6.5 Chip Results ....


104
104
105
105
107
110
112
114
117










7 CONCLUSIONS ......... ... .. 118

7.1 Overall Conclusions ......... .. .. 118
7.2 Contribution Summary ......... .. .. 119

APPENDIX

A PROOF: THE DIFFERENCE OF THE MITLTI-SCALE GAMMA FILTER'S
ADJACENT TAPS IMPLEMENT A CONTINUOUS WAVELET DECOMPOSITION121

B LEAK(Y INTEGRATE-AND-FIRE (LIF) T69K(-AS CHIP PINOUT .. .. .. 124

REFERENCES ......._._.. ........_._.. 126

BIOGRAPHICAL SK(ETCH ....._._. .. .. 131










LIST OF TABLES

Table page

2-1 Feature extractors misclassifications. . .. .. 57

2-2 Feature extractors misclassifications compared to Spike2. .. .. .. 64

:3-1 Spike sorting performance percent error. ...... .. 75

:3-2 Bandwidth reduction sorting error comparison .... .. 76

4-1 Spike sorting performance (percent error) from leaky integrate-and-fire (LIF)
feature extraction chip. ......... . .. 89

4-2 Bandwidth (pulses/s) from LIF feature extraction chip. ... .. .. 89

4-3 Bandwidth reduction, power consumption, and sorting error comparison .. 90










LIST OF FIGURES

Figure page

1-1 Typical extracellular spike waveform with high signal to noise ration (SNR). 16

1-2 Sketch of a neuron with the parts labelled. ...... .. 16

1-3 Waveforms recorded front a linear silicon hexatrode front a pyramidal cell with
hypothesized position of the hexatrode along the sonmatodendritic axis. .. .. 17

1-4 Block diagram of wireless front-end neural recording system. .. .. .. 19

1-5 Block diagram for four degrees of data reduction. .... .. 21

1-6 ITF overall neural data reduction approaches. ..... .. 3:3

1-7 An example of an input signal and it's hiphasic representation. .. .. .. .. :35

1-8 Block diagram of hiphasic encoding with integrate-and-fire (IF) neuron. .. .. :36

1-9 Block diagram of hiphasic encoding with leaky integrate-and-fire (LIF) neuron. :37

1-10 Feature extraction and subsequent sorting intplenientation schemes. .. .. .. :38

2-1 Block diagram of hiphasic encoding. . ..... .. 42

2-2 Neural simulator spike signatures from six different neurons at two different time
periods. ......... .... . 44

2-3 Neural simulator spike signatures convolved with a Gaussian to determine the
distance between it and the templates. . . .. 45

2-4 Spike sorting error as a function of the pulse train distance Gaussian o-. .. .. 46

2-5 Neural simulator signal with all six neurons and the hiphasic pulse train output
front the leaky integrate-and-fire (LIF) circuit for a bandwidth of 455 pulses/s .48

2-6 A segment of the Caltech dataset. ......... .. 50

2-7 Neural waveform recorded front rat00:3. Colunin two is zoomed in front column
one. ............ ..... .... .... 51

2-8 Spike2's templates for neurosiniulator data. ..... .. 5:3

2-9 Spike2's principal component analysis (PCA) for neurosiniulator data.. .. .. 54

2-10 Actual classified spikes for the Caltech simulated data. ... .. .. 55

2-11 Spike2's templates for the Caltech simulated data. .... .. 56

2-12 Spike2's principal component analysis for the Caltech simulated data. .. .. 56










2-13

2-14

2-15

2-16

2-17

2-18

2-19

2-20

2-21


Feature extractor's templates for the Caltech simulated data..

Spike2 example template for rat data. .

Spike2's PCA analysis for rat data.

Spike2's histogram analysis for rat data.

Spike2's templates.

Feature extractor templates.

Pile plots of Spike2's sorted spikes.

Pile plot of feature extractor's correctly sorted spikes referenced

Pile plot of feature extractor's correctly sorted spikes with those
that neuron overlaid with a dashed black line.


to Spike2's results.

misclassified as


2-22 Pile plot of feature extractor's correctly classified spikes overlaid with the spikes
from that class but misclassified as another class in that classes color

3-1 Bandwidth (pulses/s) changes for threshold and leakage values parameters and
integration capacitor at 10 pF..

3-2 Spike sorting error changes for threshold and leakage values parameters.

3-3 Spike sorting error as a function of bandwidth for three SNRs.

4-1 Leaky integrate-and-fire (LIF) circuit.

4-2 Voltage to current convertor circuit for LIF.

4-3 Operational transconductance amplifier OTA for voltage to current convertor
circuit for LIF.

4-4 Comparator circuit for LIF.

4-5 Reset and refractory period circuit for LIF..

4-6 Leaky circuit for LIF implemented as a Gm current source.

4-7 Layout for the LIF circuit.

4-8 Initial chip test setup: Compare UFs feature extractor with Dr. Sanchezs Spike2
results using neural simulator (487 NEB).

4-9 Intermediate chip test setup: Set UFs feature extractor chip parameters using
prerecorded data from rat that will be used in in vivo experiments (487 NEB).










4-10 Final test setup: Compare UFs analogf anip and feature extractor with TDTs
anip and Dr. Sanchezs Spike2 results (rat lah). First setup with neural simulator
then when working use rat.

5-1 Singfle-scale detector block diagram with snapshots of the waveform after each
block.


Data used for simulations.

Plot of receiver operating characteristic (ROC) curves, OdB SNR.

Single-scale spike detector circuit diagram.

Layout for onset spike detector chip.

Singfe-scale spike detector chip results with signal generator pulse
input. ....

Single-scale spike detector chip results. .....

Single-scale spike detector chip results. .....

SNR versus spike width. .....

Block diagram of the wavelet algorithm. .....

Step response of handpass filters. .....

Filtered scales and detected spikes on each scale and combined ou

Illustration of B li- -' detector principles. .....

Figure 6-4 with the threshold, shown with a dashed horizontal lint
B) and automatically set for the remaining hands. .....

Data set used for simulations. .....

Plot of ROC curves, OdB SNR. .....

Multi-scale Ganina filter circuit. .....

Multi-scale spike detector chip layout. .....


5-2

5-3

5-4

5-5

5-6


5-7

5-8

6-1

6;-2

6-:3

6-4

6-5

6;-6


6-7

6;-8

6i-9

6-10


I I 1 I I I I I


waveform


as


.. 100

.. 102

. 10:3

.. 105

.. 106

.. 107

tput. .. .. 108

.. 110


e, set for plot


111

11:3

114

115

116i


A-1 Multi-scale Ganina filter circuit.

A-2 Cascaded filter structure for continuous
(Type II).


wavelet transform (CWT)


decomposition









Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

ITLTRA-LOW POWER ANALOG CIRCITITS FOR SPIK(E FEATURE EXTRACTION
AND DETECTION FROM EXTRACELLITLAR NEURAL RECORDINGS

By

C'!wly13 Leigh Rogers

May 2007

Cl.! ny~: John G. Harris
Major: Electrical and Computer Engineering

The purpose of this dissertation is to investigate an ultra-low power implant for spike

detection and spike feature extraction in neural recording applications to dramatically

reduce the required coninunication bandwidth through the skin. Implanted systems

impose four major constraints: low power consumption, small size, robustness, and limited

bandwidth. The developed solution is two fold. For applications which do not require

spike sorting a lower power and lower bandwidth solution exists. A novel niulti-scale

continuous wavelet approach is used to decompose the signal into several frequency hands

to allow for individual thresholds at each hand to more accurately detect the presence of

a spike. For applications which require spike sorting, a spike feature extraction algorithm

was developed to extract information about the spikes so bandwidth is not wasted

transmitting information not relevant to spike sorting. The feature extractor's bandwidth

reduction was designed with a system level view to optimize the back-end spike sorting

while using nxininial bandwidth. Analog very large scale integration (VLSI) circuitry was

chosen to intplenient both the spike detection algorithm and feature extraction algorithm

to allow for an ultra-low power and compact solution for the integration of many channels

in an implanted device. Preliminary theoretical analysis and chip measurement results

show suitability for in vivo neural recording applications for both algorithms.









CHAPTER 1
INTRODUCTION

The neuron is the basic information processing unit in the brain. Neurons use

electrical pulses, called action potentials or spikes, to transmit information. Their

extracellular electrical pulses can he recorded using microelectrodes that are implanted

into the brain. These recordings are called neural recordings since they record the voltage

potential caused hv neurons. To best study neural information processing, many neurons

must he simultaneously recorded in awake behaving subjects. Sate-of-the-art recording

systems require microelectrode arrays with hundreds of electrodes implanted into the

brain.

While much is still unknown about the brain, researchers have now learned enough

to integrate neural prostheses with the brain [1]. One example of a neuroprosthetic is a

Brain-Machine Interface (BMI) [2]. Alotor-based BMIls extract information front neural

recordings collected in the motor, preniotor and parietal cortices with the goal of creating

predictive models for the subject's intent of motor movement to directly control a robotic

device. Eventually, these devices could allow paraplegics to control a robotic arnt to

feed themselves or turn the pages of a book. Neural prosthetics require long-ternt neural

recordings which necessitate wirelessly transmitting the data front the electrodes through

the skin. If a wire front the electrode passed through the skin to send the data, infection

is risked and the subject tethered. Also, having many wires coming out of the head and

being tethered restricts movement of the subject.

Current instrumentation technology and surgical procedures allow for the simultaneous

recording of hundreds of electrodes. The bottleneck is how to transfer the large bandwidth

raw data streams wirelessly. Transmitting the raw voltage signals front hundreds of

channels is not possible with the current wireless bandwidth limits. Furthermore, even if

these high data rates could be met, the power dissipation of the electronic circuitry would

severely drain the implanted power supply and exceed the power dissipation limits for

preventing tissue damage.










The presented research addresses this bottleneck with two approaches to bandwidth

reduction based on if spike sorting is required or not. A novel analog spike detection

circuit to only transmit spikes times dramatically reduces the required transmission

bandwidth for applications which do not require spike sorting. The detector is lower in

power and more compact than existing spike detection methods without compromising

performance. For applications which require spike sorting a feature extraction method was

developed which still dramatically reduces bandwidth compared to current instrumentation

but it does require more bandwidth than only transmitting spike times with the

spike detector. The feature extractor is also lower in power and more compact than

existing methods for data reduction which allow for spike sorting without compromising

performance .

1.1 Neural Signal Properties

One of the most widely recorded neural signals is the extracellular biopotential

generated electrochemically from individual neurons. After a neuron receives sufficient

stimuli from other neurons, its cell membrane depolarizes which causes extracellular ionic

currents to flow. This in turn causes voltage sensitive channels to open allowing ions

(Na+ and K(+) to pass through the neuron's membrane. The result is a change in the

extracellular single-unit potential in the shape of a spike (also commonly referred to as

an action potential) with a peak-to-peak amplitude of 50 pV-500 pV when measured

extracellularly [:3]. The amplitude varies inversely with the distance between the electrode

and the neuron. A typical extracellular spike waveform from a high signal to noise ratio

(SNR) neural recording from a rat is shown in Figure 1-1. The frequency content of spikes

is mainly between 100 Hz and 6 K(Hz [4] with widths varying between 0.4 ms and :3 ms

[5, 6] depending on how far the electrode is from the neuron and what part of the neuron

is closest to the electrode. A labelled drawing of the neuron is shown in Figure 1-2. As the

spike propagates from the soma (cell body) along the axon, the spike amplitude becomes

attenuated and the width increases [7] as shown in Figure 1-:3.












60-

40-

9 20-


>-20-

-40-

-60-

-80
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time(ms)

Figure 1-1. Typical extracellular spike waveform with high signal to noise ration (SNR).




Sorna

(Cell Body).


Figure 1-2. Sketch of a neuron with the parts labelled.


er synap ic
-
Terminal


De n drite s


Node of Ranvier

\Axon


















































Figure 1-3. Waveforms recorded from a linear silicon hexatrode from a pyramidal cell with
hypothesized position of the hexatrode along the somatodendritic axis to the right. Note
variation in wave shape along the somatodenritic axis adapted from Harris et. al. [7].










One electrode may record from as many as four to six neurons but there will be

many more distant neurons whose signals become part of the noise on the signal. Typical

recorded noise levels, which include distant neurons and the electrode's thermal noise,

are around 20 pVrms [8]. It is difficult to define SNR for a neural waveform because the

signal (the spike) has varying amplitudes and widths. Spikes with a larger amplitude and

width will have a higher SNR than smaller amplitude and narrower spikes from the same

channel with the same level of noise. This alone would so-----~ -r SNR should be defined

for each neuron within a recorded waveform. However, the signal is also transient and

non-stationary so the amplitude and width of spikes from a single neuron also change over

time. Therefore, SNR is often defined using the average of all the spikes in the waveform

with typical SNR ranges between 0 dB to 12 dB [9].

Single neurons deplete their chemical reservoirs after producing a spike, inactivating

the Na channels which can not reopen until the membrane potential returns to a negative

value near the threshold. This sets a minimum time required to replenish their reservoirs

before the neuron can spike again. This period is called the refractory period and is

typically around 1 ms [10]. Because extracellular neural recordings may contain the

response of more than one neuron, spikes from different neurons could fire very close

together, within the refractory period of a single neuron. Spikes from different neurons can

even overlap and create a superimposed waveform.

IIIw: electrodes are often used to record neurons from multiple sites within the brain.

A collective group of implanted electrodes are referred to as an electrode array and is

encased in cranioplast that coats a screw anchored into the skull. This physically keeps

the electrodes from moving in reference to the skull. The brain however is itself floating in

cerebrospinal fluid (CSF) and can shift by many microns over time. As the brain moves,

the distance between the electrode and the neurons it is recording from also change so the

spike shape (amplitude and width) can change over time as the brain shifts. This results

in SNR fluctuations in the signal. To make matter worse, unavoidable electrochemical










fro nt-en d




I bio-.data reduction
Amplifier

le lectrodes i

II


back-end



II




Figure 1-4. Block diagram of wireless front-end neural recording system.

effects at the electrode-tissue interface introduce DC offsets ranging from 1-2 V across the

recording sites [11].

1.2 Extracellular Neural Recording System Overview and Constraints

Neural recording systems usually consist of an analog front-end near the recording

site, a means to get the data off the subject, and an external processing element. The

analog front-end consists of electrodes and amplifiers. Electrodes are intrusively placed

in the brain to measure the neural potentials. This signal is in the pV range so it must

he amplified close to the recording site so noise does not corrupt the small signal. The

amplification also allows for improved processing later in the system. Data reduction is

performed and then, a wired or a wireless telemetry unit transmits the data away from the

subject for more intense processing on the back-end where power and size constraints are

less stringent. A block diagram of a wireless front end and back-end system is shown in

Figure 1-4.










To continuously record from awake subjects the electrodes must he implanted and

the signal wirelessly sent out of the body because of the risk of infection if wires were to

pass through the skin. Even in current animal neural recordings that use wires from the

electrodes out of the skin there are two problems in addition to infection. The animal

must he tethered, which restricts its natural movements, and the wires must he harnessed

in a fashion so they do not become entangled. Often commutators are used to prevent

tangling and torque applied to the prosthetic from movement of the wires but this limits

on the number of wires.

In addition to the electrodes several pieces of neural instrumentation electronics, such

as an amplifier, data reducer, and circuitry to transmit the data, should also be implanted

to avoid any wires passing through the skin. Low power circuitry is necessary due to the

difficulty of charging or changing implanted batteries as well as to prevent tissue damage.

Power can he broadcast into the device, but studies have shown that if the brain tissue

increases in temperature any more than about loC the brain tissue will be damaged [12].

Power dissipation over 80 nzW/cH72 has been reported to cause general tissue damage

[13]. Circuit area must also be small due to the limited area of implantation between the

subjects' skull and skin especially on small animals.

BMI systems place strong constraints on the wireless transmission because hundreds

of channels are currently recorded with the desire to reach thousands in the future.

Transmitting raw voltages from 100 channels at a 25 K(Hz sampling rate and 8-bits of

resolution will generate data rates around 20Mbps. Furthermore, even if these high

data rates could be met, the power dissipation of the electronic circuitry would severely

drain the implanted power supply and exceed the power dissipation limits for preventing

tissue damage. Therefore, data reduction is required. Because spikes are sparse within

neural recordings, only transmitting information about the spikes provides further

significant reduction in the required transmission bandwidth compared to sampling

and quantizingf the entire signal. After the spikes are detected, there are three 1! in r~













blo-ampltier








Figure ~ ~ ~ ~ ~ ~ ~ spk 1-. lckdaga fr or eresofdaareutin

option for bndwidt redcton tatdepen d ontheo apiatioun. Thiseslsiatol





1.2.1 Electoe


the rst ofthe ystem Thebett mpiter th nurlectordnSN ios freom theelcroeh


properties allow for the most effectier reodingos.The eetoems ebooial




compatible sor bidth cancotinue recordin ove time as pposedtion being enasulatedn wtht


Gias proeqiin, blockin ietsectordng h arabilit [1]. Elecrode maeial sizpe, tip sharntiess


arond pibltye should becarefully sealed etoe mnimized no ise and tissue dramage. while

mitainsing the abile im.ty for recise implanttion. The detis one speiic trad i ueos in



telretrode dhesg isnotm prsetdee butte ohne can referdn to i fo the lieaue [8, 14e, 15]










The total input referred noise of the neural recording system must he significantly

smaller than the smallest amplitude neural signal of interest. The peak-to-peak amplitude

of neural signals can he as low as 50 pV, so it is important that the total input noise

he below about20 pV [9]. The electrode contribution to the total noise of the systenl is

mainly comprised of neural background noise and thermal noise. Neural background noise

is the sunination of all the distant neurons' electrical potentials. The thermal noise occurs

at the nietal-electrolyte interface and is related to electrode impedance and the recording

bandwidth, which has a 1/f frequency dependence [15].

There are two major categories of electrodes: passive and active. Passive electrodes

do not contain any interfacing electronic circuitry on the electrode substrate [16] and are

usually made of metal [8, 17, 18] or glass [16]. Active electrodes include electronic circuitry

on the same substrate as the recording electrode [16]. The on-chip circuitry can nminintize

the number of leads on the chip as well as nxinintize the leakage and noise associated with

sending a very small signal (;tV) over wires. In 1975, Wise at the University of 1\ichigan

was the first to produce an active electrode [19] and advances are still being made [16, 20].

1.2.2 Amplifier

Because the neural signal has such a low amplitude, it must he amplified prior to

further processing. The amplifier must have very low noise, filter the signal, and be fully

integrated and low-power to allow for implantation. Frequencies outside the neural spike

range, between 100 Hz and 6 K(Hz [4], should be removed front the signal to reduce the

noise. If a clever scheme is not used to filter out the low frequency noise, a large off-chip

capacitor will be required making the circuitry too large to implant.

Ji et al. have reported one type of Conmplenientary nietaloxidesenticonductor (C'jIOS)

amplifier which shares the same substrate with silicon electrodes [20]. The amplifier

provides a nxid-band gain of 51 dB without amplifying the DC frequency components.

However, one 1! in r~ issue with this design is the gain variability front probe to probe or









even from channel to channel on the same probe. Gains also drift during probe use and

with ambient light levels, which will eventually saturate the amplifier.

T ii II; and Wise were one of the earliest teams to reduce the random DC component

at the electrode-electrolyte interface with a reverse-biased diode to clamp the input with

the high resistance of the junction depletion region [4]. Ji et al. and Akin et al. emploi-. I

an internal bandlimiting method by using diode-capacitor filters to form the low cut-off

frequency [20, 21].This scheme suffers from several issues: optical drift that reduces

reliability, limited dynamic range, and high variability of the lower cut-off frequency [22].

Dagtekin et al. reported a multi-channel chopper-modulated neural recording amplifier

that uses the chopper modulation technique and an unbiased location in the system as

a reference to minimize the effects of the DC drift of the neural signals [23]. No chip

performance metrics were ever published, only simulated data.

C'I .1..11~ lI. et al. employ a sub-threshold NMOS transistor as a high value shunt

resistor to attenuate the DC offset to stabilize the DC baseline [24]. This resistor creates

the lower cut-off frequency of the amplifier. Up to 400 mV of DC input can be handled

without sacrificing the AC performance. However, the amplifier is unable to reject

negative DC input values. Moheseni et al. modified C'I .1..11~ Il.'s design by employing a

subthreshold PMOS transistor to reject both negative DC and positive DC input [22]. A

laser trimmed resistor is used to accurately set the lower cut-off frequency. This design

can not tolerate a DC shift higher than 400 mV. The drawback is the extra process step

required for laser trimming.

C'I. is and Harris from the University of Florida have used clever circuitry based

on Harrison's design [11] to develop a low-noise ultra-low power fully integrated neural

amplifier (bioamplifier) to meet the requirements of an implantable device. The UF

bioamplifier provides a gain of almost 40 dB, input referred noise of only 9.56 pVms, a

C' \! RR of about 59 dB, and a power supply rejection ration (PSRR) of about 45 dB [25].

It has a low cutoff frequency of 0.3 Hz and a high cutoff frequency of 5.4 K(Hz. These










specifications allow the amplifier to perform without too much added noise which would

corrupt the neural signal.

1.2.3 Wireless Data Transmission

II I.!y neural prosthetic groups are working on wireless telemetry systems to transmit

data out of the brain and send power to the implanted circuitry. The difficult part is that

the circuitry on the subject's end must he implanted necessitating small area and small

power consumption. There are two 1!! li r~ schemes that groups have used to transmit the

data wirelessly: AM (amplitude modulation) and FM (frequency modulation). Akin et

al. [21] report a system where the digitized signal is encoded into an 8-bit pulse-position

modulation (PPM), pulse-code modulation (PC11L) and transmitted with AM. The

system dissipates 2.8 mW of power and its die area is 0.7 unl2 per channel. Huang

et al. [26] use a 9-bit PPhi to encode the data and transmit it using FM. The core

area is 3.6 unn x 4.350 unn for one channel. Both of these systems utilize inductively

coupled radio frequency (R F) telemetry for both the power and the data transfer. The

analog-to-digital convertors (ADCs) and modulators expand the die area and power

consumption considerably so they must he carefully designed to allow small area and low

power consumption necessary for implantation.

1.2.4 Spike Sorting

For in vivo extracellular neural recordings, multiple neurons are recorded from the

same channel and the spike shape can he used to distinguish amongst the individual

neuron signals through a process called spike sorting. Popular spike sorting methods are

too computationally intensive for implanted devices. Thus, the spike sorter is typically

pushed to the back-end where power and area constraints are less stringent.

Neuroscientists rely on a variety of spike sorting methods utilizing different features

of the spikes with no community wide agreement as to which spike sorter is the best.

Most researchers simply use the default spike sorting software that comes with the neural

instrumentation hardware. Tradeoffs exist between performance and the ability to run










the spike sorter realtime. Also, spike sorting performance is often directly related to

the amount of time the neuroscientist spends to setup the spike sorting parameters.

Spike sorting is a classical problem in the neuroscience coninunity, with many proposed

algorithms in the literature. A overview of the 1 in r~ spike sorting methods is provided

below. For a more in depth review of spike sorting algorithms consult Lewicki [27].

There are several major categories of spike sorters: template matching, clustering

approaches, independent component analysis (ICA), neural networks, and simple threshold

based methods. An overview of each group of spike sorters with proper references to the

literature follow. Also, some groups use combinations of the above methods to refine

the classes through the use of different features in an effort to improve spike sorting

performance .

Template matching does not require any feature extraction as an average waveshape

for the spikes front each neuron is used. A neuroscientist determine the number of distinct

neurons as well as the templates. Tools such as principal component analysis (PCA) and

histogframs can he used to see examine the templates. Plotting the first 2-3 principle

components shows how much separation exists between classes and gives the user a visual

aid to see how many distinct classes exist. Histograms show if the firing rate of each

neuron violates fundamental limits meaning that class contain spikes front more than one

neuron and the templates need to be reformed. Often template matching is performed by

using a matched filter to find which template most closely matches each spike to classify

it. A threshold is used so a waveshape is classified as noise if it does not closely match any

of the templates.

When feature extraction is used it is often followed by cluster based spike sorting

to determine the neuron for each spike. By plotting each feature on its own axis a

niulti-dintensional graph is formed where any clustering algorithm can he used to classify

the data. Popular features are the spike amplitude and width, principal components,

wavelet coefficients [28], and slope of rise and fall time. The most popular feature










extraction method is PCA [29] because it can he used to extract a compact set of

orthogonal features. Some popular options for clustering are the simple k-nicans or

nearest neighbor where each cluster location is marked as the mean of the data within the

cluster. A spike is then classified to the cluster with the closest mean Euclidean distance.

More elaborate methods, such as B li- -i la <1I1-rh 11).: use statistical information about

the neurons and their spike shapes and are best suited when the clusters have significant

overlap or differ front a spherical distribution. Alany more clustering algorithms exist and

are surveyed in books such as the classic one hv Duda, Hart and Stork [:30].

ICA is a special case of blind source separation. It separates a multivariate signal

into additive subconmponents. It assumes the number of electrodes equals the number of

sources which only approaches the truth for a large number of electrodes. It also assumes

sources are mixed linearly. ICA is used in tetrode recording as usually the number of

signals is closer to the number of neurons compared to single site electrode recordings [:31].

Some groups have applied neural networks to solve the feature extraction problem but

another method must he used to sort the data to provide the ground truths to train the

neural network [:32]. Thus, the neural network hased spike sorter can only be as good as

the spike sorter used to provide ground truths to train the neural network.

Threshold-based spike sorters are the simplest spike sorters and often precede more

complicated spike sorters as a spike detector. The thresholds can he simple voltage

thresholds (positive and negative) along with some rules such as the waveform must pass

through two thresholds within a certain time (a hoop). This then forms a voltage-time

threshold. One might also impose that the spike must first go positive and then negative.

To differentiate between spikes, different threshold rules are applied to differentiate spikes

by their amplitude, width, and/or rise or fall time [27]. Some on-line spike sorters, such as

Tucker-Davis Technolgies' (TDT's) SortSpike2 [:33], use this threshold-based spike sorter as

a spike detection step to separate spike waveforms front the raw data.










Currently real time spike sorters ignore the possibility of overlapping spikes due to

the computational complexity of overlap resolution methods. Template matching is the

only method to address overlap and it uses a subtraction method to separate out the

overlapped spikes [27].

All of the mentioned spike sorters require human-tuned parameters which affect

the accuracy of spike sorting. Generally less than three features are used to spike sort

so neuroscientists can view the feature clusters and properly set the parameters. A few

attempts at fully automatic spike sorting have been made but neuroscientists have not

embraced them as they increase sorting error. The human tuned parameters introduce a

variability in spike sorting across different neuroscientists.

A study on the variability of manual spike sorting using human-configured on-line

sorting algorithms by Wood, Black, Vargas-Irwin, Fellows, and Donoghue [34] showed a

wide variability in the number of neurons and spike detected in real data. The number of

spikes varied four fold and the number of neurons was only correct 25' of the time. To

obtain specific error values, synthetic data was used so the ground truths were known.

Average error rates of 2 :' false positives and t::0' false negatives were obtained with the

synthetic data. This variance and error in current spike sorting methods makes it difficult

to compare spike sorters with real data as accurate ground truths are not known. The best

approach currently available is for many expert spike sorters to use off-line spike sorters to

mark a data set and an average taken.

1.3 Neural Data Reduction

Transmitting the raw voltage signals from hundreds of channels is not possible

with the current wireless bandwidth limits and thus data reduction is necessary. As

neuroscientists do not want to limit the number of recorded electrodes, the data reduction

must limit the data sent m each electrode to decrease the overall required bandwidth. The

reduction in information sent must preserve the pertinent information.










Most neurons encode information in spikes, with the exception of retinal neurons. It

is debated as to whether the information is encoded in the rate of spike firings or with

individual spike times, but either way only recording spike times for a single neuron is the

ultimate data reducer and either encoding mechanism can still be evaluated. For single

spiking neurons, signal information is encoded in spike times and not the amplitude or

shape of the spike. However for in vivo extracellular neural recordings, multiple neurons

are recorded from the same channel so spike sorting may be necessary thus requiring

relevant features to be preserved.

BMI systems currently use a human tuned and computationally intensive spike

sorting process, which recovers several individual neural signals from each electrode at

the cost of additional power consumption and increased system size which necessitates

performing spike sorting outside the body. This results in a larger communication

bandwidth because windows of data around possible spikes or features from those windows

must he transmitted compared to only sending out spike times. Recent results -II---- -r

that the spike sorting step may possibly be eliminated without severe degradation of

BMI performance [2, :35] thus lending credibility to solely transmitting spike time for

data reduction for some applications such a low precision BMIls (approach 4 in Figure

1-5). Details are given in Section1.3.1. However this claim has not been exhaustively

tested and neuroscientists will continue to require spike sorting for studying the brain.

For cases where spike sorting is necessary there are data reduction schemes that retain

more information than solely transmitting spike times but at the expense of higher data

handwidths (approaches 1, 2, and :3 in Figure 1-5 and Section 1.3.1).

1.3.1 Data Reduction for Spike Sorting

If spike sorting is required, either features of the spikes (approach 2 in Figure 1-5)

or windows around the spikes (approach :3 in Figure 1-5) can he transmitted while still

obtaining significant data reduction compared to transmitting the sampled and quantized

raw waveforms (approach 1 in Figure 1-5). Of course there is a tradeoff between data










reduction and the amount of information retained. Neural signal data reduction in a

general issue with a variety of solutions being pursued. Popular data reduction methods

which allow for spike sorting include spike detection followed hv different options to reduce

the data.

One option is to wirelessly transmit a sampled and digitized clip of the raw

waveform surrounding the spike for spike sorting outside the subject where power and

size constraints are less stringent [9]. This allows for data reduction while retaining the

use of traditional spike sorting methods on the back-end. The main drawback is currently

complex digital VLSI circuitry is used to store the waveform until spike detection is

performed and perform spike detection. The power requirement of the VLSI circuitry is

currently prohibitive for implantation.

Another option is to extract and send the features themselves for spike sorting [36],

but how to get the features out wirelessly at low-power is problematic as they need to be

quantized and sent in a group with all the features from one spike along with the spike

time. Currently no group has yet to solve this issue.

1.3.2 Data Reduction with Spike Detection

The ultimate data compression scheme for neural recordings is spike detection, where

only spike times or pinned spike counts are transmitted ([37] approach 4 in Figure 1-5).

This greatly reduces the bandwidth required to transmit the neural signals because spike

occurrences are sparse within neural data. Spike detection is also the first step in most of

the data reduction methods which preserve enough information for spike sorting as shown

in Figure 1-5.

Spike detection must he as accurate as possible because missed detection errors

propagate through the system as missed information (missed neural spike). False

detections also propagate through the system as incorrect additional information (false

neural spikes) unless windows of data around the spikes are transmitted. Then, a spike

sorting process can allow some of the windows to be classified as noise to reduce the false










detection error but additional bandwidth to transmit the false alarms is still required.

Therefore, systems that transmit windows of data around the spike often use lower

thresholds to increase detection performance, but the lower the threshold the higher

the power dissipation and required bandwidth which can cause a bottleneck in wireless

data transmission so care must be taken when selecting the thresholds. One solution is

to monitor the bandwidth and adjust the thresholds to maximize its utilization while

remaining within the power limits.

Spike detection is a long standing issue in neuroscience. Popular spike detection

methods include amplitude thl~r. h~ldingr absolute value, energy based, wavelets, matched

filters, and template matching. Currently, there is no consensus in the community as

to the best approach to spike detection, particularly for robust, unsupervised, and

computationally simple methods. Each of the proposed detection techniques have

shortcomings for implanted applications.

Amplitude thresholding is the simplest and lowest in power spike detector since

it is a subset of the other methods for this binary detection problem. It is the easiest

to use with only one parameter to set, the threshold level, and it is the most common

spike detection tool used though it is often paired with additional processing to achieve

acceptable detection performance. For instance, it can be paired with requirements that

the signal pass through two thresholds (sometimes called a hoop) within a certain amount

of time to increase its performance. This can insure the slope is sufficient as it passes the

thresholds or that the threshold crossing is spike like in that the signal rises to cross the

threshold and then falls to cross the threshold again within a certain period of time. The

amplitude threshold detector's performance quickly begins to fail as SNR drops though

and it is not robust to DC drift [27].

The absolute value of the signal can be used to allow for an equal detection rate

of spikes with larger positive or negative amplitudes. The tradeoff is the detector must

be blinded after a detection so that spikes with more than one phase are not detected










twice. The blinding period limits the detector's resolution between two spikes because it

prohibits two spikes front being detected unless they are farther apart than the blinding

period. Obeid and Wolf have reported that though the absolute value spike detection

method does not perform the best among spike detectors it is very cost effective for their

setup in terms of computation cycles, performance, and required bandwidth to transmit

the data [9].

Several groups have tried to use non-linear energy operators (NEO) to detect spikes

[38-40] hut they are too sensitive to noise, so they only perform well for extremely high

SNRs. Recently a new class of niultiresolution TEOs (Teager Energy Operator, a type of

NEO) was presented with noticeable intprovenient over previous NEO's for low SNR data

[41]. The niultiresolution approach allows the detector to impose additional constraints

on the energy function to consider it a spike, which keeps the detector front being as

sensitive to noise. Compared to the previous energy detectors this method requires more

computation since, for nmultiresolution, it requires several TEO's to be evaluated in parallel

and then combined to make a final decision.

Recently wavelets have become popular because they allow for localization in both

the frequency and time domain which is important for transient and non-stationary

signals such as neural data. Several groups have developed off-line spike detectors based

on wavelets [42, 43]. With the increase in PC computational power, algorithms have

been designed for wavelet decomposition to run in almost real-time [44]. Applying near

real-tinle wavelet algorithms to spike detection has resulted in better performance than

existing single scale methods but current wavelet circuits consume too much power for

implantation.

The most complex and possibly one of the more accurate spike detection methods

is matched filters or template matching. If the signal was embedded in white noise, the

matched filter would nmaxintize the SNR and be the most accurate spike detector, but

distant neurons, correlated with the signal, creating non-white noise [45, 46]. 1\atched










filters require stable spike teniplate(s) obtained front a human expert selecting the spikes

in the waveform and taking an average of them to use as a template. For spike sorting a

separate template is required for each neuron and with spike detection a single average

template of all spikes can he used. The template is convolved with the signal and the

result is thresholded to determine the presence of a spike since signals similar in shape to

the template will produce a large response while the noise should produce a much smaller

response .

The problem with matched filters is that the filtering circuitry requires too much

power to be implanted and the spike templates tend to be unstable. An option is to use a

simple spike detector (such as amplitude threshold), with the threshold set low as to not

miss many spikes, as a preprocessor to parse out windows of data around possible spikes to

transmit. Then, outside the body, where more power is available, the windowed data could

be compared to the template and if the window differed too much the waveform could be

disregarded as noise. While a lower threshold increases the final detection performance,

the drawbacks are an increase in transmission bandwidth from sending out more noise

waveform clips to reduce missed spikes.

1.4 University of Florida's Neural Recording Bandwidth Reduction Strategies

The University of Florida has three approaches for neural bandwidth reduction for

different application needs but all three can he intpleniented in low-power analog VLSI

circuitry for implantation. The three different approaches are shown in order of decreasing

bandwidth in Figure 1-6. Neural signals have a rather sparse number of neural spikes with

the rest of the signal noise. As the noise is not important, it would be best to use the

bandwidth on the spike portions of the signal and all three methods take advantage of

this.

1.4.1 Biphasic Signal Coding with Reconstruction

The first approach the University of Florida (ITF) has chosen to reduce the data

bandwidth is to encode the data to reduce bandwidth so that in theory it can he perfectly





















implanted front-end

Approach 1 bio- asynchronous
biphasic pulse
amplifier encoding I


SApproach 2 bio- extractt features
amplifier f or spike sorting


SApproach 3 bio- -\spike
amplifier~ -detector I

elect rodes


Figure 1-6. UF overall neural data reduction approaches.


back-end

reconstruction
then traditional
spike sorting

pulse-based
spike sorting


inning if not
done at
transmission










reconstructed on the back-end and traditional spike sorting techniques can be applied.

This approach was developed by Dr. C'I. in in her Ph.D. studies at the University of

Florida, thus only a brief overview will be provided here but the reader can refer to her

dissertation [47] and conference paper [48] for additional information.

This encoding is done using a biphasic-pulse representation because pulses are digital

which are more robust to noise than analog in wireless transmission yet it is lower power

(100 p-W) than digital because it does not require an ADCs. The bandwidth can be

reduced by more than four times over a traditional ADC sampled system at 25 K(Hz with

12-bits of resolution.

The biphasic signal encoding uses pulses to represent when the integral of the

waveform (its area) surpasses a positive or negative threshold. An example of an input

signal and it's biphasic representation is shown in Figure 1-7. This automatically increases

the bandwidth during spikes as the area is larger and reduces it during noise as the area is

smaller. By setting the area threshold appropriately you can theoretically obtain perfect

reconstruction [48]. In practice, only the spike portions need to be reconstructed close to

perfect so the area threshold could be set only with consideration to perfectly reconstruct

the spike portions of the signal.

A block diagram for the biphasic encoding system is shown in Figure 1-8. If the

output of the integrator, y(t) reaches the positive threshold of the comparator, 8, the

output of the comparator raises and resets the integrator after a short delay, -r, in

the feedback loop. Similarly, if the output of the integrator y(t), reaches the negative

threshold, -0, the output of the comparator drops and also resets the integrator. The

delay, not shown in the simplified block diagram, sets a maximum bandwidth and with

more strict constraints it still allows for theoretical perfect reconstruction [47]. The timing

of two consecutive pulses must satisfy the following equation:


x(A)da = Os (1-1)



























































representation. A) Shows the


1


-0.5


-I


ill


0] 0.2 0.4 0).6 0.6


1
Tlcme (ms)


1 .2 1 A 1.6


Figure 1-7. An example of an input
input signal. B) Shows the biphasic


signal and it's biphasic
representation.






















II


Figure 1-8. Block diagram of hiphasic encoding with integrate-and-fire (IF) neuron.

where 04 E {-0, 0}.
1.4.2 Pulse-Based Feature Extraction

In this work, UF has started to pursue the second approach, a spike feature extraction
method that reduces the required bandwidth even further. 1\ost people optimize the
front-end for data reduction but we prefer to optimize the complete system by considering
the bandwidth reduction effect on spike sporting.
This spike feature extraction method also uses hiphasic pulses to encode the data, but
it only preserves features about the spike and little information about the noise hv not
following the strict constraints for reconstruction. Also, a leaky term, shown as a resistor
in Figure 1-9, is added to allow a greater reduction in bandwidth by subtracting out the
noise as well as providing synchronization (first pulses do not depend on the previous

samples (noise)) for the pulse-train output at the time of the spike.
It allows spike sorting to be directly performed on data that is wirelessly transmitted
reducing the complexity on the back end. The spike sorting uses a traditional method of

template matching but is untraditional because the waveform is pulse trains. The feature


positive pulse output


V to I


negative pulse output














V to I


Vmem


input


Figure 1-9. Block diagram of hiphasic encoding with leaky integrate-and-fire (LIF) neuron.

extractor can reduce the bandwidth more than one-order of magnitude lower than the

UF hiphasic encoding and more than two-orders of magnitude lower than traditional

ADC sampled data at 25 K(Hz with 12-bits of resolution while maintaining a similar spike

sorting error.

This data reduction approach is divided into four schemes based on intplenientation

as illustrated in Figure 1-10. All of the schemes use a pulse-train based spike sorter in

software on a PC that was designed specifically for the feature extraction algorithm.

The first intplenientation scheme uses existing front end data reduction techniques

such as using and ADC and replaces the back end spike sorting software with its

feature extraction algorithm and spike sorter algorithm intpleniented on a PC platform.

Inmplenientation scheme two takes advantage of the feature extractors bandwidth reduction

by placing the algorithm on a digital signal processor (DSP) in in the front-end and

using the same sorter as in approach one. Inmplenientation scheme three is a hybrid

approach and it intplenients the front-end feature extractor in analog to use it's low power

advantage with the same back-end spike sorter. Inmplenientation scheme four is purely

analog both in the front-end with the feature extractor and the back-end with the spike

sorter.


positive pulse output


negative pulse output









Back End



PC
Sorting


PC
Sorfing





Analog
Soraing


Front End


Purely Back-End
Sof twa re


Back and Front-End
Sof twa re



Hybrid




Fully Analog


DSP
Feature
Extraction

Analog
Feature
Extraction

Analog
Feature
Extraction


Figure 1-10. Feature extraction and subsequent sorting intplenientation schemes.

Inmplenientation scheme 1-3 will be presented in detail in ChI Ilpters 2, :3 and 4.
Inmplenientation scheme 4 will not he presented in detail as power savings with an analog
back-end is not currently necessary because the power limits are much greater than in the
front-end.

1.4.3 Spike Detection
The third bandwidth reduction approach ITF has taken (Figure 1-6) is the most
dramatic in bandwidth reduction since only the spike time is transmitted, but it does
not allow for spike sorting [49, 50]. As previously mentioned this may be appropriate for

applications such as in low-precision BlMls [2, 35]. The dramatic reduction in bandwidth
allows more electrodes to be recorded which is helpful in many applications. The ITF
approach for spike detection originated with a handpass filter and evolved to a niulti-scale
spike detection circuit based on wavelets. Both approaches are ultra-low power and robust
while the nmulti-scale spike detector allows for better performance than other simple spike










detectors that could be implanted. This approach will be presented in detail in ('! .pters 5

and 6.

1.5 Dissertation Structure

The development of a low power intplantable circuit for spike feature extraction and

another circuit for detection of spikes which both dramatically reduce the required

coninunication bandwidth out of the skin are presented in this dissertation. An

introduction to extracellular neural recordings and current systems has been presented.

The remainder of this dissertation follows UF's data reductions approaches. Beginning

with three intplenientations schemes for the feature extractor and pule train spike sorter

in ('!s Ilters 2, 3 and 4. C'!s Ilter 2 introduces the novel pulse-based feature extractor

followed by ('! .pter 3 where the bandwidth reduction is examined and ('!s Ilter 4 provides

the details of the feature extractor circuit and chip results. Data reduction with Feature

extraction is followed by data reduction with spike detection where chapter 5 introduces

the novel single-scale spike detector and ('! .pter 6 extends the single scale spike detector

to multiple scales increasing the performance with nmininmal power increase. ('!s Ilter

7 concludes the dissertation with an overview of previous chapters and a suninary of

contributions.









CHAPTER 2
PITLSE-BASED FEATURE EXTRACTION AND SPIK(E SORTING:
IMPLEMENTATION 1 BACK(-END SOFTWARE

2.1 Pulse-Based Feature Extraction

Some of the pulse-based feature extractor work presented in this chapter has been

previously published [51]. The pulse-based feature extractor can he intpleniented entirely

in software, entirely in analog circuitry, or a hybrid with analog circuitry for the front-end

feature extractor and software for the back-end spike sorter. Pulse based spike feature

extraction has been used in this work for low-power and low-bandwidth data transmission.

The circuit intplenientation entails modifying the current spike detector's current threshold

to an area threshold using a leaky integrate-and-fire neuron. Instead of transmitting the

raw waveform, the pulse-based feature extraction method encodes information about the

spike in a hiphasic pulse train. This greatly reduces the bandwidth required to transmit

the spike trains especially because spike occurrences are sparse within the neural data,

while the pulse coninunication offers lower power transmission options such as ultra-wide

hand (ITWB). The encoding scheme uses pulses based on area per time thresholds to

represent the spike while the noise is mostly discarded. Only the spikes and their time

within the spike train contain information so not transmitting information about the noise

saves power without any drop in system performance.

The noise is discarded more severely than when reconstruction is needed by using a

leaky term with hiphasic encoding. This allows the pulse trains for spikes with different

preceding noise to still synchronize which aids in spike sorting. The system diagram is

shown in Figure 2-1 with the leaky term to subtract out noise in the signal. The leaky

value sets the cutoff frequency for the low-pass filter formed with the integrator. This

leakiness along with the proper threshold settings allows for very few pulses to represent

the noise and the 1 in 4 Gly of pulses to contain information about the spikes. The leaky

component changes the equation for the constraint of two consecutive pulses to











i+ -ti l
z(A)e RC' dA = Os (2-1)
Srti +T ~ l

where Os E {-0, 0} and C is related to the integration capacitor and the R is related to the

leak value.

Once the pulse trains have been transmitted, a classifier can then perform spike

sorting outside the body where power issues are not so critical. The encoded pulses

for each spike serve as a spike signature, where a pulse-based spike sorting algorithm is

used to classify the spikes. The classifier would be trained once in the initial setup and

then could be periodically retrained if necessary by sending short segments of the raw

waveforms from one electrode at a time. One of the more difficult cases for this type of

spike sorter would be two spikes from different neurons but with the same area. However,

in this case, the taller and narrower spike would have more spikes in a shorter time period

so the two would have different spike signatures and could still be distinguished.

2.2 Data Reduction with Pulse-Based Feature Extractor

The feature extractor employs a leaky integrate-and-fire neuron to produce its

hiphasic pulse representation. A block diagram for the hiphasic encoding system is shown

in Figure 2-1. If the output of the integrator, y(t) reaches the positive threshold of the

comparator, 8, the output of the comparator raises and resets the integrator after a short

delay, -r, in the feedback loop. Similarly, if the output of the integrator y(t), reaches the

negative threshold, -0, the output of the comparator drops and also resets the integrator.

The leak term is a fixed value to filter out noise. The leak value sets the cutoff frequency

for the low-pass filter formed with the integrator. This leak value along with the proper

threshold settings allows for very few pulses to represent the noise and the 1! in 4 Gry of

pulses to contain information about the spikes. The timing of two consecutive pulses must

satisfy the following equation:

C1 t + 1 ~l
x(A)e RC dA = Os (2-2)
























Figure 2-1. Block diagram of hiphasic encoding.

where Os E {-0, 0} and C is related to the integration capacitor and the R is related to the

leak value.

2.3 Spike Sorting with Pulse Trains

For our feature extractor the signals are hiphasic pulse trains thus traditional spike

sorting algorithms can not he directly applied. The encoded pulses for each spike serve

as a spike signature, where a pulse-based spike sorting algorithm is used to classify

the spike. While distortion metrics for spike trains have been studied in areas such as

neuroscience and genetics, many methods are computationally complex and far from real

time such as the edit distance [52]. Another idea is to low-pass filter the pulse trains with

a function, such as an exponential, so more traditional signal processing can he applied

[53]. Instead of trying to reconstruct the signal, the spike sorter used for the pulse-based

feature extractor similarly convolves the pulse train with a Gaussian function, where the

a determines if the detector is more of a coincidence detector (a much smaller than the

interpulse interval) or a pulse count detector (large a). A Gaussian function was chosen as

it is more concentrated around the peak allowing the a to better control the detector type.

Once the pulse train is convolved with the Gaussian, it is then compared to each user

defined neuron template. The template with the lowest MSE is a match unless it exceeds

the maximum allowed MSE and then it is considered noise.


leak comparator


delay










Examples of the spike signatures for the six neural simulator neurons spikes (the

neural simulator data is explained in detail in Section 2.4.1.1) are shown in Figure 2-2.

As previously mentioned, comparing two pulse trains is computationally expensive so

the signature is convolved with a Gaussian to allow traditional template matching signal

processing techniques to be applied. An example of the signatures once they are convolved

with a Gaussian for the neural simulator data are shown in Figure 2-3

To show the importance of selecting the Gaussian a, Figure 2-4 shows the error versus

a value on the neural simulator data set. If the best detector was somewhere between a

coincidence detector and spike count detection, the curve would have a U-shape with a

sweet spot for a to where the distance between the pulse trains is somewhere between a

coincidence detector and a pulse count detector. The six neurons in the neural simulator

are all very distinct and the noise is low so it do not require any coincidence detecting to

classify the spikes. Thus, the curve is more like half a IT with the right end flattening out

because as a continues to increase (less and less coincidence detector) there is little change

mn error.

There is a problem with using a single value of a. Spikes with different amplitudes

have different interspike intervals and in fact within a single spike the interspike interval

changes since it is part of the signal encoding. The problem is that the value of a which

corresponds to a coincidence detector or a pulse count detector depends on the interspike

interval of the pulse train. Thus, for a single spike one a values means part of pulse train

distance will be computed using one type of detector while other parts of the spike will

be more toward the other type of detector. This is a problem because the detector type

changes based on inter-pulse interval which is useful as a feature. A useful variance of a in

time might he for the detector to become less of a coincidence detector towards the end of

the spike because the leak value only synchronizes the beginning of the spike and by the

end of the spike the accumulated noise will cause the later pulse times to deviate move.

Three different vwei~ to set a will be discussed.





















,r
o
ka

O
~
c~a
o
r
m C~
k
o




m m
rb 3
~ t~
k k






3
%
k O
cb


a

9'Z1
c~ k


m
E
k C~
k

r
% O
k cb


a
x
ma

E "
O m
k 3
cc~ r
O
m k
k O
~
cb

5DX
"




"'63

Ocb~
r
cb-be
O
k

'~ aa
m


~ cb




c~
Dl~k
~mc~
k~cb

5De~5D


Pmu
O


(Smu
O


DO


O
(


WLOOD
0 O


WOOO
00


O
(x


O
X


O
(D
X


dd E
~morn~ I


mom L
Sd d
I


O

IY) X


LOOL
00


LO0L
00d


O
(D
X


a


a e o m


L o m


O O 0 O 0
I I r


I


L 0O 0


LOOL


LO0L


o~ o
















P P








to















~~o






to a o l r-r o al










oo a








to a 3o--
o 9 I r










to a t r-r o o r o f




oo o












40

35

30

25

20

15



10


1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8

[h] sigma value x 10-6

Figure 2-4. Spike sorting error as a function of the pulse train distance Gaussian o-.


One solution is an adaptive leaky integrate-and-fire (LIF) threshold to create a

more uniform pulse rate and is inspired from the biological neuron's adaptive threshold

mechanism to keep the firing rate from saturating and information being lost [54].

Another solution is the addition of a refractory period which does not allow the LIF

to fire another pulse until after a certain period of time which sets a maximum firingf rate.

In this case though the signal is ignored during the refractory period so information is lost

and presumably the adaptive threshold method preserves more information and is thus

more desirable. The third option is to change the o- value according to the spike template

interspike interval to produce a more constant detector across the spike.

The focus of this research is feature extraction not spike sorting, but in order to

analyze the performance of the feature extractor spike sorting most he performed. Thus,

the spike sorting procedure was kept simple to purely show the feature extraction has

potential. Results in this paper were obtained by simply using the first spike from each

neuron as a template. This is a worse case template formation because often an average

over several spikes is taken to eliminate noise. A second or third spike template could be










used to estimate the pulse jitter statistics that could be used to set the o- of the Gaussian.

As the feature extractor reduces the noise level, using a single spike as a template proved

adequate in this data set.

2.4 Matlab Simulations Results

2.4.1 Data

Three data sets were used to evaluate the feature extractor. One data set came from

a neural simulator so the ground truths are known and another is simulated data using

invivo spike shapes and noise so those ground truths are also known. The other data was

recorded in vivo from a rat and an expert marked the sorting ground truths using Spike2.

2.4.1.1 Neurosimulator data

The pulse-based feature extractor algorithm was tested with neural recordings

from Bionic's 128-channel hardware neural signal simulator. The use of a neural signal

simulator allows the ground truths, the time of each spike and which neuron it came

from, to be known. The neural simulator outputs a repeated 11 s pattern of spikes from

three different action potentials with amplitudes of 100 pV 150 pV and a width of 1 ms.

The interspike interval is 1 s for 10 s and then reduces to 10 ms for 1 s of burst firingf.

To increase the number of neurons on one channel the reference was chosen as another

channel instead of ground. The referenced channel was carefully chosen to be a 5 ms

d. I we- II version of the first channel. In this manner, the simulated neural signal contains

spikes from six different neurons with no superimposed spikes which are not addressed in

this work since they are problematic for all spike sorting algorithms.

The UF bioamplifier [25], with a gain of 100 and a low cutoff frequency of 0.3 Hz and

a high cutoff frequency of 5.4 K(Hz, was used to amplify the neural simulator output. The

amplified signal was then digitized at ~ 24.4 K(Hz and 34.6 s were captured with a digital

logic analyzer. The average spike firing rate for the data set is 19 Hz. The signal's SNR is

about 30dB. A portion of the signal during bursting with all six neural spikes is shown in

Figure 2-5(A).

















a

cb~~

o, oa

'D E
O
Okg
co o~
[d
H
~ cll~,
E
~ a,
c~km
c~

k m
cc~ Ln
~Ln
C~ bD~~

r
rr k
r c~
Ln
[d ~ rl ."3
kOr
m

(d ~ '"
aa~

cb
cb~r
i

or
-
hj
cb o

CbC~H
'r cll
c\r "




(T, ~m g
Xkk
.~ O CC~

"~a
v,
co E
~ O
c~
a, d
E
~ Cb~k



~~~~Uo
C~mg~
Cb
r m
~~~~

ae-
cb~n ~
k
Inc~B
~cc~mO
O ~ O
Irj;?~N
r O
01 ~ kO
k'

harl~~
~" cb ~~
~em~


Q m


O\w> apnUldw~










2.4.1.2 Caltech simulated data

Rodrigfo Quian Quirogfa posted a database of simulated neural signals at various SNRs

as well as with different waveshapes while he was at Caltech. The database is referenced in

his paper [28] and available online at http://www.vis. caltech. edu/~rodri/data. htm.

The synthetic signals were constructed using a database of 594 different average

spike shapes compiled from recordings in the neocortex and basal gangflia. To mimic the

background noise generated by the activity of distant neurons, spikes randomly selected

from the database superimposed at random times and amplitudes for half the times of

the samples. Then, three distinct spike shapes (also preselected from the same database

of spikes) were superimposed on the noise signal at random times. The amplitude of

the three spike classes was normalized to have a peak value of 1. The noise level was

determined from its standard deviation, which was equal to 0.05, 0.1, 0.15, and 0.2 relative

to the amplitude of the spike classes. Spike times and identities were saved for subsequent

evaluation of the clustering algorithm.

The sample rate is 24 K(Hz. In all simulations, the three distinct spikes had a Poisson

distribution of interspike intervals with a mean firing rate of 20 Hz with a 2 ms refractory

period between spikes of the same class. There are a total of five data sets two easy, two

hard, and one bursting. The more difficult easy data set Easy2 was used with a noise level

of 0.15. The data set is shown in Figure 2-6 and will be referred to as the Caltech data

hereafter.

2.4.1.3 Rat data

The feature extraction algorithm was then tested on neural recordings from male

Sprague-Dauley rats chronically implanted with 50 pm polyimide insulated tungsten

microwire electrode arrays in 1 e. rrV of the forelimb region of the primary motor

cortex. The data was sampled at 25 K(Hz and bandpass filtered between 0.5 and 12 K(Hz

using hardware from Tucker-Davis Technologies. Action potential widths ranged from





















"fine seo

Fiue26 egeto h alehdtst













tofirst parse oudt segment s wite ateh aa pssbe sie hswsdn uigacnevtv


Then thse. msegmets wuere eam ined and only thos which actall containedaspk were hs t

kceptandlblldi the spikeceofte timen file. This eas thmerewr efleaamsbtsie


wyith larg negaive peakscoul wihav beien fomlwitte ifteiec ond phase dridf elntot cossth

pstive thesol. e determined that in opk2iseaned othee dtae setso ove thfrst ive spieconds

to fis as data 18 pies itth s crteia hapsil k. Thuwechrtrizin the featue extrga ctor's srting


pheromnc hee, thgeerocmpnntswr false aam and mnytoewisse detectiy ons, ill e analyzedwr






separately. This will allow false alarms from missed detections in Spike2 (our ground

truth) not to count as an error.










To increase the ground truth accuracy and remove any bias of an individual expert

several experts should mark the ground truths with their averaged results becoming

the ground truth. Marking spike times is a tedious task though, so it is difficult to

enlist human experts to mark a lot of data. For the parsing out of possible spikes a

second negative threshold could be used in addition to the positive one to decrease the

number of missed detections in the ground truths but this could also decrease the sorting

performance .

Figure 2-7 shows the original neural data waveform. SNR was calculated for the

signal in terms of power using an averaged spike shape.

x 104 x 104





-1I i -1





0 5 10 7.22 7.24 7.26 7.28

Figure 2-7. Neural waveform recorded from rat003. Column two is zoomed in from column
one.


2.4.2 Spike2

:1~.i -' [55], a popular commercial program first written in 1988, which can spike sort

offline, is used as a comparison to the feature extractor's spike sorting performance. .*1~.

first performs crude spike detection by capturing windows around events that cross a user

defined threshold(s). Then, spike sorting is performed with a combination of template

matching and a PCA based cluster cutting. templates from the neural simulator data are

shown in Figure 2-8 and PCA analysis shows the well separated six classes in Figure 2-9.

This process requires the user to select many parameters during the template setup such

as the number of templates and allowable variation within the template. :1~. ~' provides










the user with a interactive visual di pl w~ to assist in setting the spike sorting parameters.

The parameters were set by an expert in the field with the same procedures used in typical

experiments.

2.4.3 Spike Sorting Results: Neurosimulator Data

Mattlesb was used to simulate the pulse-based feature extractor and its spike sorter.

The LIF was set with a threshold and leakage value such that its spike sorting error was

minimal resulting in an error of 0.5' compared to Spike2's 6.1 Over a wider range of

threshold and leakage values the feature extractor obtains < ;:' error. The percent error

was calculated using Equation :31. Figure 2-5(B) shows examples of the hiphasic output

for spikes from two different neurons. The regions between spikes did not have any pulses.


E missed spikes + false positives
.error = (2-3)
total number spikes + false positives

There are several reasons for the feature extractor's outstanding performance. First,

the data set has a high SNR with distinct spike shapes thus it is a relatively easy spike

sorting data set, but this is also true for Spike2. Second, the feature extractor has fewer

parameters to set than Spike2 so with limited data is could be better optimized. This

is why this is not the only data set used to an~ lli. the feature extractor's performance.

Third, the feature extractor utilizes the leaky parameter to reduce noise which leads to

increased sorting error and its features are robust to noise.

2.4.4 Spike Sorting Results: Caltech Simulated Data

M~atlesb was used to simulate the pulse-based feature extractor and its spike sorter.

The LIF was set with a threshold and leakage value such that its spike sorting error

was minimal resulting in an error of 21n' This was compared with the average of three

neuroscientist's sorting results using Spike2 of 15' One of the neuroscientist sorted

the data extremely well, while the other two had much larger error. This is within the

expected variability in manual spike sorting [:34].









SWaveMark from neuro sim dat


File Edit ViIEW Templtes Control Analyse

IEditing marker code 0 IIlllk~ ~ I~ J = J1 1 1a

0.0104 untitled
SReclassify...
0. 005- 1mImIm m
0.0- -- Best match : 01

-0.005- ._.Fast

-0.010- Circular replay
re Make templates
ms -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 Ij~j by shape


01


935 166/83=13 a1 A


220.2110 = a 0 2010=O


04 158/79=17 a1









re 179/79 =IT 1


Figure 2-8. Spike2's templates for neurosimulator data.


0-1~

































~


1'Principal Compor~nt~



i;
1111. i)ir(~)L I~III~: Ij~ I~'J [ r~-
u


Figure 2-9. Spike2's principal component analysis (PCA) for neurosimulator data.










There are several reasons for the feature extractor's poorer performance with this

data set. The spike shape are more similar this time which is why the neuroscientists also

has higher error. This is shown with the true templates, Figure 2-10. Spike2's templates

and PCA are shown in in Figure 2-11 and Figure 2-12 respectively. And the templates

from the feature extractor are shown in Figure 2-13.


x lo-3 Pile Plot of Class Spikes


10r~ 10 I



0 0.5 1 1.5
x 10-
x 10-


0 0.5 1 1.5
Times)x 10-

Fiur -1.Acul lssfidspks o teCatehsiuatddaa








the sae 2-0 acrsthe temlates, itd sis too lare Cat the peaks tod comenatefrnoh





shorter amplitudes. So at the large peaks, the fine details are lost and it becomes

difficult to differentiate between classes only using that portion of spike. Analyzing

the misclassifications, shown in Table 2-1, confirms the single sigma a issue as often class 1

or 2 are misclassified as class 3.

The second problem is template selection. If the variability in the spike waveshapes

could be accounted for in the pulse domain with an average template, this could lower the


















































..


? aveldark fr~om C Easv noise015


'I lLI


]-1. 1 ,9 in;U11


O I>lu









Figur 2-1. Spke2' tempatesfor he Cltechsimuateddata


P rincipal l.orptonerus
Fle Edit Vlewv Cluster


~-
~~'
.5
:1~
"


PCA, of 3463 e enur rb 28 ,,-.11 press F I or help


Figure 2-12. Spike2's principal component analysis for the Caltech simulated data.

















-vI


-


0 2


2

0


8
x 104


0


2

0
-1


1 2 3 4 5
3 x 10i


1 2
Time (s)


3 4
x 10d


Figure 2-13. Feature extractor's templates for the Caltech simulated data.


Table 2-1: Feature extractors misclassifications.
classified as
neuron 1 2 3
actual 1 11- 5 33
neuron 2 5 I- 56
3 20 1










error. This could be incorporated by separating the class into several templates to account

for the variation in shape and then rejoin them after the classification step. Just choosing

the second spike as the template for neuron 3 reduced the sorting error by 5' showing if

the first spike happens to be far from the average spike shape your results decrease.

2.4.5 Rat Data Spike Sorting Results

As others working in this area have stated, it is difficult to test a spike sorting

algorithm on a real data set because the ground truths are not known. Thus, the new

algorithm is penalized for its errors as well the ground truth errors if it get those correct.

Therefore, some groups have solely focused on developing realistic simulated data where

the ground truths are known with complete accuracy. As the feature extractor was

designed to eventually be implanted as part of a neural recording system, it is necessary to

analyze its performance with real data compared to current spike sorting systems.

The "ground truths" are used as a first pass and then the false alarms and misclassifications

are examined to see if they were correct and the ground truths were wrong or if they are

indeed error. Statistical analysis can he used to see if the "errors" are indeed errors (for

example PCA can he performed to see how well the ground truth classes are clustered

etc.) but for the most part it is a combination of statistical analysis as well as an experts

gut instinct from years of experience. The terms error, misclassification, and false alarm

will be used rather loosely as the "ground truths" are not absolute in themselves and

contain error.

The rat data spike sorting "ground truths" obtained from an expert in the area,

using Spike2. Some missed detections occurred because only a single positive threshold

was used so there are missed spikes which can he picked out just by looking at the data,

but the positive portion of those spikes was attenuated due to noise. The templates were

formed and refined using PCA. The results templates are shown in Figure 2-14 and the

PCA clusters are shown in Figure 2-15. Each color corresponds to a different class with

black being spikes that were classified as noise. It is difficult to tell how well isolated












the PCA clusters are in 2-D but Spike2 allows the user to rotate the figure to analyze

it better. The red and blue classes and the cyan and magenta are the closest classes so


there is likely some misclassifications between those pairs due to noise in the signal. As


the SNR is poorer than the neural simulator data and the spike shapes between classes are

more similar, there is less distinction between the classes making it a more difficult sorting


problem.


_1=I I
,, I i I~ Y1 111
II I~r L_1 LLYI

j~lll.llllltj ~j~~-;Lj

Ft.:ij II

~i m :m :mj
:BE~St matck: 02


Fj



I' I ''''''.1' I 'I' I I 1 '' .1,' I '1,'' 1 ''' I':C'' '' 1 ''' I' I'',' I''':'' I' ~ t .j~;
u.l a2 D:4 0:5 11~
I:11~F-13Irf$18 ~


Figure 2-14. Spike2 example template for rat data.



Histogframs of times between spikes can also be plotted to see if more than one


neuron was classified into the same class. Because absolute refractory periods are known


Me 6dit yig'N 7"9mplates Ewa~b-ol: Inji',.t




II II II

I I I II II II I

i=~l I II II II II II I
:'- ' '

-I I I II II II I




I''I ':'I I ( r I '''l(r t I I I r r 1 ''''' I'''
I:I 1 -0:1 .lj


:ml



























__


File Edit View Cluster
















: **.























PCA of 13297 events wyith l7 points. pre-s F1 for help


Figure 2-15. Spike2's PCA analysis for rat data.



















60


11 l


_J


3'Principal Components










to be about 1 ms, if there are any spikes that fire within the refractory period of the

neuron then those must be misclassified. The converse is not true as two neurons could be

clustered together but never fire with in the refractory period of each other. Figure 2-16

shows the histogram for the sorted rat data.













Figue 216. pik2's istgramanaysisforrat ata

The difrec betee Spke clsiicto and th fetr xrco i .err
1EBecause th ground1~ trts hae ero thmevs hsdfernei ntncsarl ro
an utb xmndi ifeetmne.Afrtcmaisni osei h epae









lNext the correctly clssfedsiks il eexmie. ie losofSik2' ore








toue -6 Spike2's reulsfrmthera fature i exrcor are shw i iur -2.Ntienern
























I 1


, o =


Fllt E-St ulc*. Ttllrplj t c, C.:.rltr..1~ Anl' l a


II I II II II Ic








n I


----


. . . I : I
I -c as -G; 02 1 11J


I:oa =9-~lr 81 rM


S10 1= d lIt 1 n= eli]


Figure 2-17. Spike2's templates.


I


-~ iI 1 5 untitled j



'a 1


i le t lar..p l


= l3re~~r













I I I I




i~Vu~


3 1 2 3 4 5
X 10-4



2


0 1 2 3 4 5
x 10-4














0 1 2 3 4 5
x 10-4


3 1 2 3 4 5
X 10-4



6


0 1 2 3 4 5
x 10-4


1 2 3 4 5
X 10-4


Time (s)


Figure 2-18. Feature extractor templates.


































neuron 1 2 3 4 5 6 tot
1 4 2 48 65 119
2 0 1 0 1 2
3 197 0 5 6; 208
4 3 2 0 2 6;2 6;9
5 42 0 241 2 -11 296;
6 8 3 1 3 28 -43
tot 250 5 246 8 83 145


2,3, and 6 appear cleaner in the feature extractor while neuron 1 has more outlying spikes.

This could mean the classification for the feature extractor is tighter on neuron 2,3, and 6

but looser for neuron 5.

There are three types of errors to examine: misclassification, false alarms, and missed

detections. Table 2-2 shows the difference in classification between Spike2 and the feature

extractor (misclassifications). The table shows the pulse-based feature extractor and sorter

misclassifies certain neurons more than others. For instance neuron 2 and 4 never receive

misclassifications so their templates and rules must he distinct enough from the other

neurons. However neuron 3 had many of its spikes misclassified at neuron 1 and neuron 5

had many of its spikes misclassified as neuron 3 so these classes need better separability.


Table 2-2: Feature extractors misclassifications compared to Spike2.
classified as


actual
neuron


To try and discern which misclassification may have been legitimate (meaning were

actually an error with Spike2 or just distorted by noise enough the spike resembled two

templates) pile plots of the true neuron spikes with all of the spikes misclassified as from

that neuron are examined. Figure 2-21 shows pile plots of the feature extractor's correctly

classified spikes overlaid with the spikes misclassified as from that neuron in the color

neuron they actually belong to. 1\ost of the spikes misclassified as 2 and 5 reside close

to those detected as 2 so they may be errors from Spike2. However, most of the spikes

misclassified as 1 and 6 differ enough from those detected to show the rules for those

classes are too loose.















x 10-5


5 x 10-


10~ 10



0 ~ I0



0 1 23 456 01 23 45 6
x 10-4 x 10-4





x 10-5 3 x 10-5 2
10~ 10

5~ 5

0 0



0 1 23 456 01 23 45 6
x 10-4 x 10-4





x 10-5 1 x 10-5 6
10~ 10

5~ 5

0 'R0



0 1 23 456 01 23 45 6
x 10-4 x 10-4

Time (s)





Figure 2-19. Pile plots of Spike2's sorted spikes.













x 10 5


x 10-


x 10-4

x 10-5 3
10












x 10-4

x 10-5 1


0 1 23 456
X 10-4

x 10-5 2
10








01-




x 10-5 6
10












X 10-4


0 12 34 5 6
x 10-4


Time (s)


Figure 2-20. Pile plot of feature extractor's correctly sorted spikes referenced to Spike2's
results.














x 10 5


x 10-





0 12 34 5 6
x 10-4

x 10-5 3


0 1 23 456
x 10-4


i
~)


II


-5i


x 10-4

x 10-5 1
10













x 10-4


0 1 23 456
x 10-4


x 10


0 1 23 456
x 10-4
Time (s)


Figure 2-21. Pile plot of feature extractor's correctly sorted spikes with those misclassified
as that neuron overlaid with a dashed black line.










Figure 2-22 shows pile plots of the feature extractor's correctly classified spikes

overlaid with the spikes from that class but misclassified as another class in that classes

color. These pile plots shows how the feature extractor's errors relate to Spike2's

classification. For example for neuron 3 the spikes the feature extractor misclassified

different enough they may have come from another neuron. Those spikes from neuron

2 misclassified by the feature extractor were at the edges of the pile plot so likely if the

feature extractor rules for class 2 were loosened or the rules for class 4 were tightened

those misclassification would disappear.

This analysis shows the feature extractor and sorter's parameters could be tweaked

to improve performance but as the desired result is to outperform Spike2 (as was shown

possible with the neural simulator data) only experts can objectively state which sorting is

acceptable.

2.4.6 Future Work

Several areas of the feature extraction algorithm need to be further studied in a

different work to improve sorting performance while decreasing bandwidth. They are listed

below:

* a selection: The inter-pulse interval of each signature varies based on the waveform.
If the sigma remains constant, the type of detector (coincidence or pulse count)
changes as the inter-pulse interval changes. Implications of not adjusting the
Gaussian's a as the pulse rate changes need to be examined while methods to
adjust a explored. Also, it could be advantage to purposefully adapt a making it
more of a coincidence detector at the beginning of the spike when the pulses are more
synchronized from the leaky component.

* Methods for automatically setting leakiness and threshold for optimal
performance: It is believed the optimal leakiness value is related to the noise floor
and the optimal threshold value to the spike amplitudes. This needs to be studied an
automatic method to set these values developed such as based on SNR of the signal
the leaky and threshold values are set.

* More extensive performance measures: The pulse-based feature extractor
and spike sorter need be he tested on more extensive data sets with different SNRs
and spike shapes to ensure its performance across data. Also, a better analysis














x 10 5


x 10-


0 12 34 5 6
x 10-4


0 1 23 456
X 10-4


x 10 5


x 10-


x 10-4


0 1 23 456
X 10-4


x 10 5


x 10-


X 10-4


0 12 34 5 6
x 10-4


Time (s)


Figure 2-22. Pile plot of feature extractor's correctly classified spikes overlaid with the
spikes from that class but misclassified as another class in that classes color



69S










of performance on data sets without ground truths is needed to evaluate in vivo
performance .









CHAPTER 3
PULSE-BASED FEATURE EXTRACTOR AND SPIKE SORTING:
IMPLEMENTATION 2 SOFTWARE FRONT-END AND BACK(-END

Moving the feature extractor to the front-end allows for a dramatic decrease in

bandwidth though it is not without effect on spike sorting performance. That relationship

is examined in this chapter while the principles of operation for the feature extractor and

sorter are not repeated from C'!s Ilter 2.

3.1 Bandwidth Parameters

The front-end of the feature extractor is responsible for sending out pulses which

represent the best features for spike sorting with minimal bandwidth. It is a tradeoff

between spike sorting performance and bandwidth. In the pulse-based feature extractor

three parameters determine the bandwidth, the leak value, threshold values, and the

integration time constant. The integration time constant is normalized to a value of

10 pF and the effect of the leak value and threshold value on bandwidth and spike sorting

performance is explored.

Two extremes are to set the leakage so high that none of the signal passes through

or to set the thresholds so high that the signal never surpasses it to send any pulses. In

either of these two situations no information is preserved. Another extreme is to set the

leakage to zero and the thresholds very low and this preserves all of the information, more

than enough to allow for perfect reconstruction on the back-end so bandwidth is wasted.

Increases in the leakage and threshold values both decrease the bandwidth as shown

in Figure 3-1. While more than one leakage and threshold value will give the same

bandwidth they do not necessarily provide the same sorting error, described in Section

3.2.1 and is desirable to minimize, as they do not preserve the same information as shown

in the plot of Figure 3-2. Figure 3-3 shows sorting error versus minimum bandwidth for

the leakage and threshold combinations at three different SNRs. This plots show the

inverse relationship between bandwidth and sorting error. It also shows, that as SNR

decreases the sorting error increases as expected. The lower the SNR the less pointwise


















S1.1 O






0.709 ".~'o


3.=5 4 4.5 5 5.5 6 6.5 7
Leakage x 101

Figure 3-1. Bandwidth (pulses/s) changes for threshold and leakage values parameters and
integration capacitor at 10 pF.

separation there is between pulse trains of different neurons meaning it is easier to

misclassify one as the other increasing the errors. However, the performance of all spike

sorters drops as SNR decreases.

3.2 Matlab Simulations Results

The neural simulator data was used so the ground truths are known. For details

about the neural simulator data set please refer to Section 2.4.1.1.

3.2.1 Spike Sorting Results: Neural Simulator Data

Matlab was used to simulate the pulse-based feature extractor and its spike sorter

as in Section 2.4 but not also in regards to bandwidth. The LIF was set with a threshold

and leakage value such that its spike sorting error was similar to Spike2's which resulted

in a bandwidth of 455 pulses/s. Figure 2-5(B) shows examples of the biphasic output for

spikes from two different neurons. The regions between spikes did not have any pulses.

The biphasic output was then spike sorted with the results shown in Table 3-1 along










bogibw)

7.4

120 .--- 7.2
S100
7
Lu 80

S60 --- 6.8

S40 ---. 6.6

S20 <1
:':":...,-~- .. I~.' 6 .4
'1.4 '
1.2 7 6 2


Thresho d Leakage


Figure :3-2. Spike sorting error changes for threshold and leakage values parameters and
integration capacitor at 10 pF. The color represents log(bandwidth) (pulses/s)


with the results from S1.i -l :- The results are divided into each neuron class and the

percent correctly classified (true positives, tp) and the false positives (fp) which are spikes

incorrectly classified as from that neuron. The best case is 100I' tp and (I' fp. The

percent error is calculated with equation :31.

E missed spikes + false positives
.error =(:31)
total number spikes + false positives

As Table :3-1 shows, neuron 2 (the smallest spike) is one of the hardest to classify

for both sorters. Neuron 2 is more poorly classified with the feature extractor at the

lower bandwidth, 455 pulses/s because it did not have enough pulses to represent it

and some information was lost. The addition of an adaptive threshold [56] would help

to even out the number of pulses for different amplitude spikes to keep more similar

amounts of information for all spikes without having to increase the number of spikes

for all neurons as in the result with 680 pulses/s. The adaptive threshold will create a



























SNR=14.8dB no added noise
SSNR= 12.3dB, atn =.75
0 SNR=10.3dB, atn=.6










Spike2 performance, no added noise


500 600 700 800 900 1000
Bandwidth (pulses/s)


1100 1200


Figure 3-3. Spike sorting error as a function of bandwidth for three SNRs.










more uniform sorting performance across neurons without having to increase bandwidth

as significantly. The adaptive threshold is inspired from the biological neuron's adaptive

threshold mechanism to keep the firing rate from saturating and information being lost

[54]. Another solution is the addition of a refractory period which does not allow the LIF

to fire another pulse until after a certain period of time. In this case though the signal is

ignored during the refractory period so information is lost and presumably the adaptive

threshold method preserves more information and is thus more desirable.

Table 3-1: Spike sorting performance percent error.
Feature extractor Spike2
455 pulses/s 680 pulses/s 300 K~bps
neuron .I tp fp errors tp fp errors tp fp error
1 100 0 0 96.4 0 3.6 100 0 0
2 69.7 2.6 31.5 93.6 0.9 7.3 89.9 0 10.1
3 100 0 0 99.1 0 0.9 90.8 5.7 13.9
4 97.3 0 2.8 96.4 0.9 4.6 97.3 0 2.8
5 96.4 1.9 5.4 99.1 1.80 2.7 99.1 6;.8 7.8
6; 100 0.9 0.9 100 3.5 3.5 98.2 0 1.8
avg 93.9 0.8 6;.8 97.4 1.2 3.8 95.9 2.1 6.1


Overall at 455 pulses/s the feature extractor had 6~ ~' error compared with S.1~.

which had 6.1 error. While maintaining a similar classification error to traditional

sorting with S.1~. .', the feature extractor requires much less bandwidth with only 455

pulses/s compared to 300 K~bps for a traditional 25 K(Hz sampled signal at only 12-bits.

1 pulse/s is equivalent to 1bps. UF's biphasic output for reconstruction on the back-end

would require 71.9K( pulses/s. The pulse-based feature extractor can reduce its bandwidth

even further if more sorting error can be tolerated or increase its bandwidth to lessen

sorting errors. The two are inversely related.

At 680 pulses/s the feature extractor actually has less error than S.1~.i for this data

set showing it is competitive. More data simulations need to be performed across different

SNRs and data sets to see if the trend continues, but one possible explanation is that the

feature extractor preserves the important information in distinguishing between spikes










while eliminating extraneous information. Extra information can make it more difficult for

a neuroscientist to optimally set the spike sorting parameters in Sr1~. making it harder

to distinguish between spikes from different neurons.

A summary table comparing different data reduction techniques and their affect on

sorting error are shown in Table 3-2

Table 3-2: Bandwidth reduction sorting error comparison
ADC Biphasic with Biphasic Biphasic Spike
reconstruction feature feature detection
extraction extraction
Front-end 300 K~bps 72 K( pulses/s 455 pulses/s 680 plulses/s 2n1l'1'-
bandwidth channel bw
back-end
spike sorting 6. > 6.1~ 6.;' I:s N/A
error


The addition of an adaptive threshold [56] would help to even out the number of

pulses for different amplitude spikes to keep more similar amounts of information for all

spikes without having to increase the number of spikes for all neurons as in the result

with 680 pulses/s. The adaptive threshold will create a more uniform sorting performance

across neurons without having to increase bandwidth as significantly. The adaptive

threshold is inspired from the biological neuron's adaptive threshold mechanism to

keep the firing rate from saturating and information being lost [54]. Another solution

is the addition of a refractory period which does not allow the LIF to fire another pulse

until after a certain period of time. In this case though the signal is ignored during the

refractory period so information is lost and presumably the adaptive threshold method

preserves more information and is thus more desirable.

3.2.2 Future Work

In addition to the future work items in C'!s Ilter 2, several areas of the feature

extraction algorithm need to be further studied in to improve sorting performance while

decreasing bandwidth. They are listed below:










* Adaptive threshold versus refractory period: To equalize information across
spikes regardless of amplitude either an adaptive threshold or a refractory period
can he intpleniented. The adaptive threshold adapts the LIF coniparator threshold
proportionally to the output pulse rate in a similar manner at real neurons due to
increase dynamic range and keep from saturating and information being lost. The
refractory period keeps the LIF front firing for a certain time after each pulse. It
is simpler to intplenient but information is lost. Analysis must he done to examine
the tradeoffs between the techniques in terms of bandwidth, circuit area, power
consumption and their affect of spike sorting performance.

* How to optimally set leakiness and threshold values to optimize both
bandwidth and performance: Different applications will have different constraints
for bandwidth and performance. It would be good to have a simple way to take
a desire value he it bandwidth or spike sorting performance of limits for both the
bandwidth and spike sorting performance and automatically optimally set the
leakiness and threshold values.









CHAPTER 4
PITLSE-BASED FEATURE EXTRACTION AND SPIK(E SORTING:
IMPLEMENTATION 3 HYBRID

While a purely software intplenientation shows promise in comparison to Spike2,

a 1!! I r~~ advantage of the pulse-based feature extractor is that it can he efficiently

intpleniented using compact low-power analog hardware. This is done with approach

three, a hybrid solution, an an analog feature extractor in the front-end and a software

spike sorter at the back-end.

4.1 Circuit Design

An overview of the feature extraction algorithm was provided in C'!s Ilters 2 and 3.

Now the details of the circuit will be presented. The chip was built using the ANTI 0.5 pm

C \! OS technology.

4.1.1 Circuitry

The feature extraction algorithm was developed amenable to low-power analog

circuitry so it can he implanted as part of a neural recording system if desired. The

integration time constant is realized as a capacitor and is set to 10 pF at fabrication time.

The leakage value and threshold values can he adjusted after fabrication for the desired

bandwidth and spike sorting error.

The feature extractor circuit, shown in Figure 4-1, takes a current input and encodes

the neural signal's shape in a hiphasic pulse train using a leaky integrate and fire (LIF)

neuron, a simple extension (such as adding a G, current source or a resistor in parallel to

the integrator capacitor for the leakiness) of the hiphasic IF neuron [48]. The LIF neuron

integrates the signal and then produces a positive pulse when the integrated signal rises

above one threshold and a negative pulse when it falls below a second threshold. The

leakiness of the LIF sets an area per time threshold to filter out noise while preserving the

spikes. This allows the noise in the signal to only trigger an occasional stray pulse, and

thus keeps the bandwidth and power consumption even lower.














Vmem


ACI


Figure 4-1. Leaky integrate-and-fire (LIF) circuit.

Each block of the LIF will be explained though only the leaky component will be

discussed in detail as the other IF parts follow previous work as explained in detail in Dr.

C'I. i.'s publications [47, 48] and Dr. Li's dissertation [57].

4.1.1.1 Voltage to current converter circuit

The input to the LIF needs to be a current, however the output of the bio-amplifier

[25] which precedes the LIF provides a voltage. Thus, a voltage to current circuit must be

used which is shown in Figure 4-2. C1 rejects DC which is crucial to limit the feature

extractor's bandwidth. A common differential PMOS input and cascoded output

operational transconductance amplifier (OTA) is used in the voltage to current block

and it's schematic is Figure 4-2. A more detailed analysis of this circuit can be found in

Dr. Li's dissertation [57].

4.1.1.2 Comparator circuit

Figure 4-4 shows the comparator circuit which consists of three stages: the input

preamplifier, a decision block, and and output buffer. The input signal is sensed by the


negative pulse output
















C1

Vin Vou
OTA O




Vref



Figure 4-2. Voltage to current convertor circuit for LIF.

input pair M1/i\! and differential currents are copied to the next stage, the decision

block. In the decision block positive feedback is used to isolate the input pair to help

decrease the kickback noise. The cross-connected pair M7/118i increase the gain of the

comparator. The diode-connected pair M10/11 I I provides hysteresis to reject the noise on

the input signal. M13 provides a DC shift to guarantee the swing of the decision circuit

output is in the common-mode range of the output buffer. M14-M18 form an output

buffer to convert the final output of the comparator into a logic-level signal. An inverter

was added to the output of the buffer as an additional gain stage. A more detailed

analysis of the comparator can be found in Dr. C'I. i.'s work [25, 47].

4.1.1.3 Reset and refractory period circuit

The reset component resets Vmem after either a positive or negative pulse occurs

using a simple OR gate. The circuit also contains the option for a refractory period

(time delay when the reset cr li--< on) but it is minimized when the LIF is used for feature

extraction. The refractory period component is realized by an.l-i-inin.! 11;c current-starved

inverter depicted in Figure 4-5 and Dr. C'I. i.'s dissertation [47] includes a more thorough

analysis. The circuit is composed of the input C110OS pair M1/ \!0 with an additional

































M13 M8 M7 I14




M5 I 3 M4 M6


VSS


Figure 4-3. Operational transconductance amplifier OTA for voltage to current convertor
circuit for LIF.


series-connected PMOS transistor M3 in the pull-up. A PMOS transistor M4 is emploi-. I

for testing purpose. The control voltage Vbias--refractory adjusts the output rise time and

thus the refractory period. For feature extraction the Vbias--refractory is set for minimal

refractory period.

4.1.1.4 Leaky circuit

The addition of a leaky component is what distinguished the LIF from the IF. The

leaky component, a Gm current source, was realized using an OTA configured as a unity





















Vin+


Figure 4-4. Comparator circuit for LIF.


gain-follower. The OTA's schematic is depicted in Figure 4-6. The OTA's bias voltage

adjusts the amount of leakage. For a small input signal range such as neural signals have

the current is linear. A positive input value will sink current while a negative input value

sources current.

4.1.1.5 Chip specifics

Cadence Sp~ectreS simulations show the LIF circuit consumes about 30 pW of power.

The LIF circuit was fabricated using AMI 0.5 pm C \!OS technology. The chip is a

1.5 mm x 1.5 mm 40-pin DIP with 504 pm x 356 pm of circuit area as shown in Figure

4-7 The pinout for the leaky integrate-and-fire feature extractor chip used can be found in

APPENDIX B. The chip has differential input and uses +/- 2.5V power supplies.

4.1.2 Test Setup

Prior to testing the chip with any neural signals, bench top testing was done using a

function signal generator to ensure the chip was functioning properly. Three test setups

were used to progress towards in vivo testing of the chip Figures 4-8, 4-9, and 4-10. The









VDD


Vbias-refectory


Vin 0-- Vout






Vvss

Figure 4-5. Reset and refractory period circuit for LIF.

test setups were developed to limit the amount of time required to test the chip in the rat

lah as those resources are limited and shared by many people.

The initial chip test setup shown in Figure 4-8 allows the basic parallel recording

setup to be tested in NEB 487 without requiring a rat. The two things recorded in parallel

are the output of the ITF hio-aniplifier [25] and the output of the current TDT neural

recording system used in the Neuroprosthetics Research Group (NPG) lah. TDT is used

as the master clock and triggers the logic analyzer to record the LIF's hi-phasic output

at the desired time. The pulse times are later used for the pulse-based sorting algorithm

to compare the feature extractor and pulse-based sorter sorting performance to Spike2's

performance based on the bio-aniplifier's output. Using the neural simulator as the input

signal to the two systems allows the testing to be done outside of the rat lah and provides

a known ground truth.

The intermediate test setup shown in Figure 4-9 replaces the neural simulator with a

file output using TDT. The file out should be +/- 10 V to nxinintize added noise front the





Figure 4-6. Leaky circuit for LIF implemented as a Go, current source.


RX5 digital-to-analog convertor (DAC) and the PA5 that is used to attenuate the signal to

the desired neural signal level. This allows any neural data file to be used as input for the

chip. This is especially useful if simulated data with known ground truths is to be used.

The final test setup uses in vivo recordings and is shown in Figure 4-10. It is the

same as the initial setup except the input signal is now from a live rat and the complete

TDT system used in the NPG lah is recorded in parallel as well. This allows a comparison









































Figure 4-7. Layout for the LIF circuit.


between two complete systems as well as any offline spike sorters as the amplified raw

neural data is recorded.

4.2 Chip Results

The feature extractor chip was tested using the test setups mentioned previously in

Section 4.1.2 and are described in detail helow. From benchtop testing the LIF chip was

found to have more noise than the previous IF chip though both suffer from feedback front

the digital pulse output to the analog input even with the addition of off chip decoupling

capacitors between the digital power and ground input pins and separate decoupling

capacitors between the analog power and ground input pins. Another student laid out

the LIF chip but did not follow the IF chip layout with just adding the leaky component































Figure 4-8. Initial chip test setup: Compare UFs feature extractor with Dr. Sanchezs
Spike2 results using neural simulator (487 NEB).


so apparently the layout to minimize the digital to analog feedback was not as careful

as with the IF chip. This noticeably degrades the performance of the LIF when used for

reconstruction but is not as large a factor when the LIF chip is used for feature extraction

as the features are purposely robust to noise as well as the sorting algorithm.

4.2.1 Neural Simulator

The neural simulator provides absolute ground truths which make the performance

analysis of the feature extractor straightforward and a good intermediate step between

bench top testing and in vivo testing where there are no absolute ground truths. In order

to use the same exact input at different leakage and threshold setting for the LIF the

intermediate test setup from Figure 4-9 was used. However, to mimic the complete system,

neural simulator data recorded from the UF hio-amplifier [25] was used. The data was

rescaled to +/- 10 V to reduce noise as mentioned in Section 4.1.2 and then attenuated

with the PA5 back to it's original amplitude level. The process of digital-to-analog

conversion in the R X5 and attenuation in the PA5 introduce additional noise on the order


pulse outputs



























Figure 4-9. Intermediate chip test setup: Set UFs feature extractor chip parameters using
prerecorded data front rat that will be used in in vivo experiments (487 NEB).


of p-1V than if the direct output of the amplifier went to the LIF chip however the ability

to replay the exact same data for each leakage and threshold parameter combination

outweighed the additional noise present.

A leakage and threshold value were chosen to obtain the desired bandwidth front the

simulation results presented in Section :3.2.1. Then, a value above and below that was

chosen to test the LIF chip as simulation results do not ahr-l-~ 8 nap exactly to chip results

because of non-ideal factors. This resulted in a total of nine data set recordings.

The sorting error is shown in Table 4-1 and the bandwidth is shown in Table 4-2.

The spike sorting error is similar to the simulation results in Section :3.2.1 for the lower

bandwidth values but as the bandwidth increases the noise front the digital output

feedback to the analog input is increased. Thus, the largest bandwidth data has the most

effect. Also, at this point detection (being able to separate individual spikes) becomes

an issue because of the current intplenientation of the software. This combination of

problems results in an inability to spike sort at all with CI' .' error. The leakage level of

747.4 nA is not large enough to lose the noise and thus produced poor sorting results.

Even when the leakage value is increased to 847.3 nA the error is still larger than expected

for the bandwidth required. This is likely due to the large feedback noise front the digital

























Figure 4-10. Final test setup: Compare UFs analog anip and feature extractor with TDTs
anip and Dr. Sanchezs Spike2 results (rat lah). First setup with neural simulator then
when working use rat.


output to the analog input which adds many pulses to the signal. This was concluded

after observing individual spike signature varying more than expected for even the largest

of leakage current. The feedback noise is short but large in amplitude so it will perturb

the timing of the pulses and often add extra pulses which could account for the increased

variation in spike signatures compared to simulation results. Another noise source is the

quantization of pulse times by the logic state analyzer (LSA) being 5 ns but this was

accounted for in the simulations results.

The best performance was obtained with a leakage values of 946.1 nA and threshold

value of 1:30 mV which produced a bandwidth of 1.31 K( pulses/s and an error of 4.T7 .

This shows the LIF chip is capable of obtaining good sorting performance but requires

more bandwidth than expected. 1\ore careful layout to separate the sensitive analog

signals form the digital signals on the chip should allow the bandwidth to decrease for the

same performance.

A suninary table comparing the different bandwidth reduction techniques' bandwidth,

power consumption, and sorting performance are shown in Table 4-:3. The feature

extractor offers the lowest bandwidth and lowest power option while still being competitive

in spike sorting so it appears very promising.










Table 4-1: Spike sorting performance (percent error) from leaky integrate-and-fire (LIF)
feature extraction chip.
threshold (mV)
130 150 170
leakage 747.4 99.3 25.0 21.8
nA 847.3 27.3 18.7 15.2
946.1 4.7 13.2 21.3


Table 4-2: Bandwidth (pulses/s) from LIF feature extraction chip.
threshold (mV)
130 150 170
leakage 747.4 1.6i5k 1.19k 988
nA 847.3 1.42k 1.05k 885
946.1 1.31k 948 804


4.2.2 In Vivo with Rat

Ultimately the chip will be used in vivo with rats, but the performance analysis of the

chip using rat data is difficult because ground truths are not know. Presently, simulated

data sets from recorded data are being used to test the hardware using the intermediate

test setup from Figure 4-9. This allows a wide variety of realistic data to be tested but

because the ground truths are known a better analysis of the results can be done.

4.2.3 Future Work
In addition to the future work items in C'!s Ilters 2 and 3, the circuit design and chip
design add several items that need further work.

* How to implement the Leak: Either a simple variable resistor can be used to set
the leakage current or a ideally constant current source. The resistor would itself be
a form of adaption by taking more current for larger signals. The effect of the two
different leakage implementation on spike sorting performance as well as circuit area
and power need to be investigated.

* How to implement adaptive threshold in circuitry: Circuits have already been
developed [56] to implement the adaptive threshold as mentioned in OsI Ilpter 3 as an
improvement to the feature extractor so the previous work can be built upon.

* Feature extraction chip layout: the layout needs special attention to reduce the
feed through noise from the digital output to the analog input. An excellent well
know book on analog layout is written by Alan Hastings [58].










Table 4-3: Bandwidth reduction, power consumption, and sorting error comparison
ADC Biphasic with Biphasic Biphasic Spike
reconstruction feature feature detection
extraction extraction
Front-end 300 K~bps 72 K( pulses/s 455 pulses/s 680 plulses/s 2n1l'1'-
bandwidth channel bw
back-end
spike sorting 6. >6.1~ 6.' I:s N/A
error
power 3 mW 100 p-W 30 p-W 30 p-W 3 p-W


*In vivo rat testing: Once the feature extraction chip is redesigned to improve
performance it can be tested in vivo with the rats using the test plan outline in
Section 4.1.2. The test plan has been fully demonstrated up until the in vivo testing
and the TDT in vivo testing code has been written and tested.









CHAPTER 5
SINGLE-SCALE SPIK(E DETECTOR

UJF's third approach to neural data bandwidth reduction is the most dramatic and

requires an implantable spike detector. For implanted circuitry an analog implementation

is advantageous over a digital implementation because it has a much lower power

consumption and it can he more compact in size. Thus, the spike detection algorithm

chosen was limited to one that was amenable to an analog circuit implementation. Taking

inspiration from Smith's auditory onset detection scheme [59] a single-scale spike detection

algorithm hased on filtering was developed. Some of the single-scale spike detector work

presented in this chapter has been previously published [49].

5.1 Algorithm

The basic operating principal of the single-scale spike detector is to use the

thresholded difference of two low-pass filters to enhance the spike and stabilize the

baseline. One filter has a higher cutoff frequency to remove high frequency noise and the

other has a lower cutoff frequency to create a local average. When the difference between

the signal and the local average rises above a threshold, a spike is detected. This method

is robust to changes in the noise level as well as DC offsets, both of which are common

for long term neural recordings. The basic algorithm blocks are shown in Figure 5-1 along

with examples of the signal at every stage. Low pass filters are known to have a simple,

low-power, and small area implementation in analog using the subthreshold region of

operation [60, 61]. 1\ore details on the circuitry are provided later in Section 5.3.

5.2 M~atlab Simulations

To analyze the algorithm's performance prior to circuit design M~atlesb simulations

with data from in vivo neural recordings were used. First the data will be described and

then the detector's performance results will be shown.

5.2.1 Data

High SNR neural recordings, sampled at 20 K(Hz, were used to increase the confidence

of the ground truth spikes times determined from the data set. Then, white Gaussian






















Tmn~


Figure 5-1. Single-scale detector block diagram with snapshots of the waveform after each
block.


noise was added to give the detection problem a more realistic SNR level. A slowing

varying 1 Hz, 10 mV amplitude sinusoid was also added to the signal to simulate the

slowly varying DC offset. Figure 5-2 A) shows the original neural data waveform and B)

shows the 0 dB SNR waveform with an offset. SNR was calculated for the signal in terms

of power using an averaged spike shape.

5.2.2 Receiver Operating Characteristics (ROC) Curves

Receiver operating characteristic (ROC) curves are typically used to quantify the

performance of detection algorithms across the full range of thresholds [62]. ROC curves

plot the probability of a correct detection (also known as a hit) versus the probability

of a false detection (also known as a false alarm). There is ah-le w a trade-off between

the optimal detection of all the spikes and the erroneous detection of noise as a spike.

This spike detection problem also requires spike time estimation, meaning that the

performance curve could lie below the chance line for the detection problem. A detection

was considered correct if it occurred within 300 ps of the actual spike time. The ROC



























0.1

0.05




-0.05

-0.1


B 0.1


0.051


0L
1.35


1.45
Time (s)


Time (s)


Figure 5-2. Data used for simulations. A) Original waveform. B) 0 dB signal to noise
ration (SNR) waveform with offset. Column two is zoomed in from column one.


0 .12























i ~~~-*single-scalapiue

0.025 0.5 0.75 1 1.25 1.5 1.75 2
P(False Alarm) x10-3


Figure 5-:3. Plot of receiver operating characteristic (ROC) curves, OdB SNR.


curves in Figure 5-3 shows that as a larger percentage of spikes are detected more noise

will be falsely detected as a spike (also known as a false alarm). The optimal curve would

start with no detections and no false alarms and go straight to 1011' correct detection

with no false alarms. The ratio of correct detections to incorrect detections can he set to

the desired operating point on the ROC curve by choosing the corresponding threshold

level .

To determine the desired circuit cut-off frequencies for the spike detector, ROC curves

were constructed from nested cut-off frequency iterations around typical spike frequencies

100 Hz-6 K(Hz [4]. The circuit's cut-off frequencies were chosen with the minimum number

of false alarms at CII' correct detection to be 1.4 K(Hz and 5.3 K(Hz.

Once the cutoff frequencies were selected, the algorithm was tested using the in vivo

recordings described in Section 5.2.1. The single-scale detection method was compared

to the threshold method without any filtering at OdB SNR with the results shown in

Figure 5-:3. For comparison purposes the two methods were examined at their CII' correct

detection operating point. The single-scale method outperformed the amplitude threshold

without filtering method by over :30dB in terms of false alarm rate. Because spikes are










sparse in neural data the probability of a false alarm needs to be a fraction of a percent

not to swamp the number of correct detections.

The data used has an average spiking rate of 76 Hz so during one second of data

at CII' correct detections there should be 68 correct detections out of 76. At 0 dB the

single-scale method had a 2 x 10-4 probability Of a falSe alarm, 4 false detections per

second. The incorrect detection probability for the single-scale detector was I.' For the

amplitude threshold method there were 230 false detections per second, so its incorrect

detection percentage is much greater at '71.'

The single-scale detection method consistently outperformed the amplitude threshold

method over varying SNR values. Once the SNR became too low, -2 dB, neither method

performed well. Here the single-scale method degraded to 25' incorrect detections and

the amplitude threshold method was extremely poor at b 2' incorrect detect ions.

Second-order filters were simulated for the single-scale spike detection circuit but

their performance over first-order filters was negligible. Since second-order filters require

additional chip area and power without noticeable performance improvement, they were

not investigated further.

Analysis of the M~atlesb simulation results showed that at 911' correct detection almost

all of the false alarms came from noise riding on the second peak of the action potential.

Spike-like noise over other parts of the signal was only detected "' of the time a false

alarm occurred. The 1 in 4.r portion of false alarms could be reduced by blinding the

detector for a short period after it detects a spike. The trade-off to this would be the

detector losing resolution between spikes. The amount of time the detector is blinded

equals the minimum time required between spikes for detection.

Since the amplitude threshold is a subset of the single-scale method's circuitry, it

will consume less power. However, its lack of robustness to low SNR and slowly varying

DC offsets hinders its performance for BlMI devices. M~atlesb simulations with real neural

recordings showed that with 911' correct detections at 5dB SNR the single-scale method










outperformed the amplitude thresholdingf method with 1 incorrect detections versus the

71 incorrect detections respectively. This performance continued for lower SNR such

as 0 db with the single-scale detector having only a I.' incorrect detection rate while

that of the amplitude threshold method was '71.' Though the results of computationally

intensive methods such as matched filteringf and template matching were not examined,

they are sure to provide better results given enough information is known about the signal.

However, even if enough information about the signal was known, the power consumption

of matched filtering and template matching and their required supervision to adjust

parameters as spike shapes and noise change over time prohibits implantation.

5.3 Circuit Design

The three 1!! li.r~ factors in this circuit design relate to the need for implantation

and are low-power, small area, and robustness for the desired computation. Figure 5-4

shows the overall circuit block diagram for the single-scale spike detector. Each of the

blocks will be briefly discussed. An operational transconductance amplifier (OTA) is

configured as a follower integrator for the first-order low-pass filters. The OTAs are run in

the subthreshold region to reduce power [60, 61] and to allow the capacitors to be small

enough to fit on chip. The bias voltages are set off chip to enable adjustment of the cutoff

frequencies after fabrication.

The desired cut-off frequencies for the two filters were found to be 1.4 K(Hz and

5.3 K(Hz from M~atlab simulations described in Section 5.2.2. With the OTAs' bias voltages

set for a transconductance, gm, of 150 nA/V and desired cutoff frequencies of 1.4 K(Hz

and 5.3 K(Hz for the low-pass filters in Figure 5-4 the corresponding capacitance values are

C1 = 22.5 pF and C2 = 4.9 pF from Equation. 5-1.



C = m(5-1)
2x fe













output


input I I
L Bias
thresh
O TA Vthresh(

Vr m
Vbias2

Figure 5-4. Single-scale spike detector circuit diagram.

Neural spikes can vary in width from 0.4 ms to 3 ms [5, 6] depending on the species and

brain area so the cut-off frequencies are set to remove all of the noise outside the spike

frequency ranges for the particular application.
After the signal has been filtered, the difference of the two filtered signals is taken

using an OTA. The output is then thresholded with current, which is set with Vthress. For

complete unsupervised operation an automatic method for setting the threshold would be
needed. This thresholded signal is sent through two inverters to ensure a binary output
decision.

Cadence Sp~ectreS simulations show the circuit consumes an average of 1 pW of

power. The single-scale spike detector chip was fabricated using AMI 0.5 pm C'jIOS
technology. The chip is a 1.5 mm x 1.5 mm 40-pin DIP with 253 pm x 223 pm of circuit

area. The layout is shown in Figure 5-5.

5.4 Chip Results

Due to the difficulty of testing the chip with real neural data, the chip was first tested

with two basic signal generators to crudely approximate neural data. A square wave was

used to mimic the spike and a high frequency sine wave was used to simulate noise on the







Wi@WiW WiWWitW
8 HIM 11


Figure 5-5. Layout for onset spike detector chip.










signal. The input signal was based on three characteristics of real neurons: spike width,

time between spikes, and amplitude.

Because the circuit detects the onset of the spike, the effective spike width is the

width of the first rise in the action potential. With infinite SNR, this would mean

approximately half the action potential time, but as SNR degrades it reduces. The

chips functionality was tested with pulse widths of 100-400 ps.

The second signal characteristic is the time between spikes. Individual neurons have a

refractory period, which sets a minimum time between spikes. But, electrodes often record

from more four to six neurons so the refractory period is not a determining factor in the

minimum time between spikes. It is then optimal to detect a spike as close the previous

one as possible since superimposed spikes can not be discriminated between as is the case

with most spike detectors including this one. The filter del ws~ and the time to charge

the load capacitance are the two factors which determine the minimum detectable time

between spikes for the circuit.

Amplitude is the third characteristic of the input signal. Extracellular neural signals

have peak-to-peak amplitudes of 50 pV-500pV [3]. This small signal must first be

amplified to give a larger voltage swing for the analog spike detection circuit to be more

accurate. Tod w-, low noise, low power neural amplifiers, such as the UF bioamplifier

[25], can achieve a gain of up to 100, so the input signal amplitude ranges between 5 and

50 mV.

The result of a 35 mV square wave with a 125 ps pulse width at 25' duty cycle

combined with a 15 mV high frequency sine wave (to mimic neural noise) is shown in

Figure 5-6 as the bottom waveform. It shows that 10 ps after the input spikes the output

goes high for a short period.

The chip was tested over a wide range of input signal characteristics loosely patterned

after neural data. The threshold voltage allows the chip to be adjusted to change the false

alarm penalty, and correspondingly its probability of correct detection, in accordance with















































Figure 5-6. Singfe-scale spike detector chip results with signal generator pulse waveform
as input. Ch2, top waveform, is the input signal from a signal generator and Ch3, bottom
waveform, is the output signal.


Ch31 I.00 V |