Fear Memories in Visual Cortex

Material Information

Fear Memories in Visual Cortex Inter-Individual Differences Related to the Comt Val158met Polymorphism and Reflex Physiology
Gruss, Laura Forest
Place of Publication:
[Gainesville, Fla.]
University of Florida
Publication Date:
Physical Description:
1 online resource (14 p.)

Thesis/Dissertation Information

Master's ( M.S.)
Degree Grantor:
University of Florida
Degree Disciplines:
Committee Chair:
Committee Co-Chair:
Committee Members:
Graduation Date:


Subjects / Keywords:
Alleles ( jstor )
Anxiety ( jstor )
Catechols ( jstor )
Cognitive psychology ( jstor )
Electroencephalography ( jstor )
Fear ( jstor )
Genotypes ( jstor )
Heart rate ( jstor )
Psychology ( jstor )
Visual cortex ( jstor )
Psychology -- Dissertations, Academic -- UF
conditioning -- polymorphism
bibliography ( marcgt )
theses ( marcgt )
government publication (state, provincial, terriorial, dependent) ( marcgt )
born-digital ( sobekcm )
Electronic Thesis or Dissertation
Psychology thesis, M.S.


Classical fear conditioning is a widely used laboratory model to study aversive learning. Deficiencies in the acquisition and extinction of a learned fear response are expressed in a variety of anxiety and mood disorders. In studying underlying neurobiological mechanisms of these disorders, potential contributing factors of inter-individual differences can be better understood. In the current study we investigated genetic polymorphisms that may differentially impact the fear circuitry in an instructed fear conditioning paradigm. Specifically, we were interested in how variability in the fear circuitry may result in changes of perceptual processing in the primary visual cortex recorded through electroencephalography (EEG). Peripheral physiological measures were additionally taken to assess strength of fear engagement. The COMT (catechol-O-methyltransferase) val158met polymorphism revealed genotype differences, in that individuals homozygous on the Val allele (high enzymatic activity allele) showed significant cortical enhancement to the aversively cued stimulus (CS+) during initial conditioning. This was followed by subsequent heart rate acceleration in response to CS+ presentation. Other measures of fear engagement (skin conductance and startle response) showed no genotype differences. Two conclusions can be drawn from the results of this polymorphism: Variability in the activation of visual cortex implicates a modulatory role of the fronto-striatal dopaminergic system in aversive learning. In addition, the autonomic nervous system appears to be impacted by this polymorphism through peripheral catecholamine regulation. Future research is needed to replicate and extend these findings to a more comprehensive understanding of not only fear acquisition, but fear extinction as well. ( en )
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis (M.S.)--University of Florida, 2014.
Statement of Responsibility:
by Laura Forest Gruss.

Record Information

Source Institution:
Rights Management:
Applicable rights reserved.
Resource Identifier:
968786222 ( OCLC )
LD1780 2014 ( lcc )


This item has the following downloads:

Full Text


NeuralNetworksLearnHighlySelectiveRepresentationsinOrderto OvercometheSuperpositionCatastropheJeffreyS.Bowers,IvanI.Vankov,MarkusF.Damian,andColinJ.DavisUniversityofBristolAkeyinsightfrom50yearsofneurophysiologyisthatsomeneuronsincortexrespondtoinformation inahighlyselectivemanner.Whyisthis?Wearguethatselectiverepresentationssupportthecoactivationofmultiple“things”(e.g.,words,objects,faces)inshort-termmemory,whereasnonselective codesareoftenunsuitableforthispurpose.Thatis,thecoactivationofnonselectivecodesoftenresults inablendpatternthatisambiguous;theso-calledsuperpositioncatastrophe.Weshowthatarecurrent paralleldistributedprocessingnetworktrainedtocodeformultiplewordsatthesametimeoverthesame setofunitslearnslocalistletterandwordcodes,andthenumberoflocalistcodesscaleswiththelevel ofthesuperposition.Giventhatmanycorticalsystemsarerequiredtocoactivatemultiplethingsin short-termmemory,wesuggestthatthesuperpositionconstraintplaysaroleinexplainingtheexistence ofselectivecodesincortex. Keywords: grandmothercells,localistrepresentations,distributedrepresentations,short-termmemory, superpositioncatastrophe Supplementalmaterials: informationinahighlyselectivemanner.Thisselectivityisbest documentedinorganismswithsimplenervoussystems(e.g., Elliott&Susswein,2002 ),butitisobservedincomplexorganisms aswell,includingcellsintheinferiortemporalcortexandhippocampus.Forexample,aneuroninthehippocampusofahuman wasfoundthatstronglyrespondedtodifferentphotographsofthe actressJenniferAnistonbutnottoimagesofotherpersons,places, oranimals( QuianQuiroga,Reddy,Kreiman,Koch,&Fried, 2005 ).Whethertheseresultsareconsistentwithlocalistor“grandmothercell”codingisamatterofdispute( Bowers,2010 ; Plaut& McClelland,2010 ; QuianQuiroga&Kreiman,2010 ),butthereis nodoubtthatsomeneuronsrespondtoinformationinahighly selectivemanner(foradetailedreviewoftheneuroscience,see Bowers,2009 ). Giventhesefindings,animportantquestioniswhydosome neuronsrespondinthisway?Afamiliarexplanationwasfirst advancedby Marr(1971) ,whoarguedthatmemoriesinthehippocampusarestoredinahighlysparseformatsothatdifferent memoriesarecodedwithlargelynonoverlappingneurons.This allowsanetworktolearnquicklywithoutnewmemoriesinterferingwitholdmemories,asneededforepisodicmemoryforexample.Thatis,sparsenonoverlappingmemoriesprovideasolutionto catastrophicinterference ( McCloskey&Cohen,1989 ),orthe stability-plasticitydilemma ( Grossberg,1980 ).However,thisexplanationdoesnotexplainthemanyreportsofselectiverespondingofneuronsincortex.Forinstance, Logothetis,Pauls,and Poggio(1995) trainedtworhesusmonkeystoidentifyalargeset ofnovelcomputer-generatedobjects.Aftertraining,Logothetiset al.recordedfrom796neuronsininferiortemporalcortex.Afew (3/796;0.37%)respondedselectivelytooneobjectpresentedfrom anyviewpoint,and,morefrequently,neurons(93/796;11.6%) respondedselectivelytoasubsetofviewsofoneobjectsbutrarely (ornotatall)tohighlysimilarobjects.Thislevelofselectivityis asgreatasobservedinthehippocampus. Thesehighlyselectiveresponsesobservedinthecortexrequire anexplanationaswell.Weproposethatthetaskofactivating multiplethingsatthesametimeoverthesamesetofunitsin short-termmemory(STM)constitutesapressuretolearnhighly selectivecodesinthecortex,giventhatvariousperceptualand cognitivesystemswithinthecortexsupportSTM( Cowan,2001 ). Onthisview,superimposeddistributedpatternsoftenresultin ambiguousblendsthatcannotbeinterpreted,and,inthissituation, theonlysolutionistolearnhighlyselective(orlocalist)codes,as discussednext.TheSuperpositionCatastropheAnumberofauthorshavearguedthatthesuperpositioncatastrophelimitstheabilityofmostnetworkstocoactivatemultiple thingsatthesametimeoverthesamesetofunits( Bowers,2002 ; Page,2000 ; Rosenblatt,1961 ; VonderMalsburg,1986 ).Accordingtothishypothesis,apatternofactivationacrossasetofunits inanetworkcanprovideanunambiguousrepresentationofa ThisarticlewaspublishedOnlineFirstFebruary24,2014. JeffreyS.Bowers,IvanI.Vankov,MarkusF.Damian,andColinJ. Davis,SchoolofExperimentalPsychology,UniversityofBristol,Bristol, UnitedKingdom. ThisresearchwassupportedbyLeverhulmeGrantRJ5538,awardedto JeffreyS.Bowers. CorrespondenceconcerningthisarticleshouldbeaddressedtoJeffreyS. Bowers,UniversityofBristol,SchoolofExperimentalPsychology,12a PrioryRoad,BristolBS8-1TU,UnitedKingdom.E-mail: j.bowers@ Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.PsychologicalReview ©2014AmericanPsychologicalAssociation 2014,Vol.121,No.2,248–261 0033-295X/14/$12.00DOI: 10.1037/a0035943248


singleitem,butsuperimposingtwoormorepatternsoverthesame unitscanresultinablendpatternthatisambiguousinthatthereis nowaytoreconstructthepatternsfromtheblend. Thesuperpositioncatastrophehastypicallybeenconsideredin thedomainofvision,andthequestionhasfocusedonhowtobind togetherfeaturesofoneobjectwhenmultipleobjectsareinthe sceneatthesametime(inordertoavoidambiguousblendsof features).Forexample, Rosenblatt(1961) illustratedtheproblem inasimpleneuralnetworkcomposedoffourlocalistunits,with unit1respondingtoatriangleinanyposition,unit2respondingto asquareinanyposition,unit3respondingtoarbitraryobjectsin theuppervisualfield,andunit4respondingtoarbitraryobjectsin thelowervisualfield.AsnotedbyRosenblatt,thisnetworkcan unambiguouslydeterminetheshapeandlocationofasingleobject throughthecoactivationoftwounits(e.g.,atriangleintheupper visualfieldiscodedthroughthecoactivationofunits1and3),but itfailstorepresenttheidentityandlocationoftwoobjectswhen presentedatthesametime(e.g.,atriangleintheuppervisualfield andasquareinthelowervisualfieldwillcoactivateallfourunits, andsowillasquareintheuppervisualfieldandatriangleinthe lowervisualfield). Thesuperpositioncatastropheisalsoapotentialprobleminthe domainofSTM,inwhichsinglethings(objects,words,etc.)are encodedoneatatimebutmultiplethingshavetobemaintainedin memoryovertime.Forinstance,theRosenblattnetworkwould alsohavedifficultyinrememberingthefollowingsequenceof events:squaredisplayedinthelowervisualfieldfollowedbya triangledisplayedintheuppervisualfield.Theencodingofthe firstobjectwouldbeunambiguous,andtheitemcouldpersistin STMthroughthecontinuedactivationofunits2and4.Butwhen theseconditemisencoded,thesameambiguousblendisproduced. Intheaboveexamplesthesuperpositioncatastropheoccursina networkwithlocalistrepresentationsofshapesandlocations.The problemwouldappeartobemoreacutewhenknowledgeiscoded inadistributedformat.Forexample,consider Figure1 ,adapted from Page(2000) ,whichdepictsdistributedrepresentationsforthe namesJohn,Paul,George,Ringo,Mick,Keith,Brian,Charlie, Roger,andPete.Eachnameiscodedasapatternofactivation across20units,withfourunitsonandtheremainingunitsoff.The identityofthenamecanbedeterminedbyexaminingthefull patternofactivitybutnotbyexaminingindividualunits(e.g., activityinthefirstunitispotentiallyconsistentwithPete,Charlie, orRingo).Thepatternsinthisexamplewererandomlygenerated, andthishasresultedinsomepairsofpatternsthatareentirely dissimilar(e.g.,PeteandRogerarenotcodedbyanycommon units)andotherpairsthatarequitesimilar(e.g.,RogerandBrian arecodedbythreecommonunits).Thecriticalpointforour purposeisthatblendsofthesepatternscanbehighlyambiguous. ConsiderthepatternsforRoger,Brian,andPaul.Thetwounits thatdisambiguateRogerandBrian(i.e.,theunitthatisonfor BrianandoffforRoger,andviceversa)arebothoninthepattern thatcodesPaul.Consequently,thepatternsthatcodethecombinationofRogerandPaulorBrianandPaulareidentical,ascanbe seeninthebottomtworowsofthefigure.Thatis,theresulting blendisambiguous:Itisnotpossibletodeterminewhetheritwas producedbycombiningRogerandPaulorBrianandPaul. Although Figure1 showsthataproblemwithsuperpositioncan occurwhencombiningdistributedpatterns,itshouldbenotedthat noneoftheotherpossiblecombinationsoftwonamesforthis randomlygeneratedsetresultsinthissortofambiguity.Onemight thereforewondertowhatextentthesuperpositioncatastropheisa genuineproblem.Toillustratethesuperpositioncatastrophemore formally,weundertookanexhaustiveanalysisoftheparticular caseinwhichasetof20unitsisusedtocode20words.Asforthe aboveexample,weassumedthatthepatternsarebinary(i.e.,each unitiseitheronoroff)andthatcombiningtwopatternsinwhich agivenunitisoninoneorbothofthepatternsresultsinablend patterninwhichthatunitison.Onewaytothinkaboutthismethod ofcombiningpatternsistoimaginedrawingthecodeforeach patternonaseparateoverheadtransparency;thesuperpositionofa setofwordscorrespondstotheimageyouseewhenstackingthe correspondingoverheadsontopofeachother.Moreformally,this methodofsuperpositionisequivalenttoalogicalORoperation. Ourapproachwastoconstructvocabulariesof20wordsthrough randomsamplingandthentodeterminethesuperpositioncorrespondingtoeverypossiblelistofwords(uptoamaximumlist lengthofsixwords). Figure2 showstheproportionofblend patternsthatareambiguousasafunctionoflistlengthandthe sparsenessofcoding;thatis,thenumberofunitsthatareinvolved incodingeachword.Thesparsertheinput,thelesstheword patternsoverlap;theextremeconditionofusingonlyasingleunit touniquelycodeeachwordcorrespondstoalocalistrepresentation.Eachpointisbasedonanexhaustivesampleoflistsofagiven length.Forexample,thereare38,760waysinwhichsixitemscan besampledfromavocabularyof20words.Whenfourunitsare usedtocodeeachword,37,042(96%)oftheselistsresultin ambiguousblends(thisfigureisthemeanof100samples,each usingadifferentrandomlydefinedsetofpatternsforthe20words inthevocabulary). Figure1. Distributedpatternsacross20unitsforthenamesPete,Roger, Charlie,Brian,Keith,Mick,Ringo,George,Paul,andJohn,aswellastwo superpositions(Paul Roger)and(Paul Brian)thatareambiguous giventhattheyresultinthesameblendpatternacrossthe20units.The unitsthatareonoroffarerepresentedbydarkandlightcircles,respectively.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.249SUPERPOSITIONCATASTROPHE


Ascanbeseeninthefigure,thedegreeofambiguityinsuperpositionsincreasesascodingbecomeslesssparseandasthe numberofpatternsbeingcombined(thelistlength)increases. Wheneachwordiscodedbyfourunits,combinationsoftwo wordsareveryunlikelytogiverisetoambiguousblends;inthis respect,theexampleshownin Figure1 issomewhatunrepresentative.However,whenfivewordsarecombined,ambiguous blendsaretheruleratherthantheexception.Theproblemisworse stillwheneachwordiscodedbymanyunits(densedistributed coding),butitisapparentevenwhenonlytwounitsareusedto codeeachword(sparsedistributedcoding).Theonlycaseinwhich ambiguousblendsareavoidedisthelocalistcodingcase,inwhich eachwordiscodedbyasingledistinctunit. Theresultsoftheanalysisdepictedin Figure2 indicatethatthe superpositioncatastropheisagenuineproblemfordistributed representations,buttheseresultsdonotimplythatspecificmodels thatincludedistributedrepresentationsmustsufferfromthisproblem.Itisconceivablethattheproblemwillnotoccurinpracticein modelsthathavecontinuous(ratherthanbinary)unitsandthat combinepatternsusingoperationsotherthanthelogicalORoperator.Butitisbynomeansobviousthatthisisthecase(e.g., averagingcontinuousunitsmightleadtoanevenmoreextreme superpositioncatastrophe).Ourworkinghypothesiswasthatthe superpositioncatastrophewouldincreaseinseverityasafunction ofthedensityofthedistributedcodingandthenumberofpatterns thatmustbesimultaneouslyactivated.Thesimulationswepresent belowenabledustotestthishypothesis.1ResponsestotheSuperposition CatastropheHypothesisTherehavebeenthreegeneralresponsestothesuperposition catastrophehypothesis.Thefirst(mostcommon)responseisto ignoretheissue.Thisismadepossiblebythefactthatmostneural networksaredesignedtocodeonethingatatimeunderconditions inwhichthesuperpositioncatastropheconstraintdoesnotarise. Notonlydoesthisresponseleavetheproblemunresolved,butit mayunderminemodelsthatsucceedwithsingleitems.Thatis, evenifamodelsucceedswithsingleitems,itmaysolvethe probleminaqualitativelydifferentwaythanhumans,whonot sharethisrestriction.Forinstance,aneuralnetworkmodelofword namingthatincludesdistributedrepresentationsmightaccountfor arangeofdataonsingle-wordnaming,butifthereadingsystem supportsthecoactivationofmultipleorthographicandphonologicalforms(forthesakeofmorecomplexlanguagetasksorSTM), themodelÂ’ssolutionmaymischaracterizethebrainÂ’ssolutioneven intherestricteddomainofsingle-wordnaming. Thesecondresponsehasbeentoacceptthatthisisan importantconstraintonperceptionandcognitionandtodevelop methodstoeliminatetheprobleminneuralnetworks.The solutionstypicallyrelyonsomesortoflocalistcodinginorder tobindtogethertheappropriatefeatures.Consideragainthe Rosenblatt(1961) example.Onesolutionwouldbetoaddan additionallayeroflocalistrepresentations,so-calledconjunctivecodes,thatmaptogethershapesandlocationsinlong-term memory.Forexample,asnotedbyRosenblatt,thecoactivation ofthelocalistunits square,triangle,upper-visual-field, and lower-visual-field isambiguous(shapesarenotboundtolocation),butthecoactivationofconjunctivelocalistunitsfor triangle-in-the-upper-visual-field and square-in-the-lowervisual-field isnot.Arelatedsolutionistointroducedynamic binding,inwhichlocalistunitsareboundtogetherinshort-term memoryviasynchronousfiring(e.g., Hummel&Biederman, 1992 ).Tostickwiththe Rosenblatt(1961) example,onecan codethetriangleandthesquareunambiguouslyifthesquare unitandalowerfieldunitarecoactiveinsynchrony,andthe triangleandupper-visual-fieldunitarecoactiveinsynchrony (andoutofphasewithoneanother).Thissolvesthesuperpositioncatastrophewithoutaddinganotherlayerofconjunctive localistcodes(andavoidsacombinatorialexplosionoflocalist units).Note,thesetwoapproachesshouldbeconsideredcomplementarysolutionsratherthanalternatives,withconjunctive codesusedforbindingtogetherfeaturesoffamiliarthings(e.g., bindingtogetherfourlinesegmentsintoalocalistrepresentationofsquare),andsynchronybindinglocalistunitsthattogethercomposelessfrequentthings(bindingtogethertherepresentationsofsquareandlowervisualfield).Forpresent purposes,theimportantpointtoemphasizeisthatbothapproachestobindingaremotivatedbytheviewthatblended distributedpatternsareambiguous,andthesolutioninboth casesistoavoiddistributedblendpatterns:inthefirstcaseby employinglocalconjunctivecodes,andinthesecondcaseby activatingonethingatatime(throughsynchrony).Localist codesareagoodmediumforimplementingtemporalsynchrony ( Hummel,2000 ). Thethirdresponseistoclaimthatthesuperpositioncatastrophehasbeenoverstated.Indeed,advocatesofdistributed codescanpointtoexistingconnectionistmodelsthatcancoactivatemultiplethingsatthesameoverthesamesetofunits 1Itisimportanttodistinguishbetweenthecapacityofanetworktostore multipleitemsinlong-termmemoryusingdistributedrepresentations (whichisnotindoubt)andthecapacityofanetworktocoactivatemultiple distributedrepresentationsatthesametime.Thesuperpositioncatastrophe hypothesisconcernsthelaterissue. Figure2. Thepercentageofambiguousblendpatternsasafunctionof listlengthandnumberofactiveunitswhen20unitsareusedtocode20 words.Listlengthvariesfromonetosixwords,andnumberofactiveunits variesfromoneunitperwordto19unitsperword.Ambiguitiesincrease asafunctionofbothfactors,apartfromthecaseofwordscodedwitha singleunit(localistcoding).ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.250BOWERS,VANKOV,DAMIAN,ANDDAVIS


(e.g., Botvinick&Plaut,2006 ; McClelland,St.John,&Taraban,1989 ; Touretzky&Hinton,1988 ). BotvinickandPlaut (2006) developedaparalleldistributedprocessing(PDP)model ofimmediateserialrecallthatcoactivatedmanyletterswithout producingmeaninglessblendpatterns(upto8intheirsimulations,withbetterperformancepossible; Bowers,Damian,& Davis,2009 ).Themodelwaspresentedwithalistoflettersone atatimeandtrainedtoreproducethelettersinthesame order—aclassictestofshort-termmemory.Strikingly,the BotvinickandPlaut(2006) modelnotonlysucceededbutalso capturedarangeofempiricalfactsaboutSTM.Ontheirview, distributedrepresentationsdoindeedlimitamodel’scapacityto codeformultiplethingsatthesametime,buttheselimitations helpexplainhumanperformance. ThesuccessesofexistingmodelsclearlyshowthatPDP networkscancodeformultiplethingsatthesametimeoverthe samesetofunits.However,itisimportanttonotethatthese successesdonotunderminethepotentialsignificanceofthe superpositioncatastrophe.Thecriticalquestionis how doPDP modelslearntocodemultiplethingsatthesametime.Itisat leastpossiblethatPDPmodelslearnhighlyselectiveoreven localistcodesinordertoovercomethesuperpositioncatastrophe.ThismightbeparticularlytruewhenrecurrentPDPnetworksaretrainedtoencodemanyitemsatthesametimetaken fromalargevocabularyofitems(e.g., Botvinick&Plaut,2009 ; Bowersetal.,2009 ),suchthatthesuperpositionofdistributed patternsismaximallyambiguous. Belowwesystematicallyexplorethequestionofwhethera neuralnetworklearnshighlyselectiveorlocalistcodesin responsetothesuperpositioncatastrophe.Tothisendwe trainedasimplerecurrentPDPmodeltostoremultiplewordsat thesametimeandthencarriedout“single-unit”recordingson thehiddenunits—analogoustothesingle-cellrecordingstudies carriedoutinneuroscience—toseehowthemodelsucceeded. Forourpurposes,PDPmodelsareusefulbecausetheytypically learndistributedcodes,and,accordingly,theyprovideastrong testofourhypothesis.Furthermore,itisoftenclaimedthata keyadvantageofPDPmodelsisthattheylearnrepresentations thatarebestsuitedforagiventask( Plaut&McClelland,2000 ). Thatis,thelearnedrepresentationsareemergentratherthan “stipulated”bythemodeler.So,ifaPDPmodellearnslocalist codeswhencodingformultiplethingsatthesametime,this stronglysuggeststhatthesuperpositionconstraintprovidesa pressuretolearnselectivecoding.Suchapressurecouldnotbe demonstratedwithamorebiologicallyplausiblenetwork(e.g., Grossberg,1980 )thatwasaprioridesignedtolearnlocalor highlysparsecodes. Toexploretheimpactofthesuperpositioncatastropheperse, wemovedawayfromstudyingspecificcognitivecapacities, suchasimmediateserialrecall,thatinvolvemorethanencoding multipleitemsatthesametime.Rather,wetrainedarecurrent PDPmodeltoencodeaseriesofwordsoneatatimeandthen recallthemallatthesametime.Althoughthistaskdoesnot correspondtoanybehavioraltask,itconstitutesarelatively puresuperpositioncondition(i.e.,themodelissimplyrequired tocodeformultipleitemsatthesametimeinitshiddenlayer). Accordingly,anylocalcodesthatdevelopinthistaskwilllikely reflecttheimpactofthesuperpositionconstraintasopposedto anyadditionalcomputationalrequirementsassociatedwith morecomplextasks(suchascodingtheorderoftheto-beremembereditems,asrequiredintheserialrecalltask).Critically,wemanipulatedtheextenttowhichwordpatternswere superimposedinordertoexplorewhetherPDPmodelslearn moreselectivecodeswhentheambiguityassociatedwiththe blendpatternsismostacute.Simulation1WetrainedarecurrentPDPmodeltoencodeandrecallwords takenfromavocabularyof30wordsandmanipulatedthe expectedseverityofthesuperpositioncatastropheintwoways. First,wevariedthenumberofwordsthatthemodelhadtostore inSTM,withlistlengthvaryingfrom1to3,5,and7words. Thelongerthelist,thegreaterweexpectedthesuperposition catastrophetobe.Second,wevariedthesparsenessofcodingin theinputlayer.Thatis,weusedlocalistwordunits(eachword wasassociatedwithasingleinputunit),localistletterunits (eachletterwasassociatedwithasingleinputunit),anddistributedletterunits(eachletterwasassociatedwith3input units,andeachunitwasinvolvedincoding3differentletters). Wereasonedthattherewouldbeanincreaseinthesuperpositionacrossthesethreeinputcodingschemes,giventhatthereis acorrespondingincreaseintheoverlapintheinputpatterns acrossthesethreeconditions(asin Figure2 ).Ifthesuperpositionconstraintprovidesapressuretolearnlocalistcodes,then morelocalcodesshouldbelearnedinthelongerlistconditions andmorelocalcodingshouldbelearnedwhentheinputpatternsoverlappedmoresubstantially. Thenetworkwascomposedof30inputletterunits,200 hiddenunits,and30outputwordunits.Inputunitswerefully interconnectedwithhiddenunits,hiddenunitswerefullyconnectedwithoutputunits,andthehiddenlayerwasfullyrecurrent. Therewasalsoabiasunit,connectedtoalltheunitsinthe hiddenandoutputlayers.Whenwordswerecodedlocallyintheinput layer,therewasone unitdevotedtoeachword.Whenthewords werecodedascollectionsofletters,weorganizedtheinputlayer into10unitscodingforconsonantsintheonsetpositionofaword, 10unitscodingforvowels,and10unitscodingforconsonantsin thecodaposition,witheachwordcodedasoneonset,onevowel, andcodaunit.The30inputunitscodedforthefollowinglettersin theonset,vowel,andcodapositions,respectively:( b,c,d,f,g,h, j,k,l,m )( a,e,i,o,u,y,aa,ea,ou,oo )( n,p,q,r,s,t,v,w,x,z ). Giventhateachwordwasdefinedasthecoactivationofoneonset, onenucleus,andonecodeunit,thetotalnumberofpossiblewords was1,000(10onsets 10nucleus 10codas).Whenwordswere composedoflocalistlettercodes,eachwordwascomposedof threeactiveinputunits(e.g.,“ban”wascodedbycoactivatingthe inputunits1,11,21),andwhenwordswerecomposedofdistributedlettercodes,eachwordwascomposedofnineactiveinput units,witheachlettercodedbythreeunits(e.g.,“ban”wascoded bycoactivatingtheinputunits1,2,3,11,12,13,21,22,23).The listofwordsandtheirinputcodingschemesareshownin Table1 . Theinputlayerincludeda31stunitthatwasactivatedwhenthelist ofwordswastoberecalled.Theoutputunitscodedforwordsin alocalistmanner,withoneunitperword. Thetrainingprocedureconsistedoftwophases,encodingand retrieval.Intheencodingphase,thenetworkwaspresentedwith aseriesofwords,oneatatime,thatithadtostoreintheThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.251SUPERPOSITIONCATASTROPHE


recurrentconnectionsinitshiddenlayer.Thesecondphase consistedofasinglestepinwhichthenetworkhadtooutputall ofthewordsatthesametime.Asaconcreteexample,consider thetaskofencodingandrecallingthetwowords“ban”and “bep”inthenetworkwithlocalistlettercodingunits.Thewords “ban”(inputunits1,11,21)and“bep”(inputunits1,12,22) wouldbepresentedinsequence,eachstoredintherecurrent hiddenlayer.Then,followingtheretrievalcue(inputunit31), themodeloutputs“ban”(unit1)and“bep”(unit2)atthesame timeintheoutputlayer.Thesametwooutputunitswouldbe simultaneouslyoutputwhengiventhesequence“bep-ban” (giventhatorderdoesnotmatter). Thenetworkwastrainedwithbackpropagationthroughtime, avariantofthebackpropagationalgorithmthatissuitablefor trainingrecurrentPDPnetworks.Thelearningratewasfixedto 0.01,andamomentumof0.9wasused.Thestandardsigmoid activationfunctionwasusedinboththehiddenandtheoutput layer.Thegainoftheactivationfunctionwassetto1.Theerror attheoutputlayerwascomputedwiththecrossentropyfunction. Simulation1includedatotalof12conditions:4 listlength (1,3,5,and7words)and3 input-codingscheme(localword, localletter,anddistributedletterinputcoding).Whenthe modelwastestedinamultiplewordcondition,allthetrained listscontainedthecorrespondingnumberofwords(i.e.,when themodelwastestedonlistsofthreewords,alltrainedlists containedthreewords).Giventhatthedifficultyofthetask variedacrossconditions,wereporttheperformanceofthe modelatintervalsof100,000trainingtrials.Inaddition,given theslightvariabilityoftheresultsacrosssimulations,weran eachcondition100times.Theaverageperformanceofthe modelisreportedin Figure3 . Asisclearfrom Figure3 ,themodelneededmoretraining whentrainedtorecalllongerlistsofwordsandwhentrained withthedistributedinputcodingschemes.Still,performance wasexcellentinallcasesbyonemilliontrials.Thecritical question,however,ishowdidthemodelsucceedacrossthe variousconditions?Toaddressthisissue,wetookarandom versionofthemodel(outof100runs)aftertrainingitforone milliontrialsandrecordedtheactivationofthehiddenunitsin responsetowords,analogoustothesingle-cellrecordingstudiescarriedoutinneuroscience.Followingtraininginallofthe aboveconditionswerecordedtheactivationofall200hidden unitsinresponsetoall30words(presentedoneatatime)and displayedtheresultswithagraphicalmethodintroducedby Berkeley,Dawson,Medler,Schopflocher,andHornsby(1995) . Inthismethod,aseparatescatterplotforeachhiddenunitis created,andeachpointinascatterplotcorrespondstoaunit’s activationinresponsetoasingleinput(e.g.,aword).Levelof unitactivationiscodedalongthe x -axis,anddistinctvaluesare assignedtoeachpointalongthe y -axisinordertoprevent pointsfromoverlapping(wordswereorganizedrandomly). Thismethodprovidesasingle-unitrecordingforeachhidden unitinresponsetoallwords. Table1 The30WordsUsedinSimulation1andTheirCorrespondingWord,LocalistLetter,andDistributedLetterInputCodingSchemes WordLocalistwordinputcodingLocalistletterinputcodingDistributedletterinputcoding ban100000000000000000000000000000100000000010000000001000000000111000000011100000001110000000 beq000000000010000000000000000000100000000001000000000010000000111000000001110000000011100000 bis000000000000000000001000000000100000000000100000000000100000111000000000111000000000111000 cep010000000000000000000000000000010000000001000000000100000000011100000001110000000111000000 cir000000000001000000000000000000010000000000100000000001000000011100000000111000000001110000 cot000000000000000000000100000000010000000000010000000000010000011100000000011100000000011100 diq001000000000000000000000000000001000000000100000000010000000001110000000111000000011100000 dos000000000000100000000000000000001000000000010000000000100000001110000000011100000000111000 duv000000000000000000000010000000001000000000001000000000001000001110000000001110000000001110 for000100000000000000000000000000000100000000010000000001000000000111000000011100000001110000 fut000000000000010000000000000000000100000000001000000000010000000111000000001110000000011100 fyw000000000000000000000001000000000100000000000100000000000100000111000000000111000000000111 gaax000000000000000000000000100000000010000000000010000000000010000011100000000011101000000011 gus000010000000000000000000000000000010000000001000000000100000000011100000001110000000111000 gyv000000000000001000000000000000000010000000000100000000001000000011100000000111000000001110 haaw000000000000000100000000000000000001000000000010000000000100000001110000000011100000000111 heaz000000000000000000000000010000000001000000000001000000000001000001110000000001111100000001 hyt000001000000000000000000000000000001000000000100000000010000000001110000000111000000011100 jaav000000100000000000000000000000000000100000000010000000001000000000111000000011100000001110 jeax000000000000000010000000000000000000100000000001000000000010000000111000000001111000000011 joun000000000000000000000000001000000000100000000000101000000000000000111010000000111110000000 keaw000000010000000000000000000000000000010000000001000000000100000000011100000001110000000111 koop000000000000000000000000000100000000010000000000010100000000000000011111000000010111000000 kouz000000000000000001000000000000000000010000000000100000000001000000011110000000111100000001 laq000000000000000000000000000010000000001010000000000010000000100000001111100000000011100000 loon000000000000000000100000000000000000001000000000011000000000100000001111000000011110000000 loux000000001000000000000000000000000000001000000000100000000010100000001110000000111000000011 map000000000000000000010000000000000000000110000000000100000000110000000111100000000111000000 mer000000000000000000000000000001000000000101000000000001000000110000000101110000000001110000 mooz000000000100000000000000000000000000000100000000010000000001110000000111000000011100000001 Note .Eachwordiscodedacross30inputunits,with1and0indicatingthatthecorrespondinginputunitwasonoroff,respectively.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.252BOWERS,VANKOV,DAMIAN,ANDDAVIS


In Figure4 weplottheactivationofeachhiddenunitwhen themodelwithdistributedinputlettercodeswastrainedfor1 milliontrialsundertwoconditions;namely(a)whentrainedon wordsoneatatime,and(b)whentrainedonlistsofseven words.Asisclearfromthefigure,thepatternofactivation acrosstheunitsisverydifferentacrossconditions,withselectiverespondingevidentonlywhenthemodelwastrainedon listsofsevenwords.Thisselectivityishighlightedinthree specificcasesin Figure4c .Forexample,Unit113wasofftoall wordsapartfrom“gus.” Inordertosummarizetheresultsmoresuccinctlysothatwe candisplaytheseresultsacrossall12conditions,wedeveloped aselectivitymetricthatmeasuredtheextenttowhichagiven hiddenunitrespondedtoagivenwordselectively.Theselectivityofaunitwascomputedastheminimaldifferencein activationbetweenonewordandalltherest.Theseselectivity valuescanvaryfrom1(whenagivenworddrivesagiven hiddenunittoanactivationof1andallotherwordsleadtono activation;i.e.,1 0 1)to 1(whenagivenwordfailsto activateagivenhiddenunitandallotherwordsdrivetheunit toanactivationof 1;i.e.,0 1 1).Inthelattercase,a unitisselectivelycodingtheinputpatternbybeingoff.The idiosyncrasiesofthemodelingarchitecture,theinputset,orthe taskmayhaveallcontributedtotheemergenceoftheseOFF selectiveunits,butforpresentpurposes,theimportantpointis thatthecodesalsohighlightthecomputationaladvantageof selectivecodingwhenconfrontingthesuperpositioncatastrophe.InallofthesubsequentanalyseswetreattheseOFFunits asselective,butthesamepatternofresultsandconclusions followsifonlypositiveselectivityvaluesareconsidered. In Figure5 wedepicttheselectivityvaluesofthe200hidden unitswhenthenetworkwithadistributedinputcodingscheme wastrainedonlistsof(a)one,(b)three,(c)five,and(d)seven words.Inthelasttrainingconditionweidentifiedthewordto whichagivenhiddenunitresponded,withtheexactselectivity valuesshowninparentheses.Ascanbeseeninthefigure,the modeldidnotlearnanylocalistcodeswhentrainedtorecall singlewords,anditlearnedlocalistcodesfor21ofthe30 wordswhentrainedonlistsofsevenwordsata0.5selectivity criterion.Notethatthe0.5criterionisarelativelyconservative measureofselectivity.Itexcludesseveralunitsthatcould reasonablybecalledselective,suchasunit2,whichresponds selectivelyto“diq,”buthasanabsoluteselectivityscorelower than.5(see Figure4c ). In Figure6 wesummarizethenumberofselectivewordunits acrossall12conditionsforwhichtheabsolutevalueofthe selectivitymeasureexceeds0.5(averagingacross100simulations).Asisclearfromthefigure,thenumberoflearnedlocalist wordcodesincreasedasafunctionofthelistlengthandthe natureoftheinputcodingscheme,withlocalistword,localist letter,anddistributedletterinputcodingschemesproducing increasinglymorelocalistwordcodesinthehiddenlayer. Critically,localistcodesveryrarelyemergedwhenthemodel wastrainedtorecallsingleitems;thatis,whenthemodeldid notconfrontthesuperpositioncatastrophe.Thesefindingsare justaspredictedontheviewthatlocalistcodingemergesin responsetothesuperpositionconstraint,withthenumberof localcodesscalingwiththeseverityofthesuperpositioncatastrophe.Simulation2Akeyfeatureoftheabovesimulationsisthatthemodel includedmanymorehiddenunits(200)thanthesizeofthe trainedvocabulary(30words),and,accordingly,therewere morethanenoughresourcesforthemodeltodevoteahidden unittoeachword.Thisraisesaquestion.Whatwouldhappenif themodelwastrainedtorecallmultiplewordstakenfroma muchlargervocabulary,sothatthemodelincludedfewer hiddenunitsthantrainedwords?Thiswouldprecludetheuseof localistwordcodingschemestosolvethesuperpositioncatastrophe,and,accordingly,iflocalistwordcodeswererequiredto performthetask,themodelshouldfail.Alternatively,perhaps themodelcanadoptadistributedorsomeothersolutionunder Figure3. (a)Performanceofthenetworkwithlocalistwordcodingatthe inputlayerasafunctionofthenumberoftrainingtrialsandlistlength.(b) Performanceofthenetworkwithlocalistlettercodingattheinputlayeras afunctionofthenumberoftrainingtrialsandlistlength.(c)Performance ofthenetworkwithdistributedlettercodingattheinputlayerasafunction ofthenumberoftrainingtrialsandlistlength.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.253SUPERPOSITIONCATASTROPHE


theseconditionsandstillsucceed.InSimulation2weexplored thisissuebytrainingthemodeltorecallwordstakenfroma muchlargervocabulary. WetookthemodelfromSimulation1withdistributedinput lettercodesand200hiddenunitsandtraineditonavocabulary of300words.Accordingly,theoutputlayersizewasincreased from30to300(oneforeachlocalistwordunit).Againwe trainedthemodelonlistsof1,3,5,and7words.Giventhe greaterchallengeposedbythelargervocabulary,wetrainedthe modelforaslongas20milliontrials,inordertoprovide maximumopportunityforthemodeltosucceed.Again,given theslightvariabilityoftheresultsacrosssimulations,weran Figure4. (a)Scatterplotsofthe200hiddenunitstakenfromthenetworkwithdistributedlettercodingatthe inputlayerwhentrainedonavocabularyof30wordsoneatatime.Withineachscatterplot,eachcrossrepresents theunit’sresponsetoaparticularword.(b)Correspondingplotsofthe200hiddenunitswhennetworkwas trainedonavocabularyof30wordspresentedinlistsofsevenwords.(c)LabeledscatterplotofUnit113,Unit 116,andUnit2takenfrom Figure4b .Unit113respondstotheword“gus”withaselectivityof0.9,whileUnit 166respondstotheword“gaz”withaselectivityof 0.89.AlthoughUnit2respondstotheword“diq”more thantheotherwords,itsselectivityof0.49fallsbelowourthresholdof0.5.Asaconsequence,itisnot consideredaselectiveunitintheanalysesthatfollow.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.254BOWERS,VANKOV,DAMIAN,ANDDAVIS


eachcondition5times,andwereporttheaverageperformance ofthemodelin Figure7 .Ascanbeseenin Figure7a ,themodel didinfactstruggle.Indeed,whentrainedtorecalllistsofseven wordsthemodelreachednobetterthan20%accuracy.Still,it isinterestingtonotethatbytheendoftrainingthemodelhad greaterthan95%accuracyonlistsofthreewordsandapproxFigure5. Selectivityplotforthenetworkwithdistributedlettercodingattheinputlayertrainedon(a)singlewords, (b)listsofthreewords,(c)listsoffivewords,and(d)listsofsevenwords,asafunctionofvocabularysize.Each hiddenunitiscodedbyasquare(10perrow),anddegreeofselectivityisindicatedbythedegreeoflightnessofthe square,withlightgrayreferringtoaunitwithhighselectivityandblackreferringtoaunitthatisnonselective.In(d), theunitsthattakeonselectivityvalueabove.5arelabeledwiththelettertowhichtheyrespond,andtheprecise selectivityvalueispresentedinparentheses.Whentrainedonwordsoneatatime,allunitsarenonselective;whentrained onlists,someunitsareselective,withmoreselectiveunitsassociatedwithlongerlists.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.255SUPERPOSITIONCATASTROPHE


imately80%accuracyonlistsof5words.Wereportin Figure 7b thecrossentropyerrorthatthemodelwastrainedtominimize.Hereitisevidentthatperformanceimprovedquiteearly inthetrainingbutthattheerrordoesnotconvergetozerowhen themodelwastrainedonlistsofsevenwords.2Inordertogaininsightintohowthemodelsucceeded(tothe extentthatitdid)acrosstheconditions,weagaincarriedout single-unitrecordingsfromthehiddenlayertakenfromarandom versionofthemodel(outof5runs)aftertrainingitfor20million trials.In Figure8 weplottheactivationofeachhiddenunitwhen themodelwithdistributedlettercodeswastrainedtorecall(a)one wordand(b)sevenwords.Onceagain,nolocalcodeswere acquiredaftertrainingthemodelonwordsoneatatime.In addition,ascanbeseenclearlyin Figure8b ,therearenounitsthat codeforspecificwordsaftertrainingonsevenwords.However, therewasasetofunitsforwhichthereweretwodiscretebandsof activation,suchthatonesubsetofwordsdrovetheunittotakeon onelevelofactivationandanothersubsetofwordsdrovetheunit toanotherlevelofactivation.Furthermore,itisstraightforwardto interpretmanyofthebandsgiventhatallthewordsinagivenband oftencontainedaspecificletter.Forinstance,considerUnit60,in Figure8c ,whichshowedtwodistinctbandsofactivation.Allthe wordscontainedwithinthehighlyactivatedbandcontainedthe letter n, whereasallthewordswithintheinactivebanddidnot containtheletter n. Thatis,thisunitappearstobealocalist detectorfortheletter n. Again,inordertosummarizeouranalysesofthehiddenunits, wedevelopedaselectivitymeasure.Inthiscase,theselectivityof ahiddenunitwascomputedastheminimaldifferenceinactivation betweenwordsthatcontainedagivenletterandwordsthatdidnot containthisletter.In Figure9 wedepicttherangeofselectivity valuesacrossthe200hiddenunitswhenthemodelwastrainedon listsof(a)one,(b)three,(c)five,and(d)sevenwords.Inthelast caseweidentifywhatletteraunitselectivelyrespondedtowhen itsabsoluteselectivityvaluewasabove.5,withexactselectivity valuesshowninparentheses.In Figure10 wesummarizethe numberofselectiveletterunitsacrossconditionsaveragingacross the5runsofthesimulation.Ascanbeseeninthefigures,the modellearnednoselectivecodeswhentrainedtorecallsingle words,learnedafewlettercodeswhentrainedonlistsofthree words,andlearnedmanylocalistlettercodeswhentrainedtorecall listsoffiveandsevenwords.Indeed,whenadoptingthe0.5 selectivitycriterionthemodellearnedalmostthefullsetofpossiblelettercodeswhentrainedinthelaterconditions.Thisundoubtedlyunderestimatesthenumberofselectivecodes,giventhat someunits,suchasUnit68in Figure8c ,wereselectivebelowthis criterion.So,onceagain,thenumberoflearnedlocalistcodes scaledwiththelevelofambiguity. Whydidn’tthelocalistlettercodessupportbetterperformance whentrainedonthelongerlists?Theanswerisstraightforward: Thecoactivatedlettercodescouldnotuniquelyspecifywhatsetof wordswerepresentedtothemodel.Toillustrate,consideracase inwhichthevocabularyoftheabovemodelincludedthewords “abc,”“def,”“ghi,”and“adg,”andthemodelispresentedthelist “abc-def-ghi”torecall.Inthissituationallthelettersforthe nonpresentedword“adg”arecoactivated,andthisambiguitywill leadtoerrors.Thatis,themodelissufferingfromthesuperpositioncatastrophedespitelearninglocalistlettercodes.Thelonger thelistofwordstoremember,themoreambiguoustheblend.This isanalogoustothelocalistversionofthesuperpositioncatastrophe notedbyRosenblattanddiscussedearlier.Thesolution,asnoted byRosenblatt,istolearnlocalistcodesforcompleteitems,inthis casewords,butthemodeldidnothavethenecessaryresources. Still,itisclearthatlearninglocalistlettercodeswasbetterthan learningnolocalcodes,andthemodeldidthebestthatitcould withthelimitedresourcesatitsdisposal.GeneralDiscussionAstrikingresultfromneuroscienceisthatsomeneuronsrespondhighlyselectivelytoinformation,bothinthehippocampus andinthecortex( Bowers,2009 ).Thereisanongoingdebateasto whetherthisselectivityisconsistentwithlocalist(grandmother cell)coding(cf. Bowers,2010 , 2011 ; Plaut&McClelland,2010 ; QuianQuiroga&Kreiman,2010 ),butwhateverthecase,itis importanttodeterminewhysomeneuronsrespondinthisway. Ourmaincontributionistohighlightthepotentialrelevanceof thesuperpositioncatastrophe.Thecurrentsimulationsprovide clearevidencethatrecurrentnetworkstrainedtostoremultiple things(inthiscasewords)atthesametimeoverthesamesetof unitsoftenlearnhighlyselective(indeedlocalist)representations. Ofinterest,theconstraintsposedbythesuperpositioncatastrophe inthedomainofSTMcomplementstheconstraintsposedby catastrophicinterferenceinthedomainoflong-termmemory 2Inordertodeterminewhetherthelimitedsuccessofthemodelonlists ofsevenwordswasdependentonthespecificlearningratethatwe employedabove,wevariedthelearningrateparameterfrom.01to.02,.03, and.04andtrainedthemodeltoupto20milliontrials.Themodelwas limitedinitsabilitytorecallsevenwordsatalllearningrates(maximum performanceinallcasesdidnotexceed20%),suggestingthatthemodel’s performancewasrestrictedbyitslimitedresourcesratherthanthespecific learningparametersweemployed.Inaddition,inallcases,themodel learnedlocalistcodes.Thishighlightsthefactthattheemergenceof localistcodeswasnotcontingentonaspecificlearningrate.Detailsofour simulationsthatvariedlearningratescanbefoundat .com/site/superpositioncatastrophe/ Figure6. Thenumberofselectivewordcodesinthemodelswithlocalist word,localistletter,anddistributedletterinputcodingschemeswhen trainedonavocabularyof30wordsasafunctionofamountoftrainingand listlength.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.256BOWERS,VANKOV,DAMIAN,ANDDAVIS


(LTM).Thatis,justasdistributedrepresentationsareapoor mediumforSTM,distributedrepresentationsareapoormedium forrapidlearninginLTM(e.g., McCloskey&Cohen,1989 ). Together,theseconstraintsonSTMandLTMmayhelpexplain whyneuralcodingisselectiveinboththecortex(forthesakeof STM)andthehippocampus(forthesakeofepisodicLTM). Itisimportanttonotethatwefoundthatthetypeoflocalist codingvariedacrosssimulations,withlocalistwordcodeslearned inSimulation1andlocalistlettercodeslearnedinSimulation2. Thisvariancereflectedthedifferentcomputationalresourcesavailabletothemodelsacrossthetwosimulations.Whenthemodelhad morethanenoughhiddenunitstoencodeallthetrainedwordsin alocalistmanner,themodeldevelopedlocalistwordcodes(see Figure4c ).Thiswasaneffectivestrategy,asitallowedthemodel tosucceed100%ofthetimerecalling7wordsfollowingamillion trainingtrials.However,whentherewerenotenoughhiddenunits toencodewordsinalocalistmanner,themodelcontinuedtolearn localistcodesbutattheletterlevel.Thissolutioncouldnotsupport goodperformanceonthelongerlistsduetotheambiguitiesthat arisewithmultiplecoactivatedletters(thesuperpositioncatastrophe).Nevertheless,theresultshighlightthepressuretolearn localistcodesinresponsetothesuperpositionconstraintand indicatethatlocalistlettersprovidedabettersolutionthanpurely distributedcoding. Inonerespect,thelocalistlettercodingresultsinSimulation2 arethemostimpressive.Thatis,thelocalistlettercodesconstitutedanemergentrepresentation,giventhattheinputlayerincludeddistributedletters(eachletterwascodedasapatternof activationoverthreeunitsandeachunitwasinvolvedincoding threeletters)andtheoutputlayerincludedlocalistwordcodes(one unitperword).Itissometimesclaimedthatlocalistrepresentations are“stipulated”bythemodeler(e.g., Plaut&McClelland,2000 ), butinthiscase,thelocalistlettercodesemergedwithoutcorrespondinginputoroutputrepresentations.Thisagainhighlightsthe computationaladvantageoflocalistcodingwhenconfrontingthe taskofcodingmultiplethingsatthesametime. WhatshouldbemadeofthefactthatthemodelsinSimulation 1and2successfullyrecalledlistsofthreewordsrelyingon relativelyfewlocalistcodes(withmorelocalistcodesemerging onlywithlongerlists)?Thisfindinghighlightsthefactthatdistributedcodeshavesomelimitedcapacitytoovercomethesuperpositioncapacity(ascanalsobeseenin Figure2 ).Similarconclusionshavebeenmadebefore.Forexample, BotvinickandPlaut (2006) arguedthattheirrecurrentPDPmodelofimmediateserial recallsucceededonthebasisofdistributedcodesbylearninga biastorecallthemostlikelysequencesgivenitstraininghistory. Thisbiaswasthoughttoreducetheambiguitytosuchanextent thatthedistributedrepresentationscouldsupportSTMatalevel commensuratewithhumanperformance.However,ourfindings showthatdistributedsolutionsonlyworkunderlimitedconditions. Whenweincreasedtheseverityofthesuperpositionproblem,such thattheblendsofmultipleitemswerehighlyambiguous,our modelsreliedmuchmoreheavilyonlocalistcodes. Inaddition,whatshouldbemadeofthefactthatthemodels neverlearnedlocalistwordcodesforallthetrainedwordsin Simulation1?Eveninthemostdifficulttrainingconditionsthat producedthemostlocalistcodes,weonlyobservedapproximately 20outof30wordcodes(atleastbyourstrictstandardof.5 selectivity).Doesthiscompromiseourclaim?Notatall.Weare claimingonlythatthesuperpositioncatastropheprovidesapressuretolearnselectivecodesinPDPnetworks,andthispressure mighthelpexplainthemanysingle-cellrecordingstudiesthathave observedhighlyselectivecodingincortex.Furthermore,weare notclaimingthatthesuperpositioncatastropheistheonlyconstraintthatmightcontributetothedevelopmentofselectivecoding incortex.Forexample, OlshausenandField(2004) ; Page(2000) ; and Thorpe(2011) haveallidentifiedkeycomputationaladvantagesofhighlyselectiveorlocalistcoding,anditismetabolically expensivetohaveahighproportionofneuronsfiringatonce ( Lennie,2003 ).Ourmaincontributionistoprovideevidencethat thesuperpositionconstraintisyetanothercausalfactorthatmight helpexplaintherepeatedobservationthatsomeneuronsrespondto informationinaremarkablyselectivemanner. WewanttoemphasizeagainthatalthoughourPDPnetwork learnedlocalistcodes,wearenotcommittedtotheviewthatthe brainreliesonlocalist(grandmothercell)coding.Rather,wetake Figure7. (a)Performanceofthemodelwithdistributedlettercoding schemeandtrainedonavocabularyof300wordsasafunctionoflist lengthandnumberoftrainingtrials.Evenafter20millionsoftrials,the networkperformspoorlyonlonglists.(b)Networkerrorasafunctionof listlengthandamountoftraining.Althoughthenetworkperformedatfloor followingfivemilliontrainingtrialswhentrainedonlistsofsevenwords (see Figure7a ),thenetworkerrorreducedbyapproximatelyhalf.Theerror doesnotconvergetozerointhiscondition.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.257SUPERPOSITIONCATASTROPHE


Figure8. (a)Scatterplotsofthe200hiddenunitswhenthenetworkwithdistributedinputlettercodingscheme wastrainedonavocabularyof300wordsoneatatime.Withineachscatterplot,eachcrossrepresentstheunitÂ’s responsetoaparticularword.(b)Correspondingscatterplotswhenthenetworkwastrainedonlistsofseven words.(c)LabeledscatterplotofUnit60andUnit68takenfrom Figure8b .Unit60respondstowordsthat containtheletter n withaselectivityof0.80,whileUnit68respondstowordsthatcontaintheletter a witha selectivityof0.27.Thislatterselectivityscorefallsbelowourstringentthresholdof0.5,and,asaconsequence, theunitisnotconsideredselectiveintheanalysesthatfollow.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.258BOWERS,VANKOV,DAMIAN,ANDDAVIS


ourfindingsasevidencethatthesuperpositioncatastropheprovidesapressuretolearnhighlyselectivecodes.Accordingly,our conclusionsarenotinconsistentwiththeclaimthatthebrainrelies onpopulationsofhighlyselective(butnotgrandmother)neurons tocompute(e.g., Pouget,Dayan,&Zemel,2000 ).Still,giventhe combinationofcomputationalandbiologicalconstraints,this extremeversionofselectivityshouldnotbedismissedoutof hand. Figure9. Selectivityplotofthe200hiddenunitswhenthenetworkwithdistributedinputlettercodingscheme wastrainedonavocabularyof300words(a)oneatatime,(b)inlistsofthreewords,(c)inlistsoffivewords, and(d)inlistsofsevenwords.Eachhiddenunitiscodedbyasquare(10perrow),anddegreeofselectivityis indicatedbythedegreeoflightnessofthesquare,withlightgrayreferringtoaunitwithhighselectivityand blackreferringtoaunitthatisnonselective.In (d) theunitsthattakeonselectivityvalueabove.5arelabeled withthelettertowhichtheyrespond,andthepreciseselectivityvalueispresentedinparentheses.Whentrained onwordsoneatatime,allunitsarenonselective;whentrainedonlists,someunitsarecategorizedasselective toletters,withmoreselectiveletterunitsassociatedwithlongerlists.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.259SUPERPOSITIONCATASTROPHE


Asafinalnote,itisimportanttoemphasizethatthecurrent findingsdonotconstituteachallengeforPDPmodelsofperceptionandcognition.Rather,ourfindingshighlightaproblemwith howresearchersoftenthinkaboutPDPmodels.Onthestandard view,PDPmodelslearndistributedcodes,andthesedistributed codesaresimilartothoselearnedinthebrain.Whatwehavefound isthataPDPmodellearnedlocalistcodeswhentrainedtosupport STM,3and,asdetailedelsewhere( Bowers,2009 , 2010 , 2011 ), localistcodesareatleastconsistentwithwhatisfoundinthebrain. ThereareotherreasonstothinkthatPDPmodelsareinadequateto thetaskofexplainingcognitionandperceptionandthatalternative (“symbolic”)networkapproachesareneeded(e.g., Bowersetal., 2009 ; Hummel&Holyoak,2003 ).Nevertheless,thecurrentsimulationsshowhowusefulPDPmodelscanbeinidentifying importantconstraintsthatallneuralnetworksneedtoaddressand oneplausiblesolutiontothesuperpositioncatastrophe:namely,the developmentoflocalist,orhighlyselective,codes. 3WehaveobtainedthesameresultsinPDPmodelstrainedondifferent tasks.Forinstance,wehavefoundthatPDPmodelsofimmediateserial recallsimilartothemodelsof BotvinickandPlaut(2006) learnlocalist codes.Thisfurthersupportsourconclusionthatlocalistcodesprovidea solutiontothesuperpositioncatastrophethatwillariseinanysituationin whichmultipleitemsarecoactiveatthesametimeoverthesamesetof units.ReferencesBerkeley,I.S.N.,Dawson,M.R.W.,Medler,D.A.,Schopflocher,D.P., &Hornsby,L.(1995).Densityplotsofhiddenunitactivationsreveal interpretablebands. ConnectionScience,7, 167–186. doi:10.1080/ 09540099550039336 Botvinick,M.M.,&Plaut,D.C.(2006).Short-termmemoryforserial order:Arecurrentneuralnetworkmodel. PsychologicalReview,113, 201–233. doi:10.1037/0033-295X.113.2.201 Botvinick,M.M.,&Plaut,D.C.(2009).Empiricalandcomputational supportforcontext-dependentrepresentationsofserialorder:Replyto Bowers,Damian,andDavis(2009). PsychologicalReview,116, 998– 1002. doi:10.1037/a0017113 Bowers,J.S.(2002).Challengingthewidespreadassumptionthatconnectionismanddistributedrepresentationsgohand-in-hand. CognitivePsychology,45, 413–445. doi:10.1016/S0010-0285(02)00506-6 Bowers,J.S.(2009).Onthebiologicalplausibilityofgrandmothercells: Implicationsforneuralnetworktheoriesinpsychologyandneuroscience. PsychologicalReview,116, 220–251. doi:10.1037/a0014462 Bowers,J.S.(2010).Moreongrandmothercellsandthebiological implausibilityofPDPmodelsofcognition:AreplytoPlautandMcClelland(2010)andQuianQuirogaandKreiman(2010). Psychological Review,117, 300–306. doi:10.1037/a0018047 Bowers,J.S.(2011).Whatisagrandmothercell?Andhowwouldyou knowifyoufoundone? ConnectionScience,23, 91–95. doi:10.1080/ 09540091.2011.568608 Bowers,J.S.,Damian,M.F.,&Davis,C.J.(2009).Afundamental limitationoftheconjunctivecodeslearnedinPDPmodelsofcognition: CommentonBotvinickandPlaut(2006). PsychologicalReview,116, 986–997. doi:10.1037/a0017097 Cowan,N.(2001).Themagicalnumber4inshort-termmemory:A reconsiderationofmentalstoragecapacity. BehavioralandBrainSciences,24, 87–114. doi:10.1017/S0140525X01003922 Elliott,C.J.H.,&Susswein,A.J.(2002).Comparativeneuroethologyof feedingcontrolinmolluscs. JournalofExperimentalBiology,205, 877–896. Grossberg,S.(1980).Howdoesabrainbuildacognitivecode? PsychologicalReview,87, 1–51. doi:10.1037/0033-295X.87.1.1 Hummel,J.E.(2000).Localismasafirststeptowardsymbolicrepresentation. BehavioralandBrainSciences,23, 480–481. doi:10.1017/ S0140525X0036335X Hummel,J.E.,&Biederman,I.(1992).Dynamicbindinginaneural networkforshaperecognition. PsychologicalReview,99, 480–517. doi:10.1037/0033-295X.99.3.480 Hummel,J.E.,&Holyoak,K.J.(2003).Asymbolic-connectionisttheory ofrelationalinferenceandgeneralization. PsychologicalReview,110, 220–264. doi:10.1037/0033-295X.110.2.220 Lennie,P.(2003).Thecostofcorticalcomputation. CurrentBiology,13, 493–497. doi:10.1016/S0960-9822(03)00135-0 Logothetis,N.K.,Pauls,J.,&Poggio,T.(1995).Shaperepresentationin theinferiortemporalcortexofmonkeys. CurrentBiology,5, 552–563. doi:10.1016/S0960-9822(95)00108-4 Marr,D.(1971).Simplememory:Atheoryforarchicortex. Philosophical TransactionsoftheRoyalSocietyofLondon,SeriesB:Biological Sciences,262, 23–81. McClelland,J.L.,St.John,M.,&Taraban,R.(1989).Sentencecomprehension:Aparalleldistributedprocessingapproach. Languageand CognitiveProcesses,4, 287–335. doi:10.1080/01690968908406371 McCloskey,M.,&Cohen,N.J.(1989).Catastrophicinterferencein connectionistnetworks:Thesequentiallearningproblem.InG.H. Bower(Ed.), Thepsychologyoflearningandmotivation (pp.109–165). NewYork,NY:AcademicPress. Olshausen,B.A.,&Field,D.J.(2004).Sparsecodingofsensoryinputs. CurrentOpinioninNeurobiology,14, 481–487. doi:10.1016/j.conb .2004.07.007 Page,M.P.A.(2000).Connectionistmodelinginpsychology:Alocalist manifesto. BehavioralandBrainSciences,23, 443–467. doi:10.1017/ S0140525X00003356 Plaut,D.C.,&McClelland,J.L.(2000).Stipulatingversusdiscovering representations. BehavioralandBrainSciences,23, 489–491. doi: 10.1017/S0140525X00473358 Plaut,D.C.,&McClelland,J.L.(2010).Locatingobjectknowledgeinthe brain:CommentonBowers’s(2009)attempttorevivethegrandmother cellhypothesis. PsychologicalReview,117, 284–288. doi:10.1037/ a0017101 Figure10. Thenumberofselectivelocallettercodesinthenetworkwith distributedinputlettercodingschemewhentrainedonavocabularyof300 wordsasafunctionofamountoftrainingandlistlength.ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.260BOWERS,VANKOV,DAMIAN,ANDDAVIS


Pouget,A.,Dayan,P.,&Zemel,R.(2000).Informationprocessingwith populationcodes. NatureReviewsNeuroscience,1, 125–132. doi: 10.1038/35039062 QuianQuiroga,R.,&Kreiman,G.(2010).Measuringsparsenessinthe brain:CommentonBowers(2009). PsychologicalReview,117, 291– 297. doi:10.1037/a0016917 QuianQuiroga,R.,Reddy,L.,Kreiman,G.,Koch,C.,&Fried,I.(2005, June23).Invariantvisualrepresentationbysingleneuronsinthehuman brain. Nature,435, 1102–1107. doi:10.1038/nature03687 Rosenblatt,F.(1961). Principlesofneurodynamics:Perceptronsandthe theoryofbrainmechanisms .Washington,DC:SpartanBooks. Thorpe,S.J.(2011).Grandmothercells,neocorticaldarkmatterandvery longtermvisualmemories. JournalofVision,11 (11),Article1243. doi:10.1167/11.11.1243 Touretzky,D.S.,&Hinton,G.E.(1988).Adistributedconnectionist productionsystem. CognitiveScience,12, 423–466. doi:10.1207/ s15516709cog1203_4 VonderMalsburg,C.(1986).AmIthinkingassemblies?InG.Palm&A. Aertsen(Eds.), Braintheory (pp.161–176).Berlin,Germany:Springer.ReceivedJune28,2013 RevisionreceivedSeptember17,2013 AcceptedNovember4,2013 ThisdocumentiscopyrightedbytheAmericanPsychologicalAssociationoroneofitsalliedpublishers. Thisarticleisintendedsolelyforthepersonaluseoftheindividualuserandisnottobedisseminatedbroadly.261SUPERPOSITIONCATASTROPHE




© 2014 Laura Forest Gruss


To Joel Rosenfeld and my parents Gayle Gruss and Richard Gruss


4 ACKNOWLEDGMENTS I thank my mentor Dr. Andreas Keil for all his advice and support throughout, along with all members of the Center for the Study of Emotion and Attention who have helped me through this project. I also thank Joel Rosenfeld and my family for their never end ing emotional support and constant belief in me .


5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 7 LIST OF FIGURES ................................ ................................ ................................ .......... 8 LIST OF ABBREVIATIONS ................................ ................................ ............................. 9 ABSTRACT ................................ ................................ ................................ ................... 10 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 12 Inter Individual Differences during Aversive Learning ................................ ............. 12 Fear Circuitry in the Brain ................................ ................................ ....................... 13 Genetic Predisposition: the COMT Polymorphism ................................ .................. 15 Inter Individual Differences during Defensive Engagement ................................ .... 17 Research Goal ................................ ................................ ................................ ........ 18 2 METHODS ................................ ................................ ................................ .............. 20 Participants ................................ ................................ ................................ ............. 20 Stimuli ................................ ................................ ................................ ..................... 20 Design and Procedure ................................ ................................ ............................ 21 Data Collection and Processing ................................ ................................ .............. 22 Electroencephalography (EEG) ................................ ................................ ........ 22 Collection ................................ ................................ ................................ ... 22 Preprocessing ................................ ................................ ............................ 23 Processing ................................ ................................ ................................ . 23 Peripheral Physiological Measures ................................ ................................ .. 24 Self Report Measures ................................ ................................ ...................... 25 Genotyping ................................ ................................ ................................ ....... 25 Statistical Analyses ................................ ................................ ................................ . 26 3 RESULTS ................................ ................................ ................................ ............... 27 Electroencephalography ................................ ................................ ......................... 27 Peripheral Physiology ................................ ................................ ............................. 27 Self Report ................................ ................................ ................................ .............. 29 4 DISCUSSION ................................ ................................ ................................ ......... 31 Summary ................................ ................................ ................................ ................ 31


6 The Role of Fronto Striatal Modulation during Aversive Learning ........................... 31 COMT Polymorphism Impacts the ANS ................................ ................................ .. 33 Limitations and Future Directions ................................ ................................ ........... 34 Concluding Remarks ................................ ................................ ............................... 36 LIST OF REFERENCES ................................ ................................ ............................... 44 BIOGRAPHICAL SKETCH ................................ ................................ ............................ 50


7 LIST OF TABLES Table page 3 1 Self report measures by COMT genotype grouping.. ................................ ......... 43


8 LIST OF FIGURES Figure page 3 1 Distribution of conditioning effects of EEG scalp potentials . ............................... 37 3 2 Activation over occipita l pole to CS+/ stimuli measured through ssVEP during first half of conditioning . . ................................ ................................ .......... 38 3 3 Heart rate reactivity over conditioning phase . ................................ ..................... 39 3 4 Heart rate reactivity in accelerative window during the second block of conditioning. ................................ ................................ ................................ ....... 40 3 5 Skin conductance responding during conditioning.. ................................ ............ 41 3 6 Blink magnitude of the startle response during conditioning. .............................. 42


9 LIST OF ABBREVIATIONS ANS Autonomic nervous system COMT Catechol O methyltransferase CS Conditioned stimulus EEG Electroencephalography SNP Single nucleotide polymorphism s sVEP Steady state visually evoked potential US Unconditioned stimulus


10 Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science FEAR MEMORIES IN VISUAL CORTEX: INTER INDIVIDUAL DIFFERENCES RELATED TO THE COMT VAL158MET POLYMORPHISM AND REFLEX PHYSIOLOGY By Laura Forest Gruss August 2014 Chair: Andreas Keil Major: Psychology Classical fear conditioning is a widely used laboratory model to study aversive learning. Deficiencies in the acquisition and extinction of a learned fear response are expressed in a variet y of anxiety and mood disorders. In studyin g underlying neurobiological mechanisms of these disorders, potential contributing factors of inter individual differences can be better understood. In the current study we investigated genetic polymorphisms that may differentially impact the fear circ uitry in an instructed fear conditionin g paradigm. Specifically, we were interested in how variability in the fear circuit ry may result in changes of perceptual processing in the primary visual cortex recorded through electroencephalography ( EEG ) . Peripher al physiological measures were additionally taken to assess strength of fear engagement. The COMT (catechol O methyltransferase) val158met polymorphism revealed genotype differences, in that individuals homozygous on the Val allele (high enzymatic activity allele) showed significant cortical enhancement to the aversively cued stimulus (CS+) during initial conditioning. This was followed by subsequent heart rate acceleration in response to CS+ presentation . Other measures of fear engagement (skin conductance and startle


11 response) showed no genotype differences. Two co nclusions can be drawn from the results of this polymorphism : Variability in the activation of visual cortex implicates a modulatory role of the fronto striatal dopaminergic system in aversive le arning . In addition, the autonomic nervous system appears to be impacted by this polymorphism through peripheral catecholamine regulation. Future research is needed to replicate and extend these findings to a more comprehensive understanding of not only fear acquisition, but fear extinction as well.


12 CHAPTER 1 INTRODUCTION Inter Individual Differences during Aversive L earning The prevalence and dive rsity of anxiety and mood disorders in the U.S. has made it imperative to understand their underlying neurobiological mechanisms (Kessler et al., 2012) . Negative affect, the subjective experience of distress, is a major focus of experimental research in which acute defensive mobilization during fear and anxiety can be studied in conjunction with the associated fear circuitry in the brain. De fensive reactivity has been shown to be highly variable across situations, diagnoses, individuals as well as response levels (electrocortical, physiological and behavioral self report measures) (McTeague & Lang, 2012; Lang et al., 2014) . Beyo nd a clinical sample, inter individual differences exist in healthy individuals and can be studied using salient, motivationally relevant stimuli (Wangelin et al., 2 011) . Genetic polymorphisms within a population may shed light on how genetic dispositions contribute to inter individual differences. Single nucleotide polymorphisms (SNPs ), nucleotide mutations resulting in the substitution at a single base pair, have become a popular field of study in the advent of The International HapMap Project, establishing over 10 million SNPs in 2006. Of these, 300 600 thousand are common polymorphi sms occurring in the population that code for the majority of genetic variation. In the current study, polymorphisms implicated in differential regulation of the fear circuitry as well as brain plasticity were selected. Aversive learning is essential for survival and is thought to be mediated through the engagement of subcortical, limbic structures as well as sensory and motor cortices (Buchel et al., 1998) . Importantly, stimuli tha t have no inherent saliency can acquire aversiveness thr ough paired conditioning (CS+) and elicit a conditi oned response.


13 Classical, Pavlovian fear conditioning in the laboratory has demonstrated that learning can occur as far downstream as primary visual cortex (Stolarova et al., 20 06) , modulating visual perception. Essential to aversive learning is the expectation and awareness of contingencies as well as anticipation of the unconditioned, aversive stimulus (US) following the CS+ in delayed classical conditioning (Rescorla & Wagner, 1972) . The dopaminergic system has been implicated in playing a key modulatory role during aversive learning in terms of both a cquisition and extinction (Horvitz, 2000; Abraham et al., 2014) . Utilizing Pavlovian fe ar conditioning in the laboratory , the interplay of associative learning, perceptual processing and defensive engagement can be systematically studied. The goal of this research is to characterize the variability in response patterns during aversive learni ng in an unselected sample of individuals. Fear Circuitry in the B rain Neuroimaging studies have found that several cortical an d subcortical regions are engaged during aversive learning. Subcortical limbic structures include the amygdala, insula, striatum and hippocampus and have been linked to critical functioning of the network involved during defensive engagement (LeDoux, 2000) . The prefrontal cortex (PFC) plays an integral role in the fear network in terms of attention, associative learning and working memory through top down modulation of higher order sensory cortices (Desimone, 1996; E. K. Miller et al., 1996; Asaad et al., 1998; Barcelo et al., 2000; Rainer & Miller, 2000; E. K. Miller & Cohen, 2001; B. T. Miller et al., 2011) . In view of aversive conditioning, reentrant modulation of the visual cortex is critical in not only object recognition, but in directing attention to the aversive stimulus. Animal studies ha ve revealed monosynaptic projects between PFC and higher order visual areas, whereby PFC and visual neurons are modulated by reciprocal activity displaying


14 plasticity changes (LTP/LTD) as well as fast occurring single neuron spiking activity during conditi oning (Baeg et al., 2001; Kim et al., 2003) . Aversive learning in rats was found to be regulated through the nucleus accumbens core and dors al regions of the striatum (Wendler et al., 2014) . Taken together, this body of work suggests an integral involvement of the PFC together w ith subcortical structures such as the amygdala and striatum during aversive learning . Engagement of the fear circuitry in preparation for action is not an exclusively reflexive response but can incorporate higher cognitive functioning such as attention, a nticipation, working memory and learning through the dopaminergic system. By targeting polymorphisms that result in varying degrees of neurotrophic factors and variable dopaminergic regulation in fronto striatal regions , we can assess the degree to which v isual perception is modulated through plasticity and higher cognitive functioning during aversive learning, respectively. Furthermore, an extensively used neuroimaging tool in the cognitive neurosciences is the steady state visually evoked potential (ssVEP ). Lower tier visual areas may be entrained at a given frequency by an external luminance or contrast modulated flickering stimulus, captured as oscillatory scalp voltage changes through electroencephalography (EEG). Measurable activity is thought to refl ect the evoked, synchronous activity of large masses of neurons in response to a stimulus. The ssVEP has been shown to be modulated by a multitude of experimental manipulations including selective attention, feature selection, emotional content and fear co nditioning (Silberstein et al., 1995; Moratti et al., 2004; Muller et al., 2006; Keil et al., 2009) . Animal work from Amaral and colleagues (2003) has demonstrated amygdala projections to not only extrastriate cortex (Tigges et al., 1982) , but directly to V1 of the striate (primary


15 visual) cortex. These findings support the notion that regions involved in defensive engagement in the mammalian brain exert top down influe nce on the visual cortex, leading to enhanced perceptual processing (Lang & Bradley, 2010) . This increased neural responding through reentrant modulation (Roelfsema, 2006) can be captured via the ssVEP. Cortical modulation by aversive stimuli as captured through EEG scalp potentials are therefore thought to reflect subcortical cortical interactions. Genetic Predisposition: the COMT P ol ymorphism Single nucleotide polymorphisms (SNPs) implicated in fronto striatal dopamine regulation as well as brain plasticity (neurotrophic factors) were investigated in this study. The COMTval158met SNP has been implicated in dopaminergic regulation of t he dorsolateral prefrontal cortex (PFC). This polymorphism impacts the effectiveness of catechol O methyltransferase (COMT), a regulatory enzyme in the metabolic pathways of catecholamines (dopamine, epinephrine and norepinephrine) centrally in fronto stri atal regions and peripherally in blood supply to thoracic organs (Weinshilboum & Dunnette, 1981) . A single base pair substitution on chromosome 22 results in the allelic transcription of either methionine (Met) or valine (Val), the latter of which contributes to a version of COMT with 40% higher enzymatic activity (Weinshilboum et al., 1999) . Physiologically, increased COMT activity leads to a more efficient elimination of extracellular dopamine (DA). The t onic phase dopamine hypothesis in the context of DA regulation by COMT relates the varying enzymatic activity levels associated with COMT genotypes to systematic changes in dopamine transmission in ventral striatal and PFC neurons (Bilder et al., 2004) . Specifically, high act ivity COMT (related to Val allele) is related to decreased tonic DA activity through D2 auto receptors, therefore lowering the threshold


16 for phasic DA activity in the striatum. The presence of salient stimuli has been associated with phasic DA bursts at po stsynaptic neuronal sites (Redgrave et al., 1999) . In the PFC, on the other han d, increased COMT activity results in lowered extracellular DA availability throughout the cortex for the Val allele. Due to a paucity in DA transporters (DAT) in the cortex, which in the striatum aids in reuptake of DA, breakdown by COMT and other mechani sms (MAO B and reuptake in noradrenergic terminals) are the major route by which DA is regulated in PFC. Recent work by Hirvonen and colleagues (2010) have demonstrated though that D2 baseline levels in cortex and striatum in humans do no differ as a function of COMT genotype, suggesting that exclusively D1 receptors may be impacted by this polymorphism ( in vivo PET). Furthermore, Yavich and colleagues (2007) demonstrated that with COMT KO mice exposed to an overflow of DA, subsequent elimination of DA was twofold slower in PFC than in striatum ( in vivo voltammetry). The authors conclude that COMT contributes to only about 50% of DA regulation in PFC. The complexity of COMT variability in fronto striatal n etworks has brought about contending hypotheses of DA regulation by COMT in r egards to cognitive and emotional functioning. Excessive or insufficient amounts of dopamine may lead to fronto striatal dysfunctioning in an inverted U shaped fashion . A trade of f effect by which increased dopamine may be beneficial for working memory, but detrimental for emotional sensitivity is plausible. Experimental work to tap into these cognitive functions often utilizes working memory tasks (Egan et al., 2001; Smolka et al., 2005; Drabant et al., 2006) . Val allele carriers have been implicated in having more emotional stability, bu t less cognitive flexibility during task switching and reversal learning, whereas the Met


17 allele carriers show the opposite effect. Conflicting findings from studies targeting specific populations as well as gender differences has complicated the picture o f how COMT may impact higher cognition though (Lee & Prescott, 2014). No clear genotype of the COMT val158met polymorphism that can directly be associated with either a more flexible or stable phenotype of cognitive functioning as is often suggested in the literature . In the current study, we ai med to investigate to what extent the fronto striatal dopaminergic system may play a modulatory role in visual perception of an aversive stimulus. Little work has been done in fear conditioning relating COMT genotype s to differential response patterns. Lonsdorf and colleagues (2009) found Met allele carriers to have failed suppression of the startle reflex in extinction, suggesting Met allele carriers possess diminished cognitive control over emotions and increased proneness to anxiety. This may implicate altered response patterns for Met versus nonMet allele carriers during fear condit ioning in regards to the startle reflex, but is not informative as to how this may affect perceptual processing. Inter Individual Differences during Defensive Engagement In terms of Pavlovian fear conditioning, Moratti and colleagues (2006) have demonstrated inter individual differences in heart rate react ivity related to stimulus discrimination in primary visual cortex. Conditioning effects in the visual cortex measuring ssVEFs (neuromagnetic counterpart of ssVEPs) were restricted to individuals exhibiting heart rate acceleration during conditioning. Heart rate acceleration has been taken as an indicator of defensive engagement of the autonomic nervous system. Importantly, Hodes and colleagues (1985) established inter individual differences in heart rate reactivity wh en undergoing fear conditioning in the laboratory. Heart rate acceleration was hypothesized to reflect engagement of the fear system,


18 whereas heart rate deceleration was a marker of learned association that elicited an orienting response. In the Moratti st udy, the authors conclude that heart rate acceleration may predict visual cortical facilitation of the fear stimulus. Additional indicators of defensive engagement can be measured through increased galvanic skin conductanc e response (SCR), as well as an en hanced startle reflex measured as blink magnitude described in the defense cascade model (Lang et al., 1997) . The startle reflex may occur without awareness, whereas enhanced skin conductance responding requires contingency awareness during aversive learning (Hamm & Vaitl, 1996; Hamm & Weike, 2005) . Peripheral phy siological measures are therefore good indicators of strength of engagement of the fear circuitry through the autonomic nervous system. Genetic predispositions may shed light on underlying, neurobiological mechanisms by which reflexive physiology is modula ted, leading to inter individual differences in aversive learning. Research G oal In the current study, we proposed that conditioning effects found in the EEG, expressed as increased activity for the perceptual processing of the aversively cued stimulus (CS +), would be related to selected genetic polymorphisms. For SNPs implicated in plasticity, we predicted genotypes resulting in an increased amount of neurotrophic factors and their receptors would show greater enhancement to the CS+. For the COMT val158met polymorphism we sought to investigate to what extent variability in dopamine regulation impacts conditioning effects in the electrocortical response, specifically visual perception of the aversive stimulus. If variability in dopamine regulation is related to enhanced fear activation, then this should be additionally reflected on measures of reflexive physiology. To study this, we


19 implemented an instructed, differential fear conditioning paradigm with an unselected sample of college students.


20 CHAPTER 2 METHODS Participants Participants for the study were recruited from the undergraduate pool of students taking Introduction to Psychology towards course credit at the University of Florida. A total of 91 participants were run in the study. Of these, 72 were included in the final analysis due to 19 subjects having noisy or insufficient data throughout the entire recording period. In addition, due to differences in successful DNA analysis for each polymorphism, the grand mean varied per polymorphism of individ uals having good EEG data. Demographics are reported for all participants run through the study. Participants ranged in age from 18 23 (M = 18.9, SD = 1.3), with the majority being right handed, female Caucasians (94.5%, 67.0% and 64.8%, respectively). The breakdown of minorities was as followed: 17.6% Hispanic, 11.0% Asian, 5.5% African American and 1.1% Middle Eastern. All participants had normal or corrected vision, and reported no personal or family history of epilepsy or photic seizures. Stimuli Visual stimuli for the study were Gabor gratings, designed to engage either luminance or chromatic visual pathways in the brain. These stimuli were flickered in pattern phase reversal fashion, entraining a steady state visually evoked potential (ssVEP) res ponse in the visual cortex, described later in this section in greater detail. Luminance Gabor gratings were alternating black and white patterns, while chromatic Gabor gratings were alternating red and green patterns. Two orientations for each of the Gabo r gratings (±45°) were loaded as images by Psychtoolbox (Brainard, 1997) for each trial iteration, implemented through Matlab. The orientations served as CS+/


21 conditions during the conditioning phase. Four different visual stimuli were therefore presented throughout the experiment: Gabor gratings (2) x Orientation (2). Audito ry noise (1 1kHz) at 96dB. The unconditioned stimulus (US) was white noise presented for 1.4 sec. A brief startle of white noise was presented at three different times th roughout every trail (either at 4 sec, 5 sec or 1 sec after US offset). Design and Procedure Upon arriving in the laboratory, participants provided informed consent, were seated in the experiment room in a chair 100 cm away from the presentation monitor an d given instructions on the experimental procedure. After assessing net size, participants were given the State and Trait Anxiety (STAI) questionnaire to fill out in the absence of the experimenter. Subsequently, participants were instructed on how to use the mouthwash (1.5 fl Oz Scope® bottle) for the saliva sample: vigorously swishing half the mouthwash for a minute, then spitting into the pre labelled test tube provided, and repeating with the other half. Test tube with saliva sample was then immediately taken by the experimenter wearing non latex gloves and placed in a cold storage box in the laboratory. Electrodes for recording peripheral physiological measures were placed, in order, on the left palm (skin conductance response), on either arm (heart rat e response) and underneath the left eye (startle response). EEG net was applied next and an impedance check was run. Lights were turned off completely before starting the experiment. There were a total of three phases throughout the experiment (habituation , acquisition and extinction), each containing a total of 24 trials (4 conditions x 6 repetitions). During habituation, extinction and non CS+ trials trial length was 7000 ms in duration, whereas the CS+ trials extended for an additional 1400 ms during whi ch the


22 US was simultaneously presented, with a 100% reinforcement schedule. After each phase there was a brief pause in the experiment where participant filled out SAM ratings (Bradley & Lang, 1994) for the Gabor gratings, reporting perceived hedonic valence and arousal. After the conditioning phase participants additionally filled out the State Anxiety portion (form Y1) of the STAI Q again, to obtain a pre and post measure of self reported state anxiety. Upon conclusion of the study, participants were asked to rate the US on a scale of 1 10 (higher scores being more unpleasant) and were subsequently debriefed. It is worth n oting that the study was run over a span of 2 years, during which new EEG equipment was upgraded, resulting in the use of two different sensor net layouts (GSN 200 257 channel sensor net and HCGSN 129 channel sensor net by EGI), as well as new a presentati on monitor (change from 60Hz to 120Hz refresh rate monitor) that resulted in multiple stimulation frequencies over the span of running the experiment. Details on how this was dealt with will be described in detail in the appropriate sections below. All pro cedures were approved by the institutional review board of the University of Florida. Data Collection and Processing Electroencephalography (EEG) Collection EEG was continuously recorded from either 257 or 129 electrodes using Electrical Geodesic (EGI) se nsor nets with Cz as the reference electrode. The HCGSN 129 channel net was only used in the last semester of running due to a system upgrade, while keeping all other recording parameters and procedures the same as before. EEG signal was digitized at 250Hz and band pass filtered online from 0.1 to 48 Hz, with


23 impedances being kept below 60 k for the 257 channel nets and 40 k for the 129 channel nets. Preprocessing Offline preprocessing of the data was implemented through EMEGS software (Peyk et al., 2011) , by filtering (band pass at 12 18Hz), segmenting (epochs of 7600ms, 400ms of which being pre trigger onset) and subsequently runnin g artifact rejection. The procedure for artifact rejection identifies artifacts in individual channels by referencing Cz (recording reference) and detecting deviations based on distribution of the mean, the standard deviation as well the gradient of the vo ltage amplitude. Data were then re referenced to a global average and artifacts such as blinks, eye movements and other movements were eliminated. Channels contaminated by such artifacts were then interpolated using a statistically weighted, spherical spli ne interpolation from the entire channel set. Processing Scalp voltage data were converted to source space data through the minimum norm estimation (MNE) method by solving the inverse solution. Data from the radial component of the source modeling in the f irst shell was utilized for further analysis with a regularization parameter of 0.08. Subsequently, source space data was converted to the frequency domain by running a discrete Fourier transformation (DFT) on the last two seconds of the evoked potential l eading up to US onset (Moratti et al., 2006) . Multi ple flicker rates were used throughout the study due to equipment changes (13.3Hz, 14Hz and 15Hz) and therefore a composite value was used throughout the EEG data analysis after conversion to the frequency domain. Due to source modeling creating excessive noise in ventral regions of the model, signal to noise ratios (SNR) were computed on


24 the frequency domain data by specifying the bin of the desired frequency and dividing by the average of the surrounding 5 bins to either side of the target bin. Trials of the same condition were then averaged together forming condition specific representations of the evoked response in source space. For better temporal resolution during the acquisition and extinction phases, trials were partitioned into 2 blocks per phase r esulting in 6 trials per condition per block. For illustrative purposes, source data was projected back onto a 257 channel net in the topographies presented in figures. For subsequent analyses, including genotype distributions, an occipital cluster group w as taken averaged over the nearest 9 neighbors to Oz. Here, results for only the luminance Gabor stimulus are reported due to preliminary analyses showing no cortical conditioning effects for the chromatic stimulus. This was expected due to findings in pre vious research (Keil et al., 2013) and the chromatic stimulus is largely used here as a control condition. Pe ripheral Physiological M easures A computer running VPM software (Cook, 2002) received triggers from the Matlab computer for stimulus events and controlled data acquisition for all peripheral physiological measures. Galvanic skin conductance responding (SCR) was recorded from 8mm electrod es filled with 0.5 M NaCl paste on the hypothenar eminence of the left palm and continuously sampled at 20 Hz. SCR was converted to microSiemens, averaged to half second bins and log transformed log(SCR+1). Cardiogram (heart rate, HR) was recorded through 8mm electrodes on each forearm, including a ground electrode, and R R peaks were recorded, reducing inter beat intervals to half second bins. Raw electromyogram (EMG) from the orbicularis oculi for the startle response was recorded through 4mm Ag AgCl elec trodes, sampled at 1000 Hz , rectified and


25 integrated using a 20 ms time constant. Reduced data from the VPM software was then imported to JMP software (JMP®) for further processing. Startles el icited at each time window during a trial were converted to t scores, using mean and standard deviations within each experimental block for the t distribution and averaged. All other data were processed as change scores from the baseline (500ms prior to st imulus onset) and averaged as condition specific responses. Luminance and chromatic stimuli are averaged together as no stimuli specific differences were found in preliminary analyses of peripheral measures. Self Report M easures The State and Trait Anxiety (STAI) questionnaire was given before the start of the experiment, as well as the State Anxiety part again after undergoing conditioning as previously described in the procedure. Scores were then manually recorded in a spreadsheet and cumulative scores pe r individual computed. After each phase of the experiment SAM ratings were taken for the Gabor gratings, and again recorded and computed for each individual. US ratings were noted at the end of the experiment. Genotyping Nine single nucleotide polymorphism s (SNPs) were preselected on the basis of their implications in brain plasticity: Brain derived neurotrophic factor (BDNF rs6265), catechol O methyltransferase (COMT rs4680 and rs4633), stathmin (STMN1 rs159525), neurotrophin 3 (NTF3 rs6332), neurotrophic tyrosine kinase, receptor type 1 (NTRK1 rs6337), nerve growth factor (NGFB rs6328), nerve growth factor receptor (NGFR rs534561) and gastrin releasing peptide (GRP rs1062557). Acquired from buccal cells contained in the mouthwash, these saliva samples were analyzed by the


26 Center for Pharmacogenomics Genotyping Core Lab at Shands at the University of Florida. Statistical Analyses The focus of the analyses was on group differences. A one way ANOVA was run on the difference scores of the ssVEP amplitudes (CS+ minus CS ) for each of the selected gene polymorphisms (comparing 3 genotypes per SNP), with SNP variant being the between subject factor. In polymorphisms where significant group differences in the EEG were found, subsequent analyses of the peripheral phy siological measures were conducted using repeated measures ANOVAs with factors of condition (CS+/ ), time segment (see Results section for description for each measure) and the between subjects factor of genotype. Experimental block was not included as a f actor to keep ANOVA models within a manageable size and to account for the fact that null results were expected for habituation. Self report measures were analyzed with factors of condition (CS+/ ), scores (again, see Results s ection) and the between subject factor of genotype. All analyses were followed up with t tests for comparisons of condition (CS+/ ) or COMT group (Met/nonMet) where appropriate. Due to consistent findings of gender differences in COMT literature (Domschke et al., 2012; Lee & Prescott, 2014) , gender was added as a covariate in the analyses of all dependent measures.


27 CHAPTER 3 RESULTS Electroencephalography Group differences were found for the COMT val158met polymorphism grouped by Met and nonMet allele carriers (Met/Met and Met/Val genotypes grouped in Met allele carriers), with difference scores of the ssVEP as the dependent measure. NonMet allele carriers showed significant enhancement to the CS+ over CS at an occipital cluster of 9 immediate neighbors to Oz [F(1,63) = 12.93, p < .05]. In the scatterplot of the individual distributions, group separation is visible (Figure 1) and remains statistically signi ficant when excluding the outlier visible in the nonMet group [F(1,62 ) = 11.05, p < .05]. This enhancement to the CS+ was limited to the first block of acquisition, and disappeared during the second block [F(1,63) = 2.09, n.s.]. Follow up t tests were run on each genotype testing CS+ versus CS activation and was found to be exclusive to the nonMet group (Val/Val) [t(23) = 2.72, p < .05] (Figure 2). Results of other polymorphisms are not reported here, as the focus of this work is on the COMT polymorphism. Subsequent analyses were run on peripheral physiological measures and self report data. Peripheral P hysiology Difference scores were computed for mean heart rate (HR) reactivity in time intervals 0 3sec (D1), 3 5sec (A1) and 5 7sec (D2), with the first dif ference score being A1 D1 (accelerative window) and the second D2 A1 (decelerative window). In the heart rate waveforms (Figure 3) one can see that both COMT groups show HR deceleration to the CS+ for the first half of conditioning [F(1,54) = 4.4, p < .05] . In the second block, main effects are seen for the two difference scores [F(1,54) = 4.29, p < .05], condition


28 (CS+/ ) [F(1,54) = 4.96, p < .05], as well as a three way interaction of difference scores x condition x COMT group (Met vs nonMet). In the fol low up analysis a two way interaction for CS+ responding was found, difference scores x COMT group [F(1,54) = 8.39, p < .05]. An independent t test for the between subject factor of COMT group showed nonMet individuals to have significant HR increase to th e CS+ during the accelerative window, whereas Met individuals showed deceleration [t(54) = 2.37, p < .05] (Figure 4]. For skin conductance responding (SCR) there was a main effect of condition (CS+/ ) in the first block of acquisition for SCR to the CS+ [ F(1,55) = 13.88, p < .001] and US presentation [F(1,55) = 44.06, p < .001] (Figure 5). In addition, a significant interaction of condition x COMT group was found [F(1,55) = 5.72, p < .05], for a difference in the CS [t(55) = 2.56, p < .05, adjusted due t = 6.02, p < .05]] and not the CS+ [t(55) = 1.78, p = .08]. Here, Met allele carriers showed a steeper decline in SCR in CS trials and a larger trending response to the US in CS+ trials following the offset of the visual stim ulus. The main effect of condition disappears in the second half of acquisition and there is no interaction with COMT genotype. Startle response (STR) showed a main effect of probe position (3) during block 1 and 2 of acquisition as well as block 1 of ext inction [F(2,160) = 4.06, p < .05, F(2,158) = 11.76, p < .001, F(1, 156) = 5.74, p < .05], with the interaction of latency with condition (CS+/ ) being more informative (Figure 6). In block 1 of acquisition there was a significant interaction of probe posi tion and condition [F(2,160) = 13.47, p < .001] with the second startle (5 sec into trial) showing significant enhancement for the CS+ over


29 CS [t(80) = 4.43, p < .001]. For block 2 of acquisition an interaction of probe position and condition [F(2, 158) = 6.92, p < .05] revealed significant differences of CS+ to CS for the third startle (9.6 sec into trial) [t(80) = 3.65, p < .001], in which the CS was significantly larger. This may be a refractory effect as the startle occurred 1000 ms after the US pres entation in CS+ trials. In the first block of extinction, in addition to a main effect of probe position [F(2,156) = 5.74, p < .05] there is a main effect of condition [F(1,78) = 8.42, p < .05], and an interaction of probe position and condition [F(2,156) = 10.40, p < .001]. Here, startle 2 and 3 show significant enhancement to the CS+ [t(78) = 2.45, p < .05, t(78) = 3.87, p < .001, respectively], whereas startle 1 (4 sec into trial) shows an enhancement to CS [t(78) = 2.10, p < .05]. In the second block of extinction there was only a n interaction of probe position and condition [F(2,156) = 3.60, p < .05] in which startle 1 shows significant enhancement to the CS+ [t(80) = 2.02, p < .05]. No between subject effect of COMT genotype or gender was found or significant interactions for STR. Self R eport For ratings of the US, SAM ratings of the stimuli, and the State Anxiety portion of the STAI Q, no significant group differences were found for the COMT genotype. All these measures ind icated though that conditioning occurred (see Table 1). US ratings indicated that participants experienced the US as unpleasant (mean = 6.1 on a 1 10 scale), but no COMT group differences (F(1,62) = .80, n.s.]. State Anxiety ratings were taken before and a fter conditioning, with the repeated measures ANOVA revealing that participants reported higher state anxiety scores after conditioning [F(1,48) = 8.49, p < .05], but again no COMT group difference [F(1,48) = .1, n.s.]. For the SAM ratings, participants ov erall rated the CS+ as more arousing during conditioning [F(1,54) =


30 57.15, p < .001] and more unpleasant than the CS [F(1,54) = 37.15, p < .001], compared to their ratings during habituation. Again, no COMT group differences were found for arousal [F(1,54 ) = 1.60, n.s.] nor for valence [F(1,54) = .81, n.s.]. The Trait Anxiety portion of the STAI Q did reveal significant differences group differences [F(1,53) = 8.61, p < .05], with nonMet individuals scoring significantly higher. No gender differences were found in any of the measures.


31 CHAPTER 4 DISCUSSION Summary In the current study, we investigated inter individual differences in response patterns during aversive learning. To this end, single nucleotide polymorphisms (SNPs) were selected on the basis of their implications in fronto striatal involvement and brain plasticity. The COMT val158met polymorphism revealed significant group differences between genotypes, with individuals homozygous on the Val allele displaying significant enhancements in visual p erception of the aversively cued stimulus (CS+). This enhancement in the EEG signal was specific to initial conditioning and was followed by subsequent heart rate acceleration. With no group differences on other measures of fear engagement (startle respons e and skin conductance response), individuals grouped by COMT genotype did not differ in terms of defensive engagement . Effects found in heart rate reactivity may have implications of peripheral catecholamine regulation through COMT and its impact on the a utonomic nervous system (ANS). Findings of enhanced visual perceptual during initial conditioning may be interpreted as an attentional effect for individuals of the high enzymatic activity Val/Val genotype. The Role of Fronto Striatal Modulation during Ave rsive L earning In the present findings, we show clear cortical enhancement in the primary visual cortex during initial conditioning for individuals of the Val/Val genotype, reflected as an ssVEP amplitude increase to the CS+. This effect reliably replicate s over two separate SNP codon regions for the COMT polymorphism (rs4680 and rs4633). In addition, this conditioning effect dissipates over time, suggesting quick initial engagement of the system. As the COMT polymorphism impacts PFC and striatal dopamine a vailability, the


32 current findings support the idea that changes in primary visual cortex may be due to fronto striatal top down modulation. The manner by which s ignal integration between PFC, amygdala and hippocampus through ventral striatum may be modulat ed through variable dopamine levels is beyond the scope of this study. It is possible that enhancement in primary visual cortex may be due to faster engagement of amygdala or PFC through enhanced dopamine in ventral striatum, but this would need to be dire ctly measured through other neuro imaging techniques such as fMRI or MR spectroscopy. I ncreased perceptual processing in the primary visual cortex to the aversively cued stimulus may ultimately be interpreted as an attentional effect. Previous research asso ciates the Met allele with greater enhancement in amygdala activity and connectivity, self report measures of anxiety and startle potentiation during viewing of aversive stimuli (Drabant et al., 2006; Smolka et al., 2007; Montag et al., 2008) . Lonsdorf and colleagues (2009) in particular found failed suppression of the startle response during extinction in a fear conditioning paradigm, lending to the proposed idea that Met allele carriers are more susceptible to stress and have attenuated stress recovery (see warrior/worrier model, Goldman et al., 2005; Alexander et al., 2011) . The present study does not support earlier notions that Met allele carriers are more susceptible to stress, have increased fear respondin g or that Val allele carriers are particularly resilient to stress. In the current study, individuals homozygous on the Val allele show quick, initial cortical enhancement in the primary visual cortex to the aversively cued stimulus during conditioning. Th is finding supports the notion that the Val/Val genotype has facilitated network engagement that is reflected as increased attention during initial conditioning. It is worth noting though that individuals of the


33 Val/Val genotype did score significantly hig her on the Trait Anxiety portion of the STAI Q. Fear and anxiety patients have been shown to have deficiencies in fear responding in terms of more rapid acquisition and slower extinction (Lissek et al., 2005) . The vast majority of research in the COMT po lymorphism focuses on disorders such as (Eisenberg et al., 1999; Weinberger et al., 2001; Goldman et al., 2005; Norrholm et al., 2013) , all related to dysfunctional dopamine systems. It would be o f interest to expand the current findings of aversive learning in an unselected sample to a clinical population with fear and anxiety disorders. COMT Polymorphism I mpacts the ANS Work by Jabbi and colleagues (2007) have demonstrated that the COMT val158met polymorphism is not only implicated in central catecholamine regulation but peripheral as well, importantly as part of t he endocrine response to stress. Healthy individuals showed increased endocrine response measured via epinephrine with Met allele loading. Although the authors caution in making assumptions of stress resiliency and susceptibility, these findings show a dir ect link of variability in COMT enzymatic activity with plasma catecholamines such as epinephrine. Recent work has further expanded this concept in finding COMT genotype differences in ANS activity in children (Mueller et al., 2012) . In this study, children between the ages of 8 11 were genotyped for the COMT val158met polymorphism and exposed to a social stressor while heart rate reactivity was recorded. Individuals homozygous on the Val allele showed the greatest increase in heart rate, as well as the steepest decline in heart rate variability (HRV), a measure of parasympathetic activity. These findings suggest stronger sympathetic innervation for the Val/Val genotyp e, although no direct measure of


34 sympathetic activity was recorded. The authors conclude that the COMT polymorphism differentially impacts both branches of the ANS. Our results would suggest that it is the parasympathetic nervous system that is impacted by COMT activity, as we found no group differences for sympathetic measures. In the case of individuals homozygous on the Val allele, the reported heart rate acceleration may in fact be due to vagal release. Vagal tone normally inhibits the activity of the sinoatr ial node, the hearts pace maker and can become disinhibited without the innervation of the sympathetic nervous system. Vagal release results in an increase in heart rate, solely through the actions of the parasympathetic nervous system. Heart rate v ariability could offer better insight as to the strength of parasympathetic activity and is a point of interest for future study. Work by Hansen and colleagues (Hansen et al., 2003) has demonstrated that individuals with higher HRV perform better on tasks involving executive functioning such as working memory and attention, as well as some work indicating a correlation of high HRV to sustained attention in children (Suess et al., 1994) . Again, future studies would need to explicitly expl ore a possible relationship of COMT genotype with HRV and its effects on attention. In the current study though, individuals of the Val/Val genotype exhibited differential responding of the ANS, suggesting that peripheral catecholamine regulation is also i nfluenced by the COMT val158met polymorphism. Limitations and Future D irections We conclude from the current findings that fronto striatal dopamine varies as a function of COMT genotype, differentially impacting the fear circuitry, but in particular modula ting primary visual cortex. The COMT polymorphism is an indirect measure of this assumed influence though and future studies may benefit from utilizing


35 neuroimaging techniques to investigate fronto striatal and primary visual activity during aversive learn ing in relation to COMT genotypes. As far as visual stimuli engaging the system, only major results for the luminance stimulus were reported. Tapping into separate visual pathways through magno , parvo and konio activating stimuli may help elucidate feedb ack connections to primary visual cortex and how they are affected during aversive learning. In the same vein, COMT literature in primate research is sparse if nonexistent, and could significantly contribute to our understanding of how this polymorphism im pacts primate visual cortex through fronto striatal functioning. Although the overall sample in this study was relatively small to explore genotype differences within each polymorphism, the COMT val158met SNP showed strong conditioning effects in the visua l cortex for the Val/Val genotype in a robust fashion. Although no demographic differences were found in the current study, an exciting body of work has drawn attention to dopamine regulation in the PFC over the entire human lifespan and how it is affected by COMT variability (Tunbridge et al., 2007) . In terms of how the COMT polymorphism may differentially impact ANS functioning, we concluded that the parasympathetic branch shows allele dependent modulation. Future research is necessary to expand on these findings, as we found Trait Anxiety differences (Val allele higher scores) and Met carriers appeared to have stronger skin conductance responding to the US in CS+ trials (trending at p = .08) and stronger decline in SCR to the CS . This ma y indicate the possibility of stronger sympathetic responding, as well as greater sensitivity in responding for Met carriers. A further avenue of study for the COMT polymorphism, as well as other selected plasticity genes, could be in an experimental parad igm shift where participants are not


36 informed of the contingency. An uninstructed fear conditioning paradigm would therefore allow for a much more fine grained analysis of the temporal dynamics of aversive learning through single trial analysis. Concluding R emarks In conclusion, the current findings indicate that individuals homozygous on the high activity Val allele of the COMT val158met polymorphism demonstrate quick initial conditioning effects in primary visual cortex to the aversively cued stimulus. This may be reflective of top down modulation of fronto striatal network for this genotype, p resumably facilitating attention allocation to the aversively cued stimulus. The manner by which this variable top down modulation occurs appears to be a function of dopamine regulation, although the precise mechanism remains unclear. Furthermore, the current study points to the possibility that the COMT polymorphism may be related to peripheral catecholamine regulation, with COMT genotypes differentially impacting the ANS, sp ecifically the parasympathetic branch. Replication of the current findings is essential for translational research, as the underlying neurobiological mechanisms need to be understood before relating these to deficient fear responding in clinical population s.


37 Figure 3 1 . Distribution of conditioning effects of EEG scalp potentials captured through the ssVEP by genotype grouping of the COMT val158met polymorphism during the first half of conditioning. Positive values indicate greater amplitude enhancemen t to the CS+, negative values for the CS . Circles represent single individuals within a group.


38 Figure 3 2. Activation over occipital pole to CS+/ stimuli measured through ssVEP duri ng first half of conditioning. N onMet carriers (n=24) showed significan t enhancement (scale in V, SNR corrected) to the CS+ over CS , whereas Met carriers (n=41) showed no differentiation.


39 Figure 3 3. Heart rate reactivity over conditioning phase, split into two blocks. US onset is at 7 sec (thick line) for 1.4 sec. Red signifies response to CS+, green to CS . NonMet carriers (n=19) and Met carriers (n=37) do not differ in first block of conditioning in their response to the CS+, whereas in the second block nonMet carriers show significant increase in heart rate to the CS+.


40 Figure 3 4. Heart rate reactivity in accelerative window during the second block of conditioning. Mean difference scores are depicted here and show a clear acceleration to CS+ presentation for nonMet carriers, whereas Met carriers show a relative deceleration.


41 Figure 3 5. Skin conductance responding during conditioning. Red lines indicate Met allele carriers response during CS+ trials, blue lines indicate nonMet carriers during CS+ trials and black lines indicate responding during CS trials for both COMT genotype groups. C onditioning effects are visible during CS presentation, with much stronger responding occurring after US presentation. No group differences in COMT genotype were found for conditioning effects in SCR.


42 Figure 3 6. Blink magnitude of the startle re sponse duri ng conditioning. C hanged to T scores against the mean and standard deviation within each block. Startle latencies correspond to first (4 sec), second (5 sec) and third startle (9.6 sec). Red indicates CS+ and green CS . Clear startle potentiatio n in the first half of acquisition for the second startle is visible. Overall no COMT group differences found.


43 Table 3 1. Self report measures by COMT genotype grouping. SAM ratings are reported in change from habituation to conditioning for CS+ stimulus only. STAI Q has two ratings for State portion (S R1 before, S R2 after conditioning) and Trait scores (T). US rating was given on a 1 10 scale, 10 being very unpleasant. Group differences for COMT were only found for Trait Anxiety, with nonMet individuals reporting higher Trait Anxiety scores. SAM ratings STAI Q US Arousal Valence S R1 S R2 T nonMet 1.53(1.26) 1.76(1.58) 33.50 (10.34) 37.69(13.46) 39.16(9.77) 6.17(2.12) Met 1.47(1.90) 1.23(1.29) 32.15(9.20) 37.35(10.21) 33.29(6.92) 5.99(1.71)


44 LIST OF REFERENCES Abraham, A. D., Neve, K. A., & Lattal, K. M. (2014). Dopamine and extinction: a convergence of theory with fear and reward circuitry. Neurobiol Learn Mem, 108 , 65 77. doi: 10.1016/j.nlm.2013.11.007 Alexander, N., Osinsky, R., Mueller, E., Schmitz, A., Guenthert, S., Kuepper, Y., & Hennig, J. (2011). Genetic variants within the dopaminergic system interact to modulate endocrine stress reactivity and recovery. Behav Brain Res, 216 (1), 53 58. doi: 10.1016/j.bbr.2010.07.003 Amaral, D. G., Behniea, H., & Kelly, J. L. (2003). Topographic organization of projections from the amygdala to the visual cortex in the macaque monkey. Neuroscience, 118 (4), 1099 1120. Asaad, W. F., Rainer, G., & Miller, E. K. (1998). Neural activity in the primate prefrontal cortex during associative learning. Neu ron, 21 (6), 1399 1407. Baeg, E. H., Kim, Y. B., Jang, J., Kim, H. T., Mook Jung, I., & Jung, M. W. (2001). Fast spiking and regular spiking neural correlates of fear conditioning in the medial prefrontal cortex of the rat. Cereb Cortex, 11 (5), 441 451. B arcelo, F., Suwazono, S., & Knight, R. T. (2000). Prefrontal modulation of visual processing in humans. Nat Neurosci, 3 (4), 399 403. doi: 10.1038/73975 Bilder, R. M., Volavka, J., Lachman, H. M., & Grace, A. A. (2004). The catechol O methyltransferase poly morphism: relations to the tonic phasic dopamine hypothesis and neuropsychiatric phenotypes. Neuropsychopharmacology, 29 (11), 1943 1961. doi: 10.1038/sj.npp.1300542 Bradley, M. M., & Lang, P. J. (1994). Measuring emotion: the Self Assessment Manikin and th e Semantic Differential. J Behav Ther Exp Psychiatry, 25 (1), 49 59. Brainard, D. H. (1997). The Psychophysics Toolbox. Spat Vis, 10 (4), 433 436. Buchel, C., Morris, J., Dolan, R. J., & Friston, K. J. (1998). Brain systems mediating aversive conditioning: an event related fMRI study. Neuron, 20 (5), 947 957. Cook, E. W., 3rd. (2002). VPM reference manual . Birmingham, AL: Author. Desimone, R. (1996). Neural mechanisms for visual memory and their role in attention. Proc Natl Acad Sci U S A, 93 (24), 13494 13499. Domschke, K., Baune, B. T., Havlik, L., Stuhrmann, A., Suslow, T., Kugel, H., . . . Dannlowski, U. (2012). Catechol O methyltransferase gene variation: impact on amygdala response to aversive stimuli. Neuroimage, 60 (4), 2222 2229. doi: 10.1016/j.neuroimage.2012.02.039


45 Drabant, E. M., Hariri, A. R., Meyer Lindenberg, A., Munoz, K. E., Mattay, V. S., Kolachana, B. S., . . . Weinberger, D. R. (2006). Catechol O methyltransferase val158met genotype and neural mechanisms related to affective arousal and regulation. Arch Gen Psychiatry, 63 (12), 1396 1406. doi: 10.1001/archpsyc.63.12.1396 Egan, M. F., Goldberg, T. E., Kolachana, B. S., Callicott, J. H., Mazzanti, C. M., Straub, R. E., . . . Weinberger, D. R. (2001). Effect of COMT Val108/158 Met genotype on frontal lobe function and risk for schizophrenia. Proc Natl Acad Sci U S A, 98 (12), 6917 6922. doi: 10.1073/pnas.111134598 Eisenberg, J., Mei Tal, G., Steinberg, A., Tartakovsky, E., Zohar, A., Gritsenko, I., . . . Ebstein, R. P. (1999). Haplo type relative risk study of catechol O methyltransferase (COMT) and attention deficit hyperactivity disorder (ADHD): association of the high enzyme activity Val allele with ADHD impulsive hyperactive phenotype. Am J Med Genet, 88 (5), 497 502. Goldman, D., Oroszi, G., & Ducci, F. (2005). The genetics of addictions: uncovering the genes. Nat Rev Genet, 6 (7), 521 532. doi: 10.1038/nrg1635 Hamm, A. O., & Vaitl, D. (1996). Affective learning: awareness and aversion. Psychophysiology, 33 (6), 698 710. Hamm, A. O ., & Weike, A. I. (2005). The neuropsychology of fear learning and fear regulation. Int J Psychophysiol, 57 (1), 5 14. doi: 10.1016/j.ijpsycho.2005.01.006 Hansen, A. L., Johnsen, B. H., & Thayer, J. F. (2003). Vagal influence on working memory and attention . Int J Psychophysiol, 48 (3), 263 274. doi: 10.1016/s0167 8760(03)00073 4 Hirvonen, M. M., Nagren, K., Rinne, J. O., Pesonen, U., Vahlberg, T., Hagelberg, N., & Hietala, J. (2010). COMT Val158Met genotype does not alter cortical or striatal dopamine D2 rec eptor availability in vivo. Mol Imaging Biol, 12 (2), 192 197. doi: 10.1007/s11307 009 0257 5 Hodes, R. L., Cook, E. W., 3rd, & Lang, P. J. (1985). Individual differences in autonomic response: conditioned association or conditioned fear? Psychophysiology, 22 (5), 545 560. Horvitz, J. C. (2000). Mesolimbocortical and nigrostriatal dopamine responses to salient non reward events. Neuroscience, 96 (4), 651 656. Jabbi, M., Kema, I. P., van der Pompe, G., te Meerman, G. J., Ormel, J., & den Boer, J. A. (2007). C atechol o methyltransferase polymorphism and susceptibility to major depressive disorder modulates psychological stress response. Psychiatr Genet, 17 (3), 183 193. doi: 10.1097/YPG.0b013e32808374df JMP®. (1989 2007). SAS Institute Inc. Cary, NC.


46 Keil, A., Miskovic, V., Gray, M. J., & Martinovic, J. (2013). Luminance, but not chromatic visual pathways, mediate amplification of conditioned danger signals in human visual cortex. Eur J Neurosci, 38 (9), 3356 3362. doi: 10.1111/ejn.12316 Keil, A., Sabatinelli, D. , Ding, M., Lang, P. J., Ihssen, N., & Heim, S. (2009). Re entrant projections modulate visual cortex in affective perception: evidence from Granger causality analysis. Hum Brain Mapp, 30 (2), 532 540. doi: 10.1002/hbm.20521 Kessler, R. C., Petukhova, M., S ampson, N. A., Zaslavsky, A. M., & Wittchen, H. U. (2012). Twelve month and lifetime prevalence and lifetime morbid risk of anxiety and mood disorders in the United States. Int J Methods Psychiatr Res, 21 (3), 169 184. doi: 10.1002/mpr.1359 Kim, M. J., Chun , S. K., Kim, Y. B., Mook Jung, I., & Jung, M. W. (2003). Long term potentiation in visual cortical projections to the medial prefrontal cortex of the rat. Neuroscience, 120 (1), 283 289. doi: 10.1016/s0306 4522(03)00021 6 Lang, P. J., & Bradley, M. M. (201 0). Emotion and the motivational brain. Biol Psychol, 84 (3), 437 450. doi: 10.1016/j.biopsycho.2009.10.007 Lang, P. J., Bradley, M. M., & Cuthbert, B. N. (1997). Motivated attention: Affect, activation, and action. In P. J. Lang, R. F. Simmons & M. T. Bala ban (Eds.), Attention and orienting: Sensory and motivational processes (pp. 97 135). Hillsdale, NJ: Erlbaum. Lang, P. J., McTeague, L. M., & Bradley, M. M. (2014). Pathological anxiety and function/dysfunction in the brain's fear/defense circuitry. Restor Neurol Neurosci, 32 (1), 63 77. doi: 10.3233/RNN 139012 LeDoux, J. E. (2000). Emotion circuits in the brain. Annu Rev Neurosci, 23 , 155 184. doi: 10.1146/annurev.neuro.23.1.155 Lee, L. O., & Prescott, C. A. (2014). Association of the catechol O methyltrans ferase val158met polymorphism and anxiety related traits: a meta analysis. Psychiatr Genet, 24 (2), 52 69. doi: 10.1097/ypg.0000000000000018 Lissek, S., Powers, A. S., McClure, E. B., Phelps, E. A., Woldehawariat, G., Grillon, C., & Pine, D. S. (2005). Clas sical fear conditioning in the anxiety disorders: a meta analysis. Behav Res Ther, 43 (11), 1391 1424. doi: 10.1016/j.brat.2004.10.007 Lonsdorf, T. B., Weike, A. I., Nikamo, P., Schalling, M., Hamm, A. O., & Ohman, A. (2009). Genetic gating of human fear le arning and extinction: possible implications for gene environment interaction in anxiety disorder. Psychol Sci, 20 (2), 198 206. doi: 10.1111/j.1467 9280.2009.02280.x McTeague, L. M., & Lang, P. J. (2012). The anxiety spectrum and the reflex physiology of d efense: from circumscribed fear to broad distress. Depress Anxiety, 29 (4), 264 281. doi: 10.1002/da.21891


47 Miller, B. T., Vytlacil, J., Fegen, D., Pradhan, S., & D'Esposito, M. (2011). The prefrontal cortex modulates category selectivity in human extrastria te cortex. J Cogn Neurosci, 23 (1), 1 10. doi: 10.1162/jocn.2010.21516 Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annu Rev Neurosci, 24 , 167 202. doi: 10.1146/annurev.neuro.24.1.167 Miller, E. K., Erickson, C. A., & Desimone, R. (1996). Neural mechanisms of visual working memory in prefrontal cortex of the macaque. J Neurosci, 16 (16), 5154 5167. Montag, C., Buckholtz, J. W., Hartmann, P., Merz, M., Burk, C., Hennig, J., & Reuter, M. (2008). COMT genetic variation affects fear processing: psychophysiological evidence. Behav Neurosci, 122 (4), 901 909. doi: 10.1037/0735 7044.122.4.901 Moratti, S., Keil, A., & Miller, G. A. (2006). Fear but not awareness predicts enhanced sensory processi ng in fear conditioning. Psychophysiology, 43 (2), 216 226. doi: 10.1111/j.1464 8986.2006.00386.x Moratti, S., Keil, A., & Stolarova, M. (2004). Motivated attention in emotional picture processing is reflected by activity modulation in cortical attention ne tworks. Neuroimage, 21 (3), 954 964. doi: 10.1016/j.neuroimage.2003.10.030 Mueller, A., Strahler, J., Armbruster, D., Lesch, K. P., Brocke, B., & Kirschbaum, C. (2012). Genetic contributions to acute autonomic stress responsiveness in children. Int J Psycho physiol, 83 (3), 302 308. doi: 10.1016/j.ijpsycho.2011.11.007 Muller, M. M., Andersen, S., Trujillo, N. J., Valdes Sosa, P., Malinowski, P., & Hillyard, S. A. (2006). Feature selective attention enhances color signals in early visual areas of the human brai n. Proc Natl Acad Sci U S A, 103 (38), 14250 14254. doi: 10.1073/pnas.0606668103 Norrholm, S. D., Jovanovic, T., Smith, A. K., Binder, E., Klengel, T., Conneely, K., . . . Ressler, K. J. (2013). Differential Genetic and Epigenetic Regulation of catechol O m ethyltransferase is Associated with Impaired Fear Inhibition in Posttraumatic Stress Disorder. Front Behav Neurosci, 7 , 30. doi: 10.3389/fnbeh.2013.00030 Peyk, P., De Cesarei, A., & Junghofer, M. (2011). ElectroMagnetoEncephalography software: overview and integration with other EEG/MEG toolboxes. Comput Intell Neurosci, 2011 , 861705. doi: 10.1155/2011/861705 Rainer, G., & Miller, E. K. (2000). Effects of visual experience on the representation of objects in the prefrontal cortex. Neuron, 27 (1), 179 189. R edgrave, P., Prescott, T. J., & Gurney, K. (1999). Is the short latency dopamine response too short to signal reward error? Trends Neurosci, 22 (4), 146 151.


48 Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. In A. H. Black & W. F. Prokasy (Eds.), Classical conditioning II (pp. 64 99): Appleton Century Crofts. Roelfsema, P. R. (2006) . Cortical algorithms for perceptual grouping. Annu Rev Neurosci, 29 , 203 227. doi: 10.1146/annurev.neuro.29.051605.112939 Silberstein, R. B., Ciorciari, J., & Pipingas, A. (1995). Steady state visually evoked potential topography during the Wisconsin card sorting test. Electroencephalogr Clin Neurophysiol, 96 (1), 24 35. Smolka, M. N., Buhler, M., Schumann, G., Klein, S., Hu, X. Z., Moayer, M., . . . Heinz, A. (2007). Gene gene effects on central processing of aversive stimuli. Mol Psychiatry, 12 (3), 307 3 17. doi: 10.1038/ Smolka, M. N., Schumann, G., Wrase, J., Grusser, S. M., Flor, H., Mann, K., . . . Heinz, A. (2005). Catechol O methyltransferase val158met genotype affects processing of emotional stimuli in the amygdala and prefrontal cortex . J Neurosci, 25 (4), 836 842. doi: 10.1523/JNEUROSCI.1792 04.2005 Stolarova, M., Keil, A., & Moratti, S. (2006). Modulation of the C1 visual event related component by conditioned stimuli: evidence for sensory plasticity in early affective perception. Cere b Cortex, 16 (6), 876 887. doi: 10.1093/cercor/bhj031 Suess, P. E., Porges, S. W., & Plude, D. J. (1994). Cardiac vagal tone and sustained attention in school age children. Psychophysiology, 31 (1), 17 22. Tigges, J., Tigges, M., Cross, N. A., McBride, R. L ., Letbetter, W. D., & Anschel, S. (1982). Subcortical structures projecting to visual cortical areas in squirrel monkey. J Comp Neurol, 209 (1), 29 40. doi: 10.1002/cne.902090104 Tunbridge, E. M., Weickert, C. S., Kleinman, J. E., Herman, M. M., Chen, J., Kolachana, B. S., . . . Weinberger, D. R. (2007). Catechol o methyltransferase enzyme activity and protein expression in human prefrontal cortex across the postnatal lifespan. Cereb Cortex, 17 (5), 1206 1212. doi: 10.1093/cercor/bhl032 Wangelin, B. C., Low, A., McTeague, L. M., Bradley, M. M., & Lang, P. J. (2011). Aversive picture processing: effects of a concurrent task on sustained defensive system engagement. Psychophysiology, 48 (1), 112 116. doi: 10.1111/j.1469 8986.2010.01041.x Weinberger, D. R., Egan, M. F., Bertolino, A., Callicott, J. H., Mattay, V. S., Lipska, B. K., . . . Goldberg, T. E. (2001). Prefrontal neurons and the genetics of schizophrenia. Biol Psychiatry, 50 (11), 825 844. Weinshilboum, R. M., & Dunnette, J. (1981). Thermal stability and the biochemical genetics of erythrocyte catechol O methyl transferase and plasma dopamine beta hydroxylase. Clin Genet, 19 (5), 426 437.


49 Weinshilboum, R. M., Otterness, D. M., & Szumlanski, C. L. (1999). Methylation pharmacogenetics: catechol O methyltrans ferase, thiopurine methyltransferase, and histamine N methyltransferase. Annu Rev Pharmacol Toxicol, 39 , 19 52. doi: 10.1146/annurev.pharmtox.39.1.19 Wendler, E., Gaspar, J. C., Ferreira, T. L., Barbiero, J. K., Andreatini, R., Vital, M. A., . . . Da Cunha , C. (2014). The roles of the nucleus accumbens core, dorsomedial striatum, and dorsolateral striatum in learning: performance and extinction of Pavlovian fear conditioned responses and instrumental avoidance responses. Neurobiol Learn Mem, 109 , 27 36. doi : 10.1016/j.nlm.2013.11.009 Yavich, L., Forsberg, M. M., Karayiorgou, M., Gogos, J. A., & Mannisto, P. T. (2007). Site specific role of catechol O methyltransferase in dopamine overflow within prefrontal cortex and dorsal striatum. J Neurosci, 27 (38), 1019 6 10209. doi: 10.1523/JNEUROSCI.0665 07.2007


50 BIOGRAPHICAL SKETCH L. Forest Gruss is originally from Zurich, Switzerland. She graduated from the math and science high school Ramibuhl in Zurich and subsequently received her degree Bachelor of Science at the University of Florida in Psychology. At the start of her undergraduate career at the University of Florida, Forest took a wide range of courses as an English major, taking several science and humanities courses , but f ocused heavily on neuroscience in her last two years . Having found her niche in neuroscience, she switched her major to psychology. The decision to take a science route was not an obvious choic e, as music and literature were always two other main interests of Forest. She was very active in music, being a member of several orchestras, ensembles and choir. Ultimately , the idea of not pursuing intellectual curiosity was not an option. Forest velopmental lab at UF for a year, but found her real interest in studying cognitive neuroscience at the Center of the Study of Emotion and Attention (CSEA) under the guidance of Dr. Andreas Keil. After getting her b achelor degree, she started as a post baccalaureate under Dr. Keil and gained training as a research assistant. She then continued her work at the CSEA as a graduate student in the Behavioral and Cognitive Neuroscience (BCN) program in psychology with Dr. Keil as her mentor, where she is continuing her research in pursuit of the degree Doctor of Philosophy in P sychology.