JournalofResearch inReading,ISSN0141-0423 Volume 28,Issue3, 2005,pp274–301 Letter-position encoding and dyslexia Carol Whitney University of Maryland,USA Piers Cornelissen University of Newcastle,UK This article focuses on applying the SERIOL model of orthographic processing to dyslexia. The model is extended to include a phonological route and reading acquisition. We propose that the temporal alignment of serial orthographic and phonologicalrepresentationsisakeyaspectoflearningtoread,drivingtheformation of a phonemic encoding. The phonemic encoding and the serial representations are mutually reinforcing, leading to automatic, proficient processing of letter strings. A breakdown in any component of this system leads to the failure to form string- specificphonologicalandvisualrepresentations,resultinginimpairedreadingability. FollowingthepioneeringworkofLibermanandcolleagues(1974),researchintodyslexia hasfocusedonphonologicaldeficits.AsdiscussedbyCastlesandColheart(2004),thereis awiderangeofevidencethatdyslexicsareimpairedinphonologicalawarenesstasks,such asphonemedeletion,phonemecountingandphonemelending.Thiscorrelationhaslargely been taken to reflect causality. That is, impaired phonological awareness is thought to reflect abnormal phonological representations, which are thought to be the fundamental cause of dyslexia. However, a causal relationship between phonological awareness and readingabilityhasnotdirectlybeenestablished(Castles&Coltheart,2004).Itmayinstead bethecasethatpoorperformanceonphonologicaltasksisaresultofpoorreadingability. Indeed, several studies have shown that phoneme awareness tasks are influenced by properties of the corresponding orthographic representations (Castles et al., 2003; Stuart, 1990;Treiman&Cassar,1997),indicatingthatperformanceonphonologicaltasksmaytap into letter-based, rather than purely phonemic, representations. Thus,deficitsofphonologicalawarenessmaybeasymptom,ratherthanadirectcause, of dyslexia; it may be more accurate to say that dyslexia results from the failure to form normal grapheme-phoneme associations (Castles & Coltheart, 2004). Thus such a failure could potentially arise from the grapheme (visual) side, or from problems in forming the relevantvisual-auditoryassociations.Adeficitinanyaspectofprocessingcouldthenhave repercussionsthroughouttheentirenetworkthatisnormallyrecruitedtosub-servereading. To study dyslexia, it is useful to determine when and where processing first diverges from that of normal readers. MEG technology has revealed that this divergence occurs quiteearly,atthevisuallevel.Innormalsubjects,aleft-hemisphere(LH)infero-temporal rUnitedKingdomLiteracyAssociation2005.PublishedbyBlackwellPublishing,9600GarsingtonRoad, OxfordOX42DQ,UKand350MainStreet,Malden,MA02148,USA LETTERPOSITIONANDDYSLEXIA 275 area was preferentially activated by letter strings (as opposed to symbols or faces) at about150mspost-stimulus(Tarkiainen,Cornelissen&Salmelin,2002;Tarkiainenetal., 1999), while 80% of the dyslexic subjects did not show this activation (Helenius et al., 1999).Thusitseemsthatnormalreadershavelearnedsometypeofstring-specificvisual processing, while most abnormal readers have not. Whatisthecauseofthisdifferenceinvisualprocessing?Whatcanthisdifferencetellus abouttherootcause(s)ofdyslexia?Tounderstandwhatgoeswronginvisualprocessingin dyslexics, it is necessary to understand the nature of this processing in normal readers. What is the source of the early string-specific LH activation? What form do normal orthographiccodestake?Oneapproachtoansweringthesequestionsisviacomputational modelling of how the early, retinotopic representation of a letter string is progressively converted into an abstract encoding of letter order. A recent model of letter-position encoding,dubbedtheSERIOLmodel(Whitney,2001;Whitney&Berndt,1999),fitsthis bill.The SERIOL model was formulated via considerationof neurobiological constraints and behavioural data, and has yielded new insights into hemifield asymmetries in visual wordrecognition(Whitney,2004a,2004b;Whitney&Lavidor,2004,2005forthcoming). The purpose of this article is to provide an account of abnormal visual/orthographic processing in dyslexics, based on the SERIOL model. We consider how auditory and/or visualdeficitscouldleadtoafailuretolearnnormalstring-specificvisualprocessing,and howsuchafailurecouldcontributetodifficultiesinlearninggrapheme-phonemecorres- pondences. This discussion is a first step in applying the SERIOL model to dyslexia; it is admittedly quite sketchy and highly speculative. It is hoped that such consideration of the visual aspect of visual word recognition may inspire fruitful avenues of research in dyslexia. This article is organised as follows. In the next section, the SERIOL model is presented. Then experimental evidence is presented supporting some key proposals of the model relevant to the issues at hand. These preliminary sections review previously completed work. We then extend the model to include a phonological route, and to ad- dressreadingacquisition.Afterreviewingauditoryandvisualdeficitsindyslexia,weare then in a position to apply the model to dyslexia. We discuss how visual/orthographic processingmaybreakdownduringreadingacquisition,andhowthisprocessmayinteract with phonological encoding. Review of SERIOL model The SERIOL (Sequential Encoding Regulated by Inputs to Oscillations within Letter units)modelisatheoryofprocessingintheproficientreader.Itspecifieshowanabstract letter position encoding is generated from a retinotopic representation, and how that encodingactivatesthelexicallevel.NotethatSERIOLisnotafullmodelofvisualword recognition, as it does not address phonological processing. Rather, the model has focused on the orthographic route to the lexicon. However, this focus is not meant to imply a lesser role for phonological processing. A phonological route is assumed, but heretofore has not been elaborated. We will address the phonological route at a fairly high level later in this article. Wefirstgiveabriefoverviewoftheexistingmodel,andthendiscusstheprocessingin each layer inmoredepth. Themodel consistsoffivelayers: edge, feature, letter, bigram and word. Each layer is comprised of nodes, which represent neural assemblies. Within rUnitedKingdomLiteracyAssociation2005 276 WHITNEYandCORNELISSEN eachlayer,theactivationofaletteristakentobethetotalamountofactivityacrossnodes representing that letter. To illustrate the representations and activation patterns at each layer, we will consider the stimulus CART, for fixation between the A and R. The lowest layer of the model, the edge layer, corresponds to the early visual cortical areas, which are retinotopically mapped. At this level, there is an activation pattern arising from visual acuity. In our example, A and R would have the highest activation, and C and T would have lower activations (because there are more edge nodes representing A and R than C and T). At the feature layer, this acuity pattern is transformed intoanactivationgradient,dubbed the locational gradient, whereactivation decreases from left to right. That is, C would attain the highest activation, A the next highest,RthenextandTthelowestactivation.Attheletterlayer,thelocationalgradient interactswith oscillatory letter nodes toinduce aserial encodingof order.That is, the C node fires, then A, then R, then T. At the next layer, temporally ordered pairs of letters activatebigramnodes.Thatis,abigramnodeXYisactivatedwhenletterXfiresbefore letter Y. In our example, bigram nodes *C, CA, CR, AR, CT, AT, RT and T* (where * denotes a word boundary) become activated. The bigrams then activate lexical representations at the word layer. In our example, the word node CART would become more highly activated than any other word node. Thus the retinotopic representation is converted into an abstract, location-invariant encoding of letter order via the creation of a serial representation. This serial encoding activates lexical representations via bigram nodes, which encode relationships between letters.SeeFigure1foraschematicoftheletterthroughwordlayers.Thechoiceofthese representations and the transformations between representations are best illustrated in a top-down manner, as presented next. Bigram layer to word layer Firstweconsiderwhattypeoforthographicrepresentationcontactsthelexicallayer.The results of priming experiments in which the order of letters from the target word is manipulated in the prime (Humphreys, Evett & Quinlan, 1990; Peressotti & Grainger, 1999) place strong constraints on the nature of the highest pre-lexical representation (Grainger & Whitney, 2004). These experiments indicate that order is important, not absolutepositionwithin thestring.Forexample,GRDNprimesGARDENwhileGDRN does not, and the prime G_RD_N does not provide any more facilitation than GRDN. Thus,itisrelativepositionthatmatters.Theseresultscanbestbeaccommodatedbyunits encoding the information that letter X preceded letter Y, where X and Y are not necessarily contiguous in the string. The activation level of a bigram node is sensitive to the separation between the constituent letters. Activation decreases as the amount of time between the firing of the two letters increases. In our example, the RT node would have a higher activation than AT,whichwouldhaveahigheractivationthanCT.Bigramnodesconnecttowordnodes viastandardweightedconnections,wheretheweightsontheconnectionsintoeachword node are proportional to the bigram activation pattern for that word. Letter layer to bigram layer How are such bigram nodes activated? Priming studies of letter trigrams indicate that priming can occur across letter positions (Peressotti & Grainger, 1995). This indicates that a letter detector can be activated by the corresponding letter in any string position. rUnitedKingdomLiteracyAssociation2005 LETTERPOSITIONANDDYSLEXIA 277 CART WORD 1.0 1.0 1.0 0.7 0.7 BIGRAM CA AR RT CR AT AC Detect ordered pairs then LETTER A B C R T Z Sequential firing 0.8 1.0 0.6 0.4 GRADED INPUTS from Feature level e m Ti Figure1.Architectureoftheletter,bigramandwordlevelsoftheSERIOLmodel,withexampleofencoding thewordCART. Notes: At the letter level, simultaneous graded inputs are converted into serial firing, as indicated by the timingoffiringdisplayedundertheletternodes.Bigramnodesrecognisetemporallyorderedpairsofletters (connections shown for a single bigram). Bigram activations (shown above the nodes) decrease with increasing temporal separation of the constituent letters. Activation of word nodes is based on the con- ventionaldot-productmodel. Therefore, we assume that such position-independent letter nodes comprise the next lower layer of the model. Because these nodes only encode letter identity, positional information must be dynamically associated with such nodes. Two possibilities are that position is represented by activation level, or by firing order. Representation via activationlevelwouldrequireamonotonicallydecreasingactivationgradient(e.g.,inthe CARTexample,theletternodeCwouldhavethe highestactivation,Athenexthighest, R the next and T the lowest). However, such an activation pattern at the letter layer is inconsistentwiththewell-knownfinal-letteradvantage;thefinalletterisperceivedbetter than the internal letters, indicating a higher activation level than the internal letters. Therefore,inlinewithevidenceforleft-to-rightstringprocessing(Harcum&Nice,1975; Nice &Harcum,1976),letterorderis takentoberepresentedserially.Abigramnodeis activated when its constituent letters fire in the correct order. For example, bigram node CA is activated when C fires before A, but not vice versa. Feature layer to letter layer Howisthisserialfiringinducedattheletterlayer?Hopfield(1995)andLismanandIdiart (1995) have proposed related mechanisms for precisely controlling timing of firing, rUnitedKingdomLiteracyAssociation2005 278 WHITNEYandCORNELISSEN Base Oscillation Threshhold al 1 2 nti e ot P ell C Time Figure2.Interactionofinputlevelandtimingoffiringforacellundergoingasub-thresholdoscillationof excitability. Notes: When a relatively high level of input (top curving line) is added to the base oscillation, the cell crossesthresholdattime1(actionpotentialnotillustrated).Iflessinputwerereceived,thecellwouldcross thresholdlaterinthecycle,suchasattime2. in which nodes undergo synchronous, sub-threshold oscillations of excitability. For convenience, we designate the trough of this oscillatory cycle to be the ‘start’ of the cycle.Inputlevelthendetermineshowearlyinthecyclesuchanodecancrossthreshold andfire(seeFigure2).Nearthebeginningofthecycle,excitabilityislow,soonlyanode receiving a high level of input can cross threshold and fire. Excitability increases over time,allowingnodesreceivinglessandlessinputprogressivelytofire.Thusserialfiring attheletterlevelcanbeaccomplishedvialetternodesthatoscillateinsynchronyandtake inputintheformofanactivationgradient.Inourexample,theCnodewouldreceivethe mostinput,Athenext,RthenextandTtheleast,allowingCtofiretheearliest,Anext,R nextandfinallyT.Thusallnodesfirewithinasingleoscillatorycycle,whichistakento be on the time scale of 200ms (Lisman & Idiart, 1995). An activated letter node inhibits other letter nodes. In our example, once C starts firing, how then does A ever start firing? As a letter node continues to fire, its firing rateslows,reducinglateralinhibitiontotheothernodes.Thisallowsanewletternodeto start firing. When an active letter node receives lateral inhibition, it then becomes strongly inhibited, so that it will not refire for the remainder of the oscillatory cycle.1 Thus graded input levels and lateral inhibition create strictly sequential firing at the letter layer. This process also creates varying activation levels. The activation of a letter node dependsonboththerateanddurationoffiring.Undertheassumptionsthatahigherinput levelleadstofasterfiringandthatfiringdurationisfairlyconstantacrossletters,thereisa decreasingactivationgradientattheletterlevel.However,thenoderepresentingthefinal letter is not inhibited by a subsequent letter. It can continue to fire until the end (down- phase) of the oscillatory cycle.2 Therefore, the final letter could potentially fire longer thanthe otherletters,andreach ahigherlevel ofactivationthanthe internal letterseven though it receives less input. This is consistent with the well-known final-letter ad- vantage. As discussed below, this proposal also explains some counter-intuitive experi- mental results on letter perceptibility. Thus there must be a monotonically decreasing activation gradient across the next lower layer of the model, to provide input to the letter layer. Because this gradient rUnitedKingdomLiteracyAssociation2005 LETTERPOSITIONANDDYSLEXIA 279 decreasesfromlefttoright(i.e.byspatiallocation),theselower-levelunitsmustbetuned toretinallocation.Thus,aretinotopicrepresentationisconvertedintoaserialrepresenta- tion, creating a location-invariant encoding; location invariance is achieved by mapping space on to time. This location-invariant encoding is resumed to occur in the LH. Edge layer to feature layer Basedonthearchitectureofthevisualsystem,thereareseveralimportantcharacteristics of the edge layer that determine the nature of the transformations from the edge to the feature layer. The fibres from each retina divide, such that information reaching V1 is split by visual field, not by eye. The left visual field (LVF) projects to the right hemisphere(RH),whiletherightvisualfield(RVF)projectstothelefthemisphere(LH). Availableevidenceindicatesthatthereislittleornooverlapinthecorticalrepresentation ofthevisualfieldsalongtheverticalmeridian(Brysbaert,1994;Lavidor&Walsh,2004; Leff, 2004). That is, letters immediately to the left of fixation are only projected to the RH, while letters immediately to the right of fixation are only projected to the LH. It is wellknownthatthenumberofcorticalcellsrepresentingafixedareaofspacedecreases as distance from fixation increases. Thus, activation decreases as eccentricity increases, giving a different activation pattern at the edge level from at the feature level (where activation decreases from left to right). Therefore, the acuity pattern must be converted to the locational gradient as the edge layeractivatesthefeaturelayer.Notethatforafixatedword,theacuitypatternacrossthe letters in the RVF/LH is the same as the locational gradient (i.e. decreasing from left to right). Thus the acuity gradient can serve as the locational gradient for those letters. However, in the LVF/RH, the acuity gradient increases from left to right; its slope is in the opposite direction as required for the locational gradient. Therefore, when the edge level activates the feature level, the acuity gradient must be inverted in the LVF/RH, while it can be maintained for the RVF/LH. Next we consider the details of this processing. Obviously, activation levels cannot be increased by increasing the number of cells representing a letter. Rather, the locational gradient is created via modification of firing rates. In the LVF/RH, the acuity gradient is inverted via a combination of strong excitation and direction-specific lateral inhibition. This process is displayed in Figure 3. We propose that letter features in the LVF/RH become more highly activated by edge- layerinputsthanthoseintheRVF/LH.Thisallowsthefirstlettertoreachahighlevelof activation. This could occur either via higher bottom-up connection weights from the edge layer, or by stronger self-excitatory connections within the feature layer. Within the RH feature layer, we propose that there is strong left-to-right lateral inhibition. That is, a feature node inhibits nodes to its right. As a result, letter features corresponding to the first letter receive no lateral inhibition, and inhibition increases as letter position increases. (Actually, there would be a gradient within each letter, but for simplicity we consider the mean activation level of a letter’s features.) Thus, the features comprising the first letter attain the highest activation level (as a result of strong excitation and lack oflateralinhibition),andactivationdecreasestowardsfixation(duetosharplyincreasing lateral inhibition). IntheRVF/LH,theacuitygradientservesasthelocationalgradient.Overallexcitation is weaker than to the LVF/RH. Left-to-right inhibition is not necessary, although some weak such inhibition may steepen the slope of the gradient. The two hemispheric rUnitedKingdomLiteracyAssociation2005 280 WHITNEYandCORNELISSEN S A C A S T L T E L E fixation LVF/RH RVF/LH Figure3. Formation of the locational gradient at the feature layer, for the centrally fixated stimulus CASTLE. Notes:Thehorizontalaxisrepresentsretinallocation,whiletheverticalaxisrepresentsactivationlevel.The bold-face letters represent bottom-up input levels, which are higher in the RH than the LH. In each hemisphere,activationdecreasesaseccentricityincreases,duetotheacuitygradient.Theitalicisedletters representtheeffectleft-to-rightinhibitionwithintheRH,andRH-to-LHinhibitionintheLH.IntheRH,C inhibitsA,andCandAinhibitionS,creatingadecreasinggradient.TheRHinhibitseachletterintheLHby thesameamount,bringingtheactivationofTlowerthanthatofS.Asaresult,activation monotonically decreasesfromlefttoright. gradientsare‘spliced’togetherviacross-hemisphericinhibition.TheRHfeaturesinhibit theLHfeatures,bringingtheactivationoftheLHfeatureslowerthantheactivationofthe least activated RH features. As a result, an activation gradient that is strictly decreasing from left to right is created. Summary The following are the important assumptions about processing at each layer. Edge layer (cid:1) Retinotopic; (cid:1) representation of central vision split across hemispheres; (cid:1) activation levels based on acuity gradient. Feature layer (for a left-to-right language) (cid:1) Retinotopic; (cid:1) representation still split across hemispheres; (cid:1) activation decreases from left to right (locational gradient); (cid:1) locational gradient formed by hemisphere-specific processing: (cid:1) stronger excitation to RH than LH; (cid:1) strong left-to-right lateral inhibition within RH; (cid:1) RH inhibits LH. Letter layer (cid:1) Position-independent letter nodes, located in LH; (cid:1) letter nodes undergo synchronous, sub-threshold oscillations in excitability; rUnitedKingdomLiteracyAssociation2005 LETTERPOSITIONANDDYSLEXIA 281 (cid:1) lateral inhibition between letter nodes; (cid:1) interactionofoscillations,lateralinhibitionandlocational-gradientinputgiveaserial firing; (cid:1) letter-node activation level depends on: (cid:1) firing rate – determined by input level (from locational gradient); (cid:1) firingduration–determinedbywhennextletterstartstofire,whichisdetermined by the input level to that letter. Bigram layer (cid:1) Bigram XY activated when letter X fires and then letter Y fires; (cid:1) activation of bigram XY decreases with the amount of time between the firing of letter X and letter Y. Word layer (cid:1) Receives weighted connections from bigram layer. Evidence for SERIOL model Having discussed the details of and motivations for the model, we now review some experimental support for the particulars of the model, based on novel experiments and new analyses of previous studies. Letter perceptibility It is well known that the external letters of a string (first and last letters) are the best perceived under central fixation. However, for unilateral presentation of short strings (threeorfourletters)atlargeeccentricities,acounter-intuitivepatternarises.IntheLVF/ RH,thefirstletteristhebestperceivedofalltheletters;intheRVF/LH,thelastletteris the bestperceived(Bouma,1973; Estes,Allemeyer &Reder, 1976; Legge, Mansfield & Chung,2001).Thus,ineachvisualfield,theletterfarthestfromfixation(whereacuityis thelowest)isthemostlikelytobecorrectlyreported.Thispatternispresentevenatlong exposure durations. For example, see Figure 4. As discussed above, the induction of the serial encoding leads to differing activations at the letter level. These activation patterns depend on the interaction of the locational gradient and the oscillatory cycle. Such dynamics explain these patterns of letter per- ceptibility, as follows. Fora centrally fixated string,the initial-letter advantage and final-letter advantage arise fordifferentreasons.Theinitialletterhasanadvantagebecauseitreceivesthehighestlevel ofbottom-upinput,allowingittofirethefastest.Itreceivesthemostinputbecauseitisnot inhibited from the left at the feature level. The final letter has an advantage because it is not inhibited by a subsequent letter during the induction of serial firing. That is, it is not inhibitedfromtherightattheletterlevel.Thus,likeothers,wealsoattributetheadvantage fortheexternalletterstoalackoflateralinhibition.However,thisreducedlateralinhibition doesnotarisefromalackofmaskingataverylowlevel(asisgenerallyassumed).Rather it arises from string-specific processing, consistent with the finding that non-letter, non- digit symbols do not display an outer symbol advantage. That is, the outer symbol is the least well perceived, as would be expected on the basis of acuity (Mason, 1982). rUnitedKingdomLiteracyAssociation2005 282 WHITNEYandCORNELISSEN 100 80 ct e orr 60 c nt e c 40 er P 20 0 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 Retinal Location Figure4. Experimental results from Estes and colleagues (1974) for four-letter strings, occurring at two differentretinallocationsineachvisualfield. Notes:Exposuredurationwas2400ms.Subjectsweretrainedtomaintaincentralfixation,andtheirgazewas monitored. Thisanalysisoftheexternalletteradvantageimpliesthatitshouldbepossibletoaffect differentially the initial- and final-letter advantages. The initial-letter advantage should disappear if the amount of bottom-up input to the initial letter node is not significantly higher than to the other letters. The final-letter advantage should disappear if the final letter node starts firing late in the oscillatory cycle, and so is unable to fire for a longer time than the other letters. As we shall see, these proposals explain the counter-intuitive perceptibility patterns for lateralised presentation of short strings. First however, a more in-depth consideration of activation patterns at the feature level is required. Recall that locational gradient formation requires different processing across the hemispheres.IntheRVF/LH,theacuitygradientservesasthelocationalgradient.Inthe LVF/RH,the acuitygradientis invertedvia strongbottom-upexcitationandleft-to-right lateral inhibition. Because the locational gradient is formed by different mechanisms in eachhemisphere,theshapeoftheresultinggradientmayvarywithhemisphere,especially at large eccentricities. The acuity gradient is known to fall off less quickly as distance from fixation increases. That is, the slope of the acuity gradient is steepest near fixation, andbecomesshalloweraseccentricityincreases.BecausetheRVF/LHlocationalgradient isbasedontheacuitygradient,thisimpliesthattheRVF/LHlocationalgradientbecomes more shallow as eccentricity increases. (See right half of Figure 5.) In the LVF/RH, formation of the locational gradient depends on left-to-right lateral inhibition. For strings at large eccentricities, inhibition may be too strong at early string positions, because of their relatively low level of activation (but, as discussed below, inhibitionmaybecometooweakatlaterstringpositions,becauseoftheincreasingacuity). (See left half of Figure 5.)Thusthe prediction is that the locational gradient shouldvary withvisualfield.TheproposalofasteeperLVFlocationalgradient(acrossstringpositions 1 to 4) explains an observed interaction between string position, eccentricity and visual field (Wolford & Hollingsworth, 1974), as discussed in Whitney (2001). Now we are in a position to explain the unilateral perceptibility patterns. In particular, wewillconsidertheresultsofEstes,AllemeyerandReder(1976),giveninFigure4.Inthe rUnitedKingdomLiteracyAssociation2005 LETTERPOSITIONANDDYSLEXIA 283 C A R C T A R CART T Figure5.SchematicoflocationalgradientsforthestimulusCARTatthreedifferentpresentationlocations. Notes: The vertical axis represents activation, while the horizontal axis represents retinal location. For centralpresentation,thegradientissmoothlyandrapidlydecreasing.ForRVFpresentation,thegradientis shallowerbecausetheacuitygradientisshallower.ForLVFpresentation,theinitialletterstronglyinhibits nearbyletters,butthegradientflattensoutasacuityincreases. following,primacywillsignifythataletterisperceivedbetterthanallotherletters,whereas advantage will mean that an external letter is perceived better than the internal letters. First we consider LVF presentation. At the feature layer, there is strong left-right inhibition,causingasteeplocationalgradient.Therefore,attheletterlevel,thefirstletter can fire for a (relatively) long time, as it is not quickly cut off by the next letter. Thus thereisaninitial-letterprimacy.Asaresultofitslowinputlevel,thefiringofthefinallet- terispushedlateintotheoscillatorycycle.Therefore,itcannotfirelongerthantheother letters, andnofinal-letter advantage emerges.Thisexplainstheperceptibilitypatternfor locations (cid:2)8 to (cid:2)5. ForRVFpresentation,thereisweakleft-to-rightinhibition,whiletheacuity/locational gradient is quite shallow. Therefore the activation of the second letter’s features is quite closetothatofthefirstletter.Asaresult,attheletterlevel,thefiringofthefirstletteris rapidly cut off by the second letter, giving no initial-letter advantage. Each successive letter quickly inhibits the preceding letter, allowing the final letter to start firing early in the oscillatory cycle. Therefore the final letter can fire longer than the other letters, creating afinal-letterprimacy. This explainsthe perceptibilitypatternsforlocations5to 8. The proposed activation patterns for both visual fields are displayed in Figure 6. Thisaccountexplainstheinitial/finaldifferencewithinasingleretinallocation(at (cid:2)5 and 5 in Figure 4). In the LVF/RH, the left-to-right, feature-level inhibition creates a disadvantageforafinalletter,whereasaninitialletterdoesnotreceivethisinhibition.In the RVF/LH, the shallow locational gradient creates a disadvantage for an initial letter because its firing at the letter layer is rapidly inhibited by the second letter. For a final letter, firing at the letter layercan continue until the end of the oscillatorycycleinstead. In contrast to the patterns at the larger eccentricity, the perceptibility function is U- shaped for both (cid:2)5 to (cid:2)2 and 2 to 5. As a result of higher acuity, bottom-up input is higher overall. In the LVF/RH, this allows the final letter to start firing earlier in the cycle,creatingafinal-letteradvantage.Alongwiththeusualinitial-letter advantage,this givestheU-shapedpattern.IntheRVF/LH,theacuity/locationalgradientissteeperthan for the larger eccentricity, so the difference in input to the first and second letters is larger, creating an initial-letter advantage and giving an overall U-shape. Next we consider the implications of this account for differing exposure durations. Under the assumption that a longer exposure duration increases the overall level of bottom-up input, the above analysis suggests that the RVF final-letter primacy and the LVF initial-letter primacy should be differentially affected by variations in durations. In rUnitedKingdomLiteracyAssociation2005