enhanced attentional gain as a mechanism for cortexsubmitted 9 may 2014; accepted in ﬁnal form 6...

doi:10.1152/jn.00353.2014 112:1217-1227, 2014. First published 11 June 2014;J NeurophysiolAnna Byers and John T. Serences

cortexgeneralized perceptual learning in human visual Enhanced attentional gain as a mechanism for

You might find this additional info useful...

99 articles, 33 of which can be accessed free at:This article cites /content/112/5/1217.full.html#ref-list-1

including high resolution figures, can be found at:Updated information and services /content/112/5/1217.full.html

can be found at:Journal of Neurophysiologyabout Additional material and information http://www.the-aps.org/publications/jn

This information is current as of September 2, 2014.

American Physiological Society. ISSN: 0022-3077, ESSN: 1522-1598. Visit our website at http://www.the-aps.org/.(monthly) by the American Physiological Society, 9650 Rockville Pike, Bethesda MD 20814-3991. Copyright © 2014 by the

publishes original articles on the function of the nervous system. It is published 12 times a yearJournal of Neurophysiology

on Septem

ber 2, 2014D

ownloaded from

on Septem

ber 2, 2014D

ownloaded from

http://www.the-aps.org/publications/jn

Enhanced attentional gain as a mechanism for generalized perceptual learningin human visual cortex

Anna Byers1 and John T. Serences1,2

1Department of Psychology, University of California, San Diego, California; and 2Neurosciences Graduate Program,University of California, San Diego, California

Submitted 9 May 2014; accepted in final form 6 June 2014

Byers A, Serences JT. Enhanced attentional gain as a mechanismfor generalized perceptual learning in human visual cortex. J Neuro-physiol 112: 1217–1227, 2014. First published June 11, 2014;doi:10.1152/jn.00353.2014.—Learning to better discriminate a spe-cific visual feature (i.e., a specific orientation in a specific region ofspace) has been associated with plasticity in early visual areas (sen-sory modulation) and with improvements in the transmission ofsensory information from early visual areas to downstream sensori-motor and decision regions (enhanced readout). However, in manyreal-world scenarios that require perceptual expertise, observers needto efficiently process numerous exemplars from a broad stimulus classas opposed to just a single stimulus feature. Some previous datasuggest that perceptual learning leads to highly specific neural mod-ulations that support the discrimination of specific trained features.However, the extent to which perceptual learning acts to improve thediscriminability of a broad class of stimuli via the modulation ofsensory responses in human visual cortex remains largely unknown.Here, we used functional MRI and a multivariate analysis method toreconstruct orientation-selective response profiles based on activationpatterns in the early visual cortex before and after subjects learned todiscriminate small offsets in a set of grating stimuli that were renderedin one of nine possible orientations. Behavioral performance im-proved across 10 training sessions, and there was a training-relatedincrease in the amplitude of orientation-selective response profiles inV1, V2, and V3 when orientation was task relevant compared withwhen it was task irrelevant. These results suggest that generalizedperceptual learning can lead to modified responses in the early visualcortex in a manner that is suitable for supporting improved discrim-inability of stimuli drawn from a large set of exemplars.

fMRI; visual cortex; perceptual learning; attention

PERCEPTUAL LEARNING (PL) REFERS to an improved ability to detector discriminate a sensory stimulus after repeated exposure(Gibson 1963, 1969). This behavioral improvement is thoughtto be supported by at least two neural mechanisms: an increasein the fidelity of early sensory responses (sensory modulationaccount) and/or an increase in the efficiency with which earlysensory information is relayed to, or “read out” by, down-stream decision mechanisms (enhanced readout account).Consistent with the sensory modulation account, learning for aspecific feature can enhance the gain of informative sensorysignals (Zohary et al. 1994; Gilbert et al. 2001; Schoups et al.2001; Furmanski et al. 2004; Yang and Maunsell 2004; Bao etal. 2010; Jehee et al. 2012) and reduce correlated noise sharedbetween visually responsive neurons (Adab and Vogels 2011;Bejjanki et al. 2011; Gu et al. 2011). Conversely, readoutmodels can explain improved perceptual performance without

the need to invoke learning-related changes in the fidelity ofresponses in early sensory areas (Dosher and Lu 1999, 2007,2009; Lu and Dosher 1999; Petrov et al. 2005; Lu et al. 2010,2011; Huang et al. 2012) and empirical data suggest thattraining can modify the degree to which informative sensoryneurons are weighted during decisionmaking, particularlywhen a large set of training stimuli are employed [i.e., thetrained stimuli change from session to session (Law and Gold2008, 2009, 2010; Gold et al. 2010)].

However, most prior investigations that documentedchanges in early sensory modulation trained subjects to dis-criminate a single feature value (often in a single location), sothe degree to which sensory modulation can support PL acrossmultiple stimulus exemplars is largely unknown. This is animportant question, as many real-world instances of PL requireidentifying unique members from a broad class of possiblefeatures: a radiologist must find a small fracture irrespective ofits precise location or its exact orientation, and a baggagescreener at an airport must detect potential weapons of allshapes and sizes hidden in a piece of luggage. Intuitively, thismore general form of PL might be supported primarily bychanges in the readout of unmodified sensory signals (e.g.,Law and Gold 2008), as any plasticity in sensory areas thatfacilitates the processing of one specific exemplar might inter-fere with the processing of other exemplars (Schoups et al.2001; Fahle 2004, 2009). On the other hand, it is possible thatlearning modulates early sensory responses in a more flexiblemanner that facilitates the processing of all stimuli from thetrained set of exemplars, perhaps by enhancing the overall gainof feature-selective responses in early visual areas. Here, wetested this latter hypothesis in human subjects using functionalmagnetic resonance imaging (fMRI) and a multivariate analy-sis technique that assesses how PL modifies the characteristicsof population-level orientation-selective tuning profiles in hu-man visual cortex.

METHODS

Subjects. Thirteen subjects (7 males and 6 females) between theages of 18 and 28 (M � 21.5, SD � 2.8) were recruited from theUniversity of California, San Diego to participate in the experiment,and all gave written informed consent in accordance with the Institu-tional Review Board at University of California, San Diego. Of these13 subjects, 5 were excluded from analysis: 1 did not complete bothscanning sessions, 2 could not do the task for unknown reasons (withfinal day orientation offsets of 24 and 40°, which were �3–10 timeslarger than the other subjects), and 2 showed significant motionartifacts (i.e., multiple abrupt translations of 5 mm or more in bothscanning sessions). Subjects performed a brief 20-min familiarizationsession before the initial scanning session, a 2-h pretraining fMRI

Address for reprint requests and other correspondence: J. Serences, Dept. ofPsychology, Neuroscience Graduate Program, 9500 Gilman Dr., La Jolla, CA,92093-0109 (e-mail: [email protected]).

J Neurophysiol 112: 1217–1227, 2014.First published June 11, 2014; doi:10.1152/jn.00353.2014.

12170022-3077/14 Copyright © 2014 the American Physiological Societywww.jn.org

on Septem

ber 2, 2014D

ownloaded from

mailto:[email protected]

scan, ten 1-h behavioral training sessions in the laboratory, and a final2-h posttraining fMRI scan. Subjects were compensated $10/h for thetraining sessions and $20/h for the scan sessions.

Stimulus and task. All stimuli were created using MATLAB (v.7.8;The Math Works, Natick, MA) with the Psychophysics Toolbox (v.3;Brainard 1997; Pelli 1997). The physical properties of the stimulusdisplay were identical across two attention conditions: 1) attend to theorientation of a grating stimulus (attend-orientation) and 2) attend toa central rapid serial visual presentation (RSVP) stream of letters(attend-RSVP). Each trial started with a 200-ms fixation interval,followed by a peripheral circular square-wave grating and a centrallypresented RSVP stream (Fig. 1). The grating had a radius of a 10°visual angle and a spatial frequency of 0.5 cycles per degree andflickered at a rate of 5 Hz. The inner blank aperture around fixationextended to a 2° visual angle. The orientation of the grating waspseudorandomly selected on each trial from a set of nine possibleorientations that were evenly distributed across 180° of orientationspace; �3° of orientation jitter was randomly added to the first gratingon each trial to ensure that subjects could not make their judgmentwith respect to a fixed orientation. The spatial phase of each of thegratings was also randomly determined on each trial to ensure thatmeasured signals were related to orientation and not to specificluminance patterns on the display screen. After 2,000 ms, the gratingwas removed from the screen for 400 ms, and then a second flickeringgrating was presented for an additional 2,000 ms. On attend-orienta-tion blocks, the subject’s task was to press one button if the secondstimulus was rotated clockwise from the first stimulus, and anotherbutton if it was rotated counterclockwise from the first stimulus(buttons were pressed with the first 2 fingers of the right hand). TheRSVP stream appeared within the inner blank aperture of the grating,with each letter subtending at a 1.2° visual angle. The letter streamwas presented during the same temporal epochs in which the orientedgratings were presented (i.e., 2,000 ms on, 400 ms off, 2,000 ms on).

Either an “X” or a “Y” was presented within each epoch of RSVP, andon attend-RSVP blocks the subject’s task was to press one button ifthe target letters matched (i.e., 2 “Xs” or 2 “Ys”) and another buttonif the two target letters did not match. Subjects had 2,000 ms torespond after the stimuli were removed from the screen, and thisresponse window also served as the intertrial interval (ITI).

At the beginning of each training session, the clockwise or coun-terclockwise rotational offset between the first and the second gratingswas set to 30°, and performance was adjusted using the QUESTalgorithm (Watson and Pelli 1983). The orientation offset required toachieve 75% accuracy was taken as the threshold for each session, andlearning was operationalized as the change in the angular offsetbetween the first and second gratings that was required to achievethreshold accuracy during each testing session. The exposure durationof each letter in the RSVP stream was manually adjusted to achieve�75% accuracy to ensure that the task remained equally challengingacross training sessions.

Training regimen. Before the initial scan session, subjects per-formed 4 blocks of 36 trials of the task (2 blocks each of theattend-orientation and attend-RSVP conditions) to familiarize themwith the task. After the initial scan session, subjects completed ten 1-htraining sessions. During each training session, subjects performedfour blocks of the attend-orientation condition and four blocks of theattend-RSVP condition, with each condition alternating on a block-by-block basis. Auditory feedback (short beeps, 100-ms duration) wasprovided to indicate correct and incorrect responses during the famil-iarization and training phases but not during either scan session. Afterthe 10th training session, subjects completed the second scan session.

fMRI sessions. Subjects performed eight blocks of the behavioraltask in each of the two scan sessions, and each block lasted 5 min.Eight null trials were randomly interleaved with 36 task trials in eachblock. For the first scan session, the offset between the two presen-tations of the grating was based on the staircased offset from thefamiliarization session (M � 8.3°, SE � 4.7°) and was manuallyadjusted at the end of each block based on subject performance tomaintain consistent performance across runs (M � 8.1°, SE � 0.95°).The letter exposure duration was also based on the exposure durationfrom the familiarization session (M � 157.1 ms, SE � 6.1 ms), butthis number was decreased slightly in the scanner (M � 140.7 ms, SE �9.5 ms), due to relatively high accuracy on this task during thepretraining session. For the second scan session, the orientation offsetwas based on the final training session (M � 2.9°, SE � 0.64°) andagain was manually adjusted at the end of each block to maintainconsistent performance across runs (M � 5.1°, SE � 0.33°). The letterexposure duration was also based on the last training session (M �33.0 ms, SE � 0.0 ms), and this number was also manually adjustedto maintain consistent performance (M � 85.4 ms, SE � 5.0 ms).

Functional localizer scans. Two 6.6-min functional localizer scanswere run in each scan session to independently identify voxels inretinotopically organized early visual areas that responded to stimuliin the spatial location occupied by the visual stimulus in the main task.On each trial, a contrast-reversing checkerboard stimulus (spatial fre-quency � 1 cycle/°, temporal frequency � 8 cycles/°) was presented inthe center of the screen for 10 s. On half of the 16 trials, the radius of thestimulus was a 5° visual angle and on the other half the radius of thestimulus was a 10° visual angle (matching the size of the stimulus usedin the main learning experiment). Subjects were instructed to press abutton whenever the contrast of the flickering stimulus dimmedslightly. Contrast-dimming target events occurred four times per trialand each dimming lasted for 66.7 ms. The occurrence of each targetwas determined pseudorandomly, with the constraint that the targetappeared at least 1 s after stimulus onset and 1 s before stimulusoffset, and each target was separated from the previous one by at least1 s. Trials were separated by a 10-s ITI.

Retinotopic mapping. A standard meridian mapping procedure wasused to identify visual areas V1, V2v/V2d, and V3v/V3d. Subjects passivelyviewed a contrast-reversing checkerboard stimulus flickering at 8 Hz and

ISI (400ms)

2nd Interval (2000ms)

1st Interval (2000ms)

Fig. 1. Schematic of the experimental paradigm. In both task types [attend-orientation/attend- rapid serial visual presentation (RSVP)], subjects viewed a2-s presentation of a annular square-wave grating with a rapid serial visualpresentation (RSVP) letter stream in the center, followed by a 400-ms blankinterstimulus interval (ISI), and then another 2-s exposure to a grating and acentral RSVP stream. While visible, the gratings flickered at 5 Hz to drive astrong visual response (and the phase of the grating’s spatial frequency wasrandomly selected on each flicker). On attend-orientation blocks, subjectsindicated whether the second grating was rotated clockwise or counterclock-wise from the first grating. On attend-RSVP blocks, subjects monitored theRSVP stream for a target letter (either “X” or “Y”) and indicated whether thetarget letter in the second stream matched (i.e., 2 “Xs” or 2 “Ys”) ormismatched (i.e., an “X” followed by a “Y” or vice versa) the target letter inthe first stream. Subjects responded via a manual button press with either theirindex finger (to indicate a counterclockwise rotation of the gratings or a matchbetween target letters) or their middle finger (to indicate a clockwise, rotationof the gratings or a mismatch between target letters).

1218 SENSORY GAIN AND GENERALIZED PERCEPTUAL LEARNING

J Neurophysiol • doi:10.1152/jn.00353.2014 • www.jn.org

on Septem

ber 2, 2014D

ownloaded from

covering 60° of polar angle and that alternated every 20 s between thehorizontal and vertical meridians (Engel et al. 1994; Sereno et al. 1995). Astandard general linear model (GLM) with regressors corresponding toepochs of vertical meridian and horizontal meridian stimulation was used toidentify voxels in the visual cortex that responded to each respective regionof the visual field. Each regressor was generated by convolving a boxcarmodel of each stimulus time series (“0” if no stimulus, “1” if stimuluspresent) with a standard difference-of-two gamma function model (time topeak of positive response: 5 s; time to peak of negative response: 15 s; ratioof positive and negative responses: 6; positive and negative response disper-sion: 1). The retinotopic mapping data were then projected onto a computa-tionally inflated rendering of each subjects’ gray/white matter boundary tofacilitate the identification of the horizontal and vertical meridians to defineeach cortical area. Because we used a large central stimulus, voxels in ventraland dorsal aspects of V2 and V3 were combined.

fMRI data acquisition and preprocessing. Scanning was performedusing a 3T GE MR750 MRI scanner with an eight-channel head coilat the University of California, San Diego’s Keck Center for Func-tional MRI. Anatomical images were collected using a T1-weightedsequence that produced images with a 1-mm3 resolution (TR/TE �11/3.3 ms, TI � 1,100 ms, 172 slices, flip angle � 18°). Functionalimages were collected using a gradient echo EPI pulse sequence with35 slightly oblique slices, which covered the whole brain. The sliceswere collected in ascending interleaved order and were 3-mm thick[TR � 2,000 ms, TE � 30 ms, flip angle � 90°, image matrix � 64(AP) � 64 (RL), with FOV � 192 mm (AP) � 192 mm (RL), voxelsize � 3 � 3 � 3 mm]. All EPI images were slice-time corrected,motion-corrected (both within and between scans), and high-passfiltered (3 cycles/run). BrainVoyager QX (v2.3; Brain Innovation,Maastricht, The Netherlands) was used to perform the data analysis inconjunction with custom analysis scripts written in MATLAB (ver-sion 7.11.0584; The Math Works).

Analysis of functional localizer data. Data from the functionallocalizer scans were analyzed using a GLM that contained separateregressors marking the temporal interval in which each of the twostimulus types was presented (small checkerboard and large checker-board). Each regressor was generated by convolving a boxcar model ofthe stimulus time series (“0” if no stimulus, “1” if stimulus) with astandard difference-of-two gamma function model (time to peak ofpositive response: 5 s; time to peak of negative response: 15 s; ratio ofpositive and negative responses: 6; positive and negative response dis-persion: 1). A contrast was then performed to identify voxels thatresponded more to the large 10° checkerboard compared with the small5° checkerboard; this contrast was intended to exclude voxels thatrespond strongly to the location of the central RSVP stream. Voxelswithin each of the visual areas were included in subsequent analyses ifthey passed a threshold of P � 0.05, after using BrainVoyager’s falsediscovery rate algorithm to correct for multiple comparisons.

Estimating trial-by-trial blood oxygen level-dependent responseson attention scans. Before estimating the trial-by-trial magnitude ofthe blood oxygen level-dependent (BOLD) response during the mainattention task, we first estimated the shape of the hemodynamicresponse function (HRF) separately for each visual area in eachsubject. This subject- and area-specific fitting was done because theshape of the HRF can vary substantially across subjects and brainregions (Zarahn et al. 1997; Gonzalez-Castillo et al. 2012). Data wereaveraged across all trials in the main scans, and the mean stimulus-locked HRF was computed separately for each visual area in eachsubject across a 24-s window using a Finite Impulse Response model(Dale 1999). We then used a gradient descent algorithm implementedin MATLAB to fit each HRF estimate using a standard difference-of-two gamma function model with time to peak, onset time, and theratio of positive and negative components as free parameters (usingthe same seed parameters listed in Analysis of functional localizerdata). The parameters that yielded the best fit were used to produce acustom HRF model for each visual area in each subject. We thencreated a design matrix with a boxcar model of the onset/offset time

of each 4.4-s stimulus epoch on each trial within a scan, yielding 36regressors of interest (4 instances of each of the 9 orientations) alongwith 1 additional regressor as a constant term. Each boxcar regressormodel was then convolved with the custom HRF estimate for thatsubject, and a GLM was used to estimate the relative amplitude of theBOLD response on each trial. This estimate of response amplitude oneeach trial in each voxel was then used as input to the orientationencoding model described below.

Estimating orientation-selective BOLD tuning profiles using aforward encoding model. Forward encoding models adopt a set of apriori assumptions about the important features or stimulus labels thatcan be distinguished using hemodynamic signals within a region ofinterest (ROI) (Brouwer and Heeger 2009, 2011; Dumoulin andWandell 2008; Gourtzelidis et al. 2005; Kay and Gallant 2009; Kay etal. 2008; Mitchell et al. 2008; Naselaris et al. 2009; Schönwiesner andZatorre 2009; Scolari, et al. 2012; Thirion et al. 2006; reviewed inNaselaris et al. 2011; Serences and Saproo 2012). The features orlabels in the model are then used to predict the pattern of BOLDresponses. Unlike multivoxel pattern analysis (MVPA) decodingtechniques, this approach allows us to quantify changes in the shapeof orientation population-response profiles that are related to PL. Thistechnique thus complements MVPA approaches, as MVPA providesan index of how much overall information is encoded in a pattern ofresponses and how that level of information changes as a result of PL(or other experimental manipulations: see Serences and Saproo 2012).However, a general increase in the information content of a responsepattern might be supported by several different types of modulation (i.e.,a decrease in the bandwidth of orientation-selective population tuningprofiles, an increase in gain of orientation-selective population tuningprofiles, etc.). Thus, by directly reconstructing population tuning profilesusing an a priori encoding model, we can quantify the modulatorypatterns that occurred as a result of generalized PL.

Here, we used a forward model adapted from Brouwer and Heeger(2009, 2011) to estimate the response across a series of orientationchannels as a function of PL and the focus of attention. The modelassumes that the BOLD response in a given voxel reflects the pooledactivity across a large population of orientation-selective neurons andthat the distribution of neural tuning preference is biased within agiven voxel either due to large-scale feature maps (Freeman et al.2011) or to subvoxel anisotropies in cortical columns (Kamitani andTong 2005; Swisher et al. 2010). In either case, BOLD responsesmeasured from many voxels in early visual cortex exhibit a modestbut robust orientation preference (Haynes and Rees 2005; Kamitaniand Tong 2005; Serences et al. 2009; Freeman et al. 2011).

For each subject and scan session, we first split the data into twosets (training and test sets), using a hold-one-out cross-validationmethod. Adopting the terminology and formulations of Brouwer andHeeger (2009, 2011) for consistency, let m be the number of voxels ina given visual area, n1 be the number of observations (trials) in thetraining set, n2 be the number of trials in the test set, and k be thenumber of hypothetical orientation channels. Let B1 (m � n1 matrix)be the training set, and B2 (m � n2 matrix) be the test set. The trainingdata in B1 were then mapped onto the matrix of hypothetical channeloutputs (C1, k � n) by the weight matrix (W, m � k) that wasestimated using a linear model of the form:

B1 � WC1 (1)

where the ordinary least-squares estimate of W is computed as:

W � B1C1T(C1C1

T)�1. (2)

The channel responses (C2, k � n2) were then estimated for the testdata (B2) using the weights estimated in Eq. 2:

C2 � �WTW��1WTB2. (3)

Equations 1 and 2 are similar to a traditional univariate GLM in thateach voxel gets a weight for each feature in the model (in this case, 1

1219SENSORY GAIN AND GENERALIZED PERCEPTUAL LEARNING


on Septem

ber 2, 2014D

ownloaded from

weight for each orientation channel). Equation 3 then estimateschannel responses on each trial based on the estimated weightsassigned to each voxel and the vector of responses observed across allvoxels on a given trial in the test set.

The construction of the basis set matrix C has a large impact on theresulting channel response estimates. In the present experiment, weused a half-sinusoidal function that was raised to the 5th power toemulate the approximate shape of single-unit tuning functions in V1,where the average 1/sqrt(2) half-bandwidth of orientation tuned cellsis �20° (Schiller et al. 1976; Ringach et al. 2002; Gur et al. 2005).This function was then replicated six times, with the six copies evenlydistributed across orientation space. The power of the sinusoidalfunction necessitated the use of at least six copies to adequately coverorientation space (Freeman and Adelson 1991). While we selected theshape of the basis function based on existing physiology studies, allresults that we report are robust to reasonable variations in this value(i.e., raising the half-sinusoid to the 5th, 6th, 7th, or 8th power, all ofwhich are reasonable choices based the large range of single-unitbandwidths observed in early visual areas). The analysis describedabove was then repeated with the center position of each of the sixbasis functions shifted across all 180° of orientation space in 1°increments to generate channel tuning functions with 1° resolution(which ultimately produced a 180° tuning function). Note that becauseof overlap between adjacent basis functions, each point along these180-point tuning curves was not independent from neighboring points.However, this approach was adopted to maximize the smoothness ofthe orientation tuning functions, which in turn facilitates quantifyingthe amplitude, bandwidth, and baseline level using the model fittingapproach described below. Last, the 180-point channel responsefunction estimated on each trial was circularly shifted so that thechannel matching the orientation of the stimulus on that trial waspositioned in the center of the tuning curve, thereby aligning allchannel response profiles to a common stimulus-centered referenceframe (which is plotted as 0° on the x-axis by convention, see Figs.5–6). The channel responses for each subject were then averagedacross subjects within each scanning session so that group channelresponse functions for the different attention and training conditionscould be compared.

Quantifying channel responses by fitting a von Mises function.After the channel responses were recentered to a common referenceframe, data from each subject and each condition (attend-orientation/attend-RSVP, pretraining/posttraining) were independently fit with avon Mises function:

f�� ae��cos��x��1� � b (4)

with amplitude (a), mean (�), bandwidth (�) , and baseline (b) asindependent free parameters that reflect distinct attributes of thefunction. The amplitude was restricted to a range from 0 and 2, themean was restricted to a range from 60 and 120°, the bandwidth wasrestricted to a range between 0 and 8, and the baseline was restrictedto a range from �3 and 3. The von Mises function (Eq. 4) was thenfit to the data from each subject 150 times using randomly choseninitial seed values for each parameter on each iteration (across therange of allowable values) to ensure that the fitting algorithm did notsettle in a local minimum. The set of parameters for each subject thatyielded the lowest root mean squared error across the 150 iterationswere then used in subsequent analyses.

RESULTS

Behavior during training sessions outside of scanner. Thestaircasing procedure did not converge for one subject duringthe first of their 10 behavioral training sessions, rendering datafrom session 1 highly variable (that subject had an orientationoffset of 22.2°, where the mean across all other subjects was4.7°). Therefore, statistical comparisons of the behavioral datawere performed excluding this subject (although the size of thelearning effect only increases if this subject is included, andtheir offsets during sessions 2-10 were stable and showedlearning effects). The angular offset between the gratingsneeded to perform the task at criterion was steadily reducedacross training sessions [see Fig. 2A; one-way repeated-mea-sures ANOVA, F(9,54) � 2.18, P � 0.038], and there was asignificant decrease in the average orientation offset betweensession 1 (4.7°) and session 10 [3.0°; paired t-test, t(6) �3.4150, P � 0.014]. Response time did not significantly changeacross sessions 1-10 in the attend-orientation task [one-wayrepeated-measures ANOVA, F(9,54) � 0.48, P � 0.88]. Ac-curacy also did not significantly change across sessions 1-10[see Fig. 2B; one-way repeated-measures ANOVA, F(9,54) �0.8, P � 0.62], suggesting that staircasing was successful inmaintaining a criterion level of accuracy.

Observers also improved in the attend-RSVP condition, suchthat the exposure duration of each letter in the RSVP stream

65 65

CA

DB

**

*

Fig. 2. Behavioral performance for training sessionsoutside of the scanner. A: average orientation offsetin degrees between the 2 gratings during each of the10 behavioral training sessions (at a fixed accuracythreshold, see B). Note that for 1 subject, the stair-casing procedure did not converge, leading to anoverinflated estimate of their orientation offset onsession 1; hence, all statistical analyses were per-formed excluding this subject (and data from thatsubject are excluded from this plot as well, see maintext for more details). B: accuracy on the attend-orientation condition was maintained at a fixed rateacross all behavioral testing sessions. C: averageletter exposure duration in ms/letter across each ofthe 10 behavioral training sessions. D: accuracy onthe attend-RSVP condition decreased across train-ing sessions (see text for more details). All errorbars � 1SE. *P � 0.05, significance across ses-sions.



on Septem

ber 2, 2014D

ownloaded from

steadily decreased across the 10 training sessions [see Fig. 2C;one-way repeated-measures ANOVA, F(9,54) � 22.89, P �0.0002]. The average letter exposure duration also significantlydecreased from 128.57 ms/letter during session 1 to 33.33ms/letter during session 10 [paired t-test, t(6) � 8.40, P �0.0002]. However, accuracy also changed across the 10 ses-sions in the RSVP task [see Fig. 2D; one-way repeated-measures ANOVA, F(9,54) � 5.08, P � 0.0001], suggestingthat the reduction in exposure duration across sessions is not aperfect indicator of the amount of learning that occurred in theRSVP task (although this was not an issue when comparingpre- and posttraining behavioral performance in the scanner asall comparisons were carried out within session, see below).Response times for the attend-RSVP condition did not signif-icantly change across training sessions [one-way repeated-measures ANOVA, F(9,54) � 0.82, P � 0.6].

Behavior in the scanner. In the scanner, the average orien-tation offset was 8.1° (SE � 0.95°) in the pretraining scansession and 5.1° (SE � 0.33°) in the posttraining scan session[Fig. 3A; paired t-test, t(7) � 3.84, P � 0.006]. For the RSVPtask, the pretraining letter presentation speed was 140.63 ms/letter (SE � 9.5 ms/letter) and the posttraining letter presen-tation speed was 85.42 ms/letter [SE � 5.0 ms/letter; Fig. 3C;paired t-test, t(7) � 6.33, P � 0.0004]. We speculate that theslightly elevated orientation offsets and letter exposure dura-tions in the scanner, compared with outside the scanner, weredue to the lower quality visual display in the scanner. Accuracyon attend-orientation trials was slightly lower during the post-training scan session compared with the pretraining session[see Fig. 3B; orientation task: 91% pretraining SE � 1%, 88%posttraining SE � 1%, paired t-test, t(7) � 2.78, P � 0.03].However, this small change in accuracy was not a major causeof the observed learning effects, as four subjects had balancedaccuracy across pre- and posttraining scan sessions and they allshowed at least 4° improvement with training. Accuracy in theattend-RSVP condition did not significantly differ across scansessions {see Fig. 3D; letter task: 89% pretraining (SE � 2%)and 88% posttraining [SE � 2%, repeated-measures t-test,t(7) � 0.8769, P � 0.4]}.

fMRI results. First, we computed the mean univariate HRFfor each attention condition based on the z-scored BOLD timeseries averaged across all voxels in each ROI (see METHODS;Fig. 4). As all three areas showed a similar pattern of re-sponses, we performed statistics on the data averaged acrossV1, V2, and V3 (compare panels in Fig. 4). A three-wayrepeated-measures ANOVA with time (0–24 s), attention con-dition, and training (pre- vs. posttraining) revealed a maineffect of time [F(12,84) � 41.13, P � 0.0001], a main effectof attention condition [F(1,7) � 36.91, P � 0.0005], and aninteraction between time and attention [F(12,84) � 12.24, P �0.0001]. However, there was no main effect of training[F(1,7) � 0.11, P � 0.74] and no interaction between trainingand time [F(12,84) � 7.99, P � 0.65] or training and attentioncondition [F(1,7) � 2.91, P � 0.133]. Overall, this patternsuggests that the HRFs were significantly modulated by atten-tion but not by learning.

We then used an encoding model (see METHODS) to determinehow learning modulated orientation-selective response profilesin V1, V2, and V3. Across both scan sessions and attentionconditions, the averaged orientation-selective response profilespeaked at the target orientation and fell off in a Gaussian-likemanner (see Figs. 5 and 6). We fit the tuning curve for eachsubject with a von Misses function and submitted the bestfitting parameters to a three-way repeated-measures ANOVAwith factors for attention condition (attend-orientation vs. at-tend-RSVP), training (scan 1 vs. scan 2), and visual area (V1,V2, V3). There was no main effect of visual area and nothree-way interaction among visual area, attention condition,and training condition for any of the fit parameters (amplitude,bandwidth, baseline, or mean, see Table 1 for all statistics).This suggests that all areas showed a similar pattern of mod-ulation across task manipulations (see Figs. 6 and 7 for datafrom individual visual areas). As a result, we hereafter focusour statistical analysis on the von Mises parameters fit to tuningfunctions that were computed after combining all voxels fromV1-V3. Because the size of the orientation offset differedacross pre- and posttraining scan sessions, a comparison be-tween pre- and posttraining within each attention condition

CA

DB

* *

*

Fig. 3. Behavioral performance inside thescanner. A: average orientation offset (atthreshold) in degrees between the 2 gratingsduring the pretraining (dark gray) and post-training (light gray) scan sessions. B: accu-racy in the attend-orientation condition was�90% during both the pre- and posttrainingscan sessions, although accuracy in the post-training scan decreased slightly (see maintext). C: average letter exposure duration inms/letter during the pretraining (hatchedbars) and posttraining (open bars) scan ses-sions. D: accuracy on the attend-RSVP con-dition was maintained at �90% during boththe pre- and posttraining scan sessions. Allerror bars � 1SE. *P � 0.05 significance.



on Septem

ber 2, 2014D

ownloaded from

could not be directly assessed due to the sensory differences inthe display. Thus we focused on evaluating a two-way repeat-ed-measures ANOVA with attention condition and training asfactors to compare with the magnitude of attention effectsduring pre- and posttraining sessions as our main marker of PL.The ANOVA revealed a significant overall main-effect ofattention condition on the amplitude of the orientation-selec-tive tuning profiles, such that the amplitude in the attend-orientation condition was higher than the amplitude in theattend-RSVP condition [collapsed across scanning sessions;

Fig. 5, A and B; F(1,7) � 10.03, P � 0.015]. In addition, therewas a significant interaction between attention condition andtraining on tuning function amplitude, indicating that the mag-nitude of the attentional modulation (i.e., attend-orientation vs.attend-RSVP) increased with training [Fig. 5C; F(1,7) � 7.98,P � 0.025]. This interaction was driven by a significantdifference between response amplitudes in the attend-orienta-tion and attend-RSVP conditions posttraining but not pretrain-ing [t(7) � 3.5, P � 0.01 and t(7) � 1.25, P � 0.25,respectively]. Finally, there was an overall main effect of

0 4 8 12 16 20 24

0

0.25

0.5

0.75

1

BO

LD re

spon

se (z

)B

OLD

resp

onse

(z)

Time (s) Time (s)

RSVP-PreOrient-Pre

RSVP-PostOrient-Post

0 4 8 12 16 20 24

0

0.25

0.5

0.75

1V1-V3 Averaged V1

0 4 8 12 16 20 24

0

0.25

0.5

0.75

1

0 4 8 12 16 20 24

0

0.25

0.5

0.75

1 V3V2

*†× *†×

*†× *†×

Fig. 4. Mean hemodynamic response func-tions (HRFs) for the attend-orientation (blacklines) and attend-RSVP (blue lines) condi-tions across training conditions (solid lines:pretraining, dashed lines: posttraining). Topleft: HRFs averaged across voxels in all 3regions of interest. Top right and bottom leftand right: HRFs for each individual visualarea (V1, V2, V3, respectively). BOLD,blood oxygen level dependent. *Significanteffect of attention condition. †Significantmain effect of time. xSignificant attention bytime-course interaction. Note that learning didnot modulate the time-course or the magni-tude of attention effects. All error bars � 1SE.All P � 0.05.

BA

C D E* × *

Fig. 5. Orientation-selective tuning curves com-puted across V1-V3 and averaged across sub-jects, centered on the target orientation on agiven trial and resulting parameter estimatesfrom the best fitting von Mises function. A:pretraining tuning curves during the attend-ori-entation (gray) and attend-RSVP conditions(blue). B: posttraining tuning curves for thesame 2 conditions. C: mean amplitude of thebest-fitting von Mises function the orientationtuning curves in the attend-orientation (gray)and attend-RSVP conditions (blue) during thepretraining (left) and posttraining (right) scansessions. Gray asterisk indicates a significanteffect of attention condition. xSignificant inter-action between attention condition and trainingcondition. P � 0.05. D: mean bandwidth of thebest-fitting von Mises function in the attend-orientation (gray) and attend-RSVP (blue) tasksduring the pretraining (left) and posttraining(right) scan sessions. E: mean baseline of thebest-fitting von Mises function in the attend-orientation (gray) and attend-RSVP (blue) tasksduring the pretraining (left) and posttraining(right) scan sessions. Blue asterisk indicates asignificant effect of attention condition at P �0.05. All error bars are � 1SE.



on Septem

ber 2, 2014D

ownloaded from

attention condition on the baseline parameter, such that base-line estimates were higher in the attend-RSVP condition thanthe attend-orientation condition [Fig. 5E; F(1,7) � 7.84, P �0.026]. No learning-related changes in any of the other param-eters were observed.

DISCUSSION

Learning to better discriminate a specific stimulus featurehas been associated with enhanced sensory responses in earlyvisual cortex (Gilbert et al. 2001; Schoups et al. 2001; Fur-manski et al. 2004; Yang and Maunsell 2004; Jehee et al. 2012)and with the improved readout of sensory information bydownstream sensorimotor and decision mechanisms (Dosherand Lu 1998, 2009; Ghose et al. 2002; Law and Gold 2008,2009, 2010; Xiao et al. 2008; Zhang et al. 2010b, 2013; Goldet al. 2010; Lu et al. 2011; Huang et al. 2012). However, littleis known about how these mechanisms support PL in situationswhere observers must learn to make difficult discriminationsbetween stimuli that are drawn from a large set of exemplars.Indeed, modified early sensory responses that are highly selec-tive for supporting the discrimination of a specific featuremight be predicted to impair the discriminability of otherstimuli (e.g., Schoups et al. 2001; Fahle 2004, 2009). However,here we show that generalized PL leads to an enhancement offeature-selective attentional modulations in early visual cortex,supporting a version of the sensory modulation account ofgeneralized PL.

Even though our data support a sensory modulation accountof generalized PL, they do not speak to the possibility of

additional contributions due to changes in the efficiency ofsensory readout. Indeed, sensory modulation and changes inthe efficiency of readout are not mutually exclusive. Forexample, Law and Gold (2008) trained subjects using a largestimulus set and their data provide support for enhancedreadout in the absence of changes in early sensory responses inmotion selective area MT. Their task involved discriminatingstimuli moving in one of two directions, where the exactdirection and location of the motion stimulus were changed ona session-by-session basis to optimize the response of cells inMT and the lateral intraparietal area (LIP). Law and Gold didnot observe training-related changes in MT, an area thatrepresents the quality of motion-evoked sensory signals. In-stead, they found an increase in motion-driven responses inLIP, where cells are thought to play a role in the “readout” (orsensorimotor transformation) of sensory signals that are gen-erated in MT. However, there are several differences betweenour study and theirs that might contribute to the differences thatwe observe. First, we used fMRI to assess the response profileacross orientation channels tuned to all stimulus features (theentire range from 0 to 180°) whereas Law and Gold justassessed the activity of cells that were selectively tuned to thetrained stimulus feature (or to the antipreferred feature). Thuswe mapped out the full tuning profile and they mapped twopoints along this function. This could be a key difference, asour results suggest that the overall amplitude increase infMRI-based orientation tuning profiles is due to a combination

Fig. 6. Orientation-selective tuning curves for each individual visual area (V1,V2, V3) and averaged across subjects, centered on the orientation of thegrating stimulus presented on each trial. Left: pretraining tuning curves for V1,V2, and V3 for the attend-orientation (gray) and attend-RSVP (blue) condi-tions. Right: posttraining orientation-selective tuning curves for V1, V2, andV3 for the attend-orientation (gray) and attend-RSVP (blue) conditions. Allerror bars are � 1SE.

Table 1. Statistical comparisons for each fit parameter acrossconditions and visual areas

Comparison Error F Statistic P Value

AmplitudeVisual Area (2, 14) 2.587 0.1107Day (1, 7) 0.009 0.9273Attn Cond (1, 7) 10.025 0.0158Vis � Day (2, 14) 0.711 0.5082Vis � Attn (2, 14) 0.241 0.7888Day � Attn (1, 7) 7.984 0.0256Vis � Day � Attn (2, 14) 2.903 0.0882

BaselineVisual Area (2, 14) 3.246 0.0695Day (1, 7) 0.199 0.6687Attn Cond (1, 7) 7.84 0.0265Vis � Day (2, 14) 0.43 0.6591Vis � Attn (2, 14) 3.113 0.0761Day � Attn (1, 7) 4.121 0.0819Vis � Day � Attn (2, 14) 0.969 0.4037

BandwidthVisual Area (2, 14) 3.051 0.0795Day (1, 7) 1.058 0.3378Attn Cond (1, 7) 0.009 0.9263Vis � Day (2, 14) 0.17 0.8456Vis � Attn (2, 14) 1.044 0.378Day � Attn (1, 7) 3.988 0.086Vis � Day � Attn (2, 14) 1.816 0.199

MeanVisual Area (2, 14) 0.348 0.7121Day (1, 7) 0 0.9856Attn Cond (1, 7) 0.742 0.4175Vis � Day (2, 14) 0.338 0.7191Vis � Attn (2, 14) 1.554 0.2458Day � Attn (1, 7) 0.962 0.3593Vis � Day � Attn (2, 14) 1.591 0.2384

Significant effects shown in italics.



on Septem

ber 2, 2014D

ownloaded from

of enhanced responses in orientation channels tuned around therelevant stimulus feature and to attenuated responses in orien-tation channels tuned progressively farther away from therelevant stimulus (compare Fig. 5, A to B). Thus measuring theentire tuning profile might provide improved sensitivity todetect learning-related changes in the amplitude of sensoryresponses. In addition, and perhaps more importantly, ourdisplay contained multiple competing stimuli (a central RSVPstream and a peripheral grating), and competition betweenmultiple stimuli has been previously shown to amplify themagnitude of attentional modulations in visual cortex (e.g.,Moran and Desimone 1985; Motter 1993; Desimone and Dun-can 1995; Kastner et al. 1998; Awh et al. 2003; Reynolds andDesimone 2003; Serences et al. 2004; Gál et al. 2009). Thusthe degree of interstimulus competition might be a key factorin determining the magnitude of sensory modulation duringPL. In either case, we believe that the current evidence, bothfrom our study and from others, firmly suggests that PLenhances sensory modulation as well as enhanced readout, andfuture studies will need to directly determine how much eachmechanism contributes to behavioral indexes of PL, particu-larly as a function of the degree of competition betweenrelevant and irrelevant stimuli in the visual field.

The present observation of a training-related increase infeature-selective attention effects is consistent with severalprevious studies that focused on PL for highly specific visualfeatures (reviewed in Seitz and Dinse 2007; Roelfsema et al.2010; Sasaki et al. 2010; Byers and Serences 2012). In partic-ular, the present data are consistent with previous suggestionsthat top-down attentional modulations play a key role in bothcontext-dependent effects on sensory processing, PL, and at-tenuating responses evoked by task-irrelevant features (Gilbertet al. 2000; Crist et al. 2001; Schoups et al. 2001; Li et al. 2004;Vidnyánszky and Sohn 2005; Gál et al. 2009; Mukai et al.2011; Gilbert and Li 2012). For example, Gilbert and Li (2013)summarize evidence for top-down attentional control and re-current feedback connections as a mechanism for changingearly sensory responses based on task goals; a pattern that ispresent even in the pretraining orientation-selective response

profiles described here. However, our data suggest a furtherkey role of task context in gating posttraining sensory re-sponses, as enhanced responses to trained stimuli were primar-ily elicited when the orientation stimulus is task relevant andnot when the orientation stimulus is merely present (i.e., in theattend RSVP condition; Gilbert et al. 2000; Li et al. 2004;Gilbert and Li 2012). Furthermore, the nature of the neuralchanges associated with PL may be dependent on the contextin which subjects were trained, suggesting that training withmore irrelevant distractors could lead to even larger attentionalmodulations, both before and after learning (Luck et al. 1993;Motter 1993; Kastner and Ungerleider 2000; Li et al. 2004;Vidnyánszky and Sohn 2005; Gál et al. 2009; Gilbert and Li2012).

Recently, Jehee et al. (2012) used fMRI to show that theability of individual voxels in V1-V4 to discriminate betweentwo similar orientations was improved when the orientedstimulus was attended compared with when it was unattended.However, this improvement was specific to both the trainedorientation and stimulus location, so the implication of theirresults to situations in which training is focused on improvingthe discriminability of a larger class of stimuli is unclear. Incontrast, we addressed this complementary question by ex-ploiting the joint information encoded by large-scale popula-tions to show that generalized orientation learning improvesthe amplitude of feature-based modulations in early visualcortex. Thus the current results demonstrate that PL modifiespopulation-level orientation tuning profiles via enhanced atten-tional gain and that these population-level modulations providecomputational constraints on models of PL in more complexscenarios.

In the present study, we observed a qualitative attenuation ofresponses to the grating stimulus when subjects were perform-ing the attend-RSVP task (see blue line in Fig. 5B; althoughnote that we cannot compare pre- and posttraining responseamplitude levels quantitatively in the attend-RSVP conditiondue to learning-related changes in the nature of the sensorystimulus, see RESULTS). At first glance, this observation seemsinconsistent with prior work showing PL even for task-irrele-

Fig. 7. Parameter estimates from the best fitting vonMises function for each individual visual area (V1,V2, V3). Left: amplitude of best-fitting von Misesfunction for the attend-orientation (gray) and at-tend-RSVP (blue) tasks during the pretraining (left)and posttraining (right) scan sessions. Middle:bandwidth of the best-fitting von Mises function inthe attend-orientation (gray) and attend-RSVP(blue) tasks during the pretraining (left) and post-training (right) scan sessions. Right: baseline of thebest-fitting von Mises function in the attend-orien-tation (gray) and attend-RSVP (blue) tasks duringthe pretraining (left) and posttraining (right) scansessions. All error bars are � 1SE.



on Septem

ber 2, 2014D

ownloaded from

vant stimuli when they are paired with the presentation of aconcurrent relevant or rewarding stimulus (Seitz and Watanabe2005; Seitz et al. 2009; Frankó et al. 2010). However, task-irrelevant PL seems to occur only when the irrelevant stimulusdoes not strongly compete or interfere with the processing ofthe relevant attended stimulus (Watanabe et al. 2001; Seitz etal. 2006, 2009; Tsushima et al. 2008; Choi et al. 2009; Seitzand Watanabe 2009; Huang and Watanabe 2012; Leclercq andSeitz 2012a,b). This criterion is not likely met in our study, asthe grating was highly salient and it surrounded the relativelylow-salience central RSVP stream. Thus this stimulus config-uration likely led to a high degree of interstimulus competition,and in turn this competition may have blocked task-irrelevantPL effects on the oriented grating when subjects were attendingthe central RSVP stream.

In previous reports (e.g., Regan and Beverley 1985; Seungand Sompolinsky 1993; Schoups et al. 2001), including workfrom our own laboratory (Scolari and Serences 2009, 2010;Scolari et al. 2012), the performance of a fine discriminationtask between similar orientations has been shown to mediatethe gain of “off-channel” neural populations, or those neuronsthat are tuned just away from the to-be-discriminated feature.From an information-theoretic viewpoint, this is optimal be-cause these off-channel neurons have a steeper tuning functionaround the discriminada and should thus undergo a largerchange in firing rate to the two stimulus alternatives (seeScolari and Serences 2009 for more discussion). Thus wemight have expected such off-channel modulations in thepresent study to be enhanced with PL. However, in experiment1 of Scolari et al. (2012), we demonstrated that off-channelgain, at least as assessed using fMRI and the same analysisapproach and general paradigm used here, is not observedwhen the direction of the orientation offset is unknown inadvance (i.e., the subject does not know whether to attend toand expect a clockwise or a counterclockwise rotation). Inexperiment 2 in the same article, we then demonstrated a robustoff-channel gain effect when we precued subjects about therotational offset of the oriented stimuli in advance. Given thatthe present task closely mimics experiment 1 from Scolari et al.(2012) in which no precue was given, we did not expect to seeoff-channel gain in early visual areas in the present study andindeed no robust evidence for such an effect was observed.

Although we observed an increase in the overall amplitudeof orientation-selective responses in early visual areas (Fig.5C), we did not observe any training-related changes in themean BOLD response computed across all voxels in each ROI(Fig. 4). Several previous reports also failed to observe changesin the mean BOLD response with PL (Yotsumoto et al. 2008;Zhang et al. 2010a; Jehee et al. 2012). Note, however, that thepresence of enhanced feature-based attention effects (Figs. 5and 6) can still be observed even if there is no net change in themagnitude of the mean HRF that is computed based on theaverage across all voxels. For example, we observed enhancedresponses in channels tuned to the relevant stimulus andattenuated responses in channels tuned away from the relevantstimulus (Figs. 5B and 6). Thus increases and decreases in thechannel responses roughly cancel out and we would not predicta change in overall HRF amplitude. In addition, others havesuggested that learning leads to improved coding for specificfeatures even in the absence of a change in mean HRFamplitude, as PL might reduce response variability (which

would effectively increase the signal-to-noise ratio of the HRFwithout changing its overall magnitude; Jehee et al. 2012).Finally, it is possible that the amplitude of the mean BOLDresponse is modulated by PL but that early learning-inducedincreases in BOLD amplitude return to baseline with furthertraining (Yotsumoto et al. 2008). Thus we may not haveobserved changes in HRF amplitude because our posttrainingfMRI session took place after the HRF returned to pretrainingbaseline levels.

In summary, we used a forward encoding model to makeinferences about how PL influences the shape of population-level orientation tuning functions based on the multivariateresponse profiles in visual cortex (Serences and Saproo 2012).We show that PL leads to a selective increase in the amplitudeof orientation-selective tuning profiles in V1-V3 with training.This amplitude increase corresponds to an increase in thedynamic-range (or entropy) of orientation-selective responseprofiles within these regions. In turn, increasing the dynamicrange of feature-selective response profiles should, in theory,increase the information content of these population-responseprofiles and thus the separability of responses evoked bystimuli rendered at different orientations (i.e., an amplitudeincrease should magnify the difference between mean re-sponses evoked by two nearby orientations; Saproo and Ser-ences 2010, 2014).

ACKNOWLEDGMENTS

We thank Sirawaj Itthipuripat, Thomas C. Sprague, and Tiffany Ho foruseful discussions.

GRANTS

This work was supported by National Institute of Mental Health GrantR01-092345 (to J. T. Serences).

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the author(s).

AUTHOR CONTRIBUTIONS

Author contributions: A.B. and J.T.S. conception and design of research;A.B. performed experiments; A.B. analyzed data; A.B. and J.T.S. interpretedresults of experiments; A.B. prepared figures; A.B. drafted manuscript; A.B.and J.T.S. edited and revised manuscript; A.B. and J.T.S. approved finalversion of manuscript.

REFERENCES

Adab HZ, Vogels R. Practicing coarse orientation discrimination improvesorientation signals in macaque cortical area v4. Curr Biol 21: 1661–1666,2011.

Awh E, Matsukura M, Serences JT. Top-down control over biased compe-tition during covert spatial orienting. J Exp Psychol Hum Percept Perform29: 52–63, 2003.

Bao M, Yang L, Rios C, He B, Engel S. Perceptual learning increases thestrength of the earliest signals in visual cortex. J Neurosci 30: 15080–15084,2010.

Bejjanki VR, Beck JM, Lu Z, Pouget A. Perceptual learning as improvedprobabilistic inference in early sensory areas. Nat Neurosci 14: 642–648,2011.

Brainard DH. The Psychophysics Toolbox. Spat Vis 10: 433–436, 1997.Brouwer G, Heeger D. Decoding and reconstructing color from responses in

human visual cortex. J Neurosci 29: 13992–14003, 2009.Brouwer GJ, Heeger DJ. Cross-orientation suppression in human visual

cortex. J Neurophysiol 106: 2108–2119, 2011.



on Septem

ber 2, 2014D

ownloaded from

Byers A, Serences JT. Exploring the relationship between perceptual learningand top-down attentional control. Vision Res 74: 30–39, 2012.

Choi H, Seitz AR, Watanabe T. When attention interrupts learning: inhibi-tory effects of attention on TIPL. Vision Res 49: 2586–2590, 2009.

Crist R, Li W, Gilbert C. Learning to see: experience and attention in primaryvisual cortex. Nat Neurosci 4: 519–525, 2001.

Dale A. Optimal experimental design for event-related fMRI. Hum BrainMapp 8: 109–114, 1999.

Desimone R, Duncan J. Neural mechanisms of selective visual attention.Annu Rev Neurosci 18: 193–222, 1995.

Dosher BA, Lu ZL. Perceptual learning reflects external noise filtering andinternal noise reduction through channel reweighting. Proc Natl Acad SciUSA 95: 13988–13993, 1998.

Dosher BA, Lu ZL. Mechanisms of perceptual learning. Vision Res 39:3197–3221, 1999.

Dosher BA, Lu ZL. The functional form of performance improvements inperceptual learning: learning rates and transfer. Psychol Sci 18: 531–539,2007.

Dosher BA, Lu ZL. Hebbian reweighting on stable representations in percep-tual learning. Learn Percept 1: 37–58, 2009.

Dumoulin SO, Wandell BA. Population receptive field estimates in humanvisual cortex. Neuroimage 39: 647–660, 2008.

Engel AS, Rumelhart DE, Wandell BA, Glover GH, Chichilnisky E,Shadlen MN. fMRI of human visual cortex. Nature 369: 525, 1994.

Fahle M. Perceptual learning: a case for early selection. J Vis 4: 879–890,2004.

Fahle M. Perceptual learning and sensomotor flexibility: cortical plasticityunder attentional control? Philos Trans R Soc B Biol Sci 364: 313–319,2009.

Frankó E, Seitz AR, Vogels R. Dissociable neural effects of long-termstimulus-reward pairing in macaque visual cortex. J Cogn Neurosci 22:1425–1439, 2010.

Freeman J, Brouwer G, Heeger D, Merriam E. Orientation decodingdepends on maps, not columns. J Neurosci 31: 4792–4804, 2011.

Freeman W, Adelson E. The design and use of steerable filters. IEEE TransPattern Anal Mach Intell 13: 891–906, 1991.

Furmanski C, Schluppeck D, Engel S. Learning strengthens the response ofprimary visual cortex to simple patterns. Curr Biol 14: 573–578, 2004.

Gál V, Kozák LR, Kóbor I, Bankó EM, Serences JT, Vidnyánszky Z.Learning to filter out visual distractors. Eur J Neurosci 29: 1723–1731,2009.

Ghose GM, Yang T, Maunsell JH. Physiological correlates of perceptuallearning in monkey V1 and V2. J Neurophysiol 87: 1867–1888, 2002.

Gibson EJ. Perceptual learning. Annu Rev Psychol 14: 29–56, 1963.Gibson EJ. Principles of Perceptual Learning and Development. New York:

Appleton-Century-Crofts, 1969.Gilbert C, Ito M, Kapadia M, Westheimer G. Interactions between atten-

tion, context and learning in primary visual cortex. Vision Res 40: 1217–1226, 2000.

Gilbert C, Sigman M, Crist R. The neural basis of perceptual learning.Neuron 31: 681–697, 2001.

Gilbert CD, Li W. Adult visual cortical plasticity. Neuron 75: 250–264, 2012.Gold JI, Law CT, Connolly P, Bennur S. Relationships between the

threshold and slope of psychometric and neurometric functions duringperceptual learning: implications for neuronal pooling. J Neurophysiol 103:140–154, 2010.

Gonzalez-Castillo J, Saad ZS, Handwerker AD, Inati SJ, Brenowitz N,Bandettini P. Whole-brain, time-locked activation with simple tasks re-vealed using massive averaging and model-free analysis. Proc Natl Acad SciUSA 109: 5487–5492, 2012.

Gourtzelidis P, Tzagarakis C, Lewis SM, Crowe AD, Auerbach E, JerdeAT, Ugurbil K, Georgopoulos AP. Mental maze solving: directional fMRItuning and population coding in the superior parietal lobule. Exp Brain Res165: 273–282, 2005.

Gu Y, Liu S, Fetsch CR, Yang Y, Fok S, Sunkara A, DeAngelis GC,Angelaki DE. Perceptual learning reduces interneuronal correlations inmacaque visual cortex. Neuron 71: 750–761, 2011.

Gur M, Kagan I, Snodderly D. Orientation and direction selectivity ofneurons in V1 of alert monkeys: functional relationships and laminardistributions. Cereb Cortex 15: 1207–1221, 2005.

Haynes J, Rees G. Predicting the orientation of invisible stimuli from activityin human primary visual cortex. Nat Neurosci 8: 686–691, 2005.

Huang CB, Lu ZL, Dosher B. Co-learning analysis of two perceptual learningtasks with identical input stimuli supports the reweighting hypothesis.Vision Res 61: 25–32, 2012.

Huang T, Watanabe T. Task attention facilitates learning of task-irrelevantstimuli. PLoS One 7: e35946, 2012.

Jehee JF, Ling S, Swisher JD, van Bergen RS, Tong F. Perceptual learningselectively refines orientation representations in early visual cortex. J Neu-rosci 32: 16747–16753, 2012.

Kamitani Y, Tong F. Decoding the visual and subjective contents of thehuman brain. Nat Neurosci 8: 679–685, 2005.

Kastner S, De Weerd P, Desimone R, Ungerleider L. Mechanisms ofdirected attention in the human extrastriate cortex as revealed by functional.MRI Sci 282: 108, 1988.

Kastner Ungerleider S. Mechanisms of visual attention in the human visualcortex. Annu Rev Neurosci 23: 315–341, 2000.

Kay K, Naselaris T, Prenger R, Gallant J. Identifying natural images fromhuman brain activity. Nature 452: 352–355, 2008.

Kay KN, Gallant JL. I can see what you see. Nat Neurosci 12: 245, 2009.Law C, Gold J. Shared mechanisms of perceptual learning and decision

making. Top Cogn Sci 2: 226–238, 2010.Law CT, Gold JI. Neural correlates of perceptual learning in a sensory-motor,

but not a sensory, cortical area. Nat Neurosci 11: 505–513, 2008.Law CT, Gold JI. Reinforcement learning can account for associative and

perceptual learning on a visual-decision task. Nat Neurosci 12: 655–663,2009.

Leclercq V, Seitz AR. The impact of orienting attention in fast task-irrelevantperceptual learning. Atten Percept Psychophys 74: 648–660, 2012a.

Leclercq V, Seitz AR. Fast task-irrelevant perceptual learning is disrupted bysudden onset of central task elements. Vision Res 61: 70–76, 2012b.

Li W, Piëch V, Gilbert C. Perceptual learning and top-down influences inprimary visual cortex. Nat Neurosci 7: 651–657, 2004.

Lu Z, Hua T, Huang C, Zhou Y, Dosher BA. Visual perceptual learning.Neurobiol Learn Mem 95: 145–151, 2011.

Lu Z, Liu J, Dosher BA. Modeling mechanisms of perceptual learning withaugmented Hebbian re-weighting. Vision Res 50: 375–390, 2010.

Lu ZL, Dosher B. Characterizing human perceptual inefficiencies with equiv-alent internal noise. J Opt Soc Am A Opt Image Sci Vis 16: 764–778, 1999.

Luck SJ, Fan S, Hillyard SA. Attention-related modulation of sensory-evoked brain activity in a visual search task. J Cogn Neurosci 5: 188–195,1993.

Mitchell TM, Shinkareva SV, Carlson A, Chang KM, Malave VL, MasonAR, Just MA. Predicting human brain activity associated with the meaningsof nouns. Science 320: 1191–1195, 2008.

Moran J, Desimone R. Selective attention gates visual processing in theextrastriate cortex. Science 229: 782–784, 1985.

Motter BC. Focal attention produces spatially selective processing in visualcortical areas V1, V2, and V4 in the presence of competing stimuli. JNeurophysiol 70: 909–919, 1993.

Mukai I, Bahadur K, Kesavabhotla K, Ungerleider LG. Exogenous andendogenous attention during perceptual learning differentially affect post-training target thresholds. J Vis 11: 1–15, 2011.

Naselaris T, Kay KN, Nishimoto S, Gallant JL. Encoding and decoding infMRI. Neuroimage 56: 400–410, 2011.

Naselaris T, Prenger RJ, Kay KN, Oliver M, Gallant JL. Bayesianreconstruction of natural images from human brain activity. Neuron 63:902–915, 2009.

Pelli D. The VideoToolbox software for visual psychophysics: transformingnumbers into movies. Spat Vis 10: 437–442, 1997.

Petrov AA, Dosher B, Lu ZL. The dynamics of perceptual learning: anincremental reweighting model. Psychol Rev 112: 715–743, 2005.

Regan D, Beverley KI. Postadaptation orientation discrimination. J Opt SocAm A 2: 147–155, 1985.

Reynolds JH, Desimone R. Interacting roles of attention and visual saliencein V4. Neuron 37: 853–863, 2003.

Ringach D, Shapley R, Hawken M. Orientation selectivity in macaque V1:diversity and laminar dependence. J Neurosci 22: 5639–5651, 2002.

Roelfsema PR, van Ooyen A, Watanabe T. Perceptual learning rules basedon reinforcers and attention. Trends Cogn Sci 14: 64–71, 2010.

Saproo S, Serences J. Spatial attention improves the quality of populationcodes in human visual cortex. J Neurophysiol 104: 885–895, 2010.

Saproo S, Serences JT. Attention improves transfer of motion informationbetween V1 and MT. J Neurosci 34: 3586–3596, 2014.

Sasaki Y, Nanez JE, Watanabe T. Advances in visual perceptual learningand plasticity. Nat Rev Neurosci 11: 53–60, 2010.



on Septem

ber 2, 2014D

ownloaded from

Schiller PH, Finlay BL, Volman SF. Quantitative studies of single-cellproperties in monkey striate cortex. III. Spatial frequency. J Neurophysiol39: 1334–1351, 1976.

Schönwiesner M, Zatorre RJ. Spectro-temporal modulation transfer functionof single voxels in the human auditory cortex measured with high-resolutionfMRI. Proc Natl Acad Sci USA 106: 14611–14616, 2009.

Schoups A, Vogels R, Qian N, Orban G. Practising orientation identificationimproves orientation coding in V1 neurons. Nature 412: 549–553, 2001.

Scolari M, Byers A, Serences JT. Optimal deployment of attentional gainduring fine discriminations. J Neurosci 32: 7723–7733, 2012.

Scolari M, Serences J. Basing perceptual decisions on the most informativesensory neurons. J Neurophysiol 104: 2266–2273, 2010.

Scolari M, Serences JT. Adaptive allocation of attentional gain. J Neurosci29: 11933–11942, 2009.

Seitz A, Watanabe T. A unified model for perceptual learning. Trends CognSci 9: 329–334, 2005.

Seitz A, Watanabe T. The phenomenon of task-irrelevant perceptual learning.Vision Res 49: 2604–2610, 2009.

Seitz AR, Dinse HR. A common framework for perceptual learning. CurrOpin Neurobiol 17: 148–153, 2007.

Seitz AR, Kim D, Watanabe T. Rewards evoke learning of unconsciouslyprocessed visual stimuli in adult humans. Neuron 61: 700–707, 2009.

Seitz AR, Nanez JE, Holloway S, Tsushima Y, Watanabe T. Two casesrequiring external reinforcement in perceptual learning. J Vis 6: 966–973,2006.

Serences JT, Saproo S. Computational advances towards linking BOLD andbehavior. Neuropsychologia 50: 435–446, 2012.

Serences JT, Saproo S, Scolari M, Ho T, Muftuler LT. Estimating theinfluence of attention on population codes in human visual cortex usingvoxel-based tuning functions. Neuroimage 44: 223–231, 2009.

Serences JT, Yantis S, Culberson A, Awh E. Preparatory activity in visualcortex indexes distractor suppression during covert spatial orienting. JNeurophysiol 92: 3538–3545, 2004.

Sereno MI, Dale AM, Reppas JB, Kwong KK, Belliveau JW, Brady TJ,Rosen BR, Tootell RB. Borders of multiple visual areas in humans revealedby functional magnetic resonance imaging. Science 268: 889–893, 1995.

Seung HS, Sompolinsky H. Simple models for reading neuronal populationcodes. Proc Natl Acad Sci USA 90: 10749–10753, 1993.

Swisher J, Gatenby J, Gore J, Wolfe B, Moon C, Kim S, Tong F. Multiscalepattern analysis of orientation-selective activity in the primary visual cortex.J Neurosci 30: 325–330, 2010.

Thirion B, Duchesnay E, Hubbard E, Dubois J, Poline JB, Lebihan D,Dehaene S. Inverse retinotopy: inferring the visual content of images frombrain activation patterns. Neuroimage 33: 1104–1116, 2006.

Tsushima Y, Seitz A, Watanabe T. Task-irrelevant learning occurs onlywhen the irrelevant feature is weak. Curr Biol 18: R516–R517, 2008.

Vidnyánszky Z, Sohn W. Learning to suppress task-irrelevant visual stimuliwith attention. Vision Res 45: 677–685, 2005.

Watanabe T, Náñez JE, Sasaki Y. Perceptual learning without perception.Nature 413: 844–848, 2001.

Watson A, Pelli D. QUEST: A Bayesian adaptive psychometric method.Percept Psychophys 33: 113–120, 1983.

Xiao LQ, Zhang JY, Wang R, Klein AS, Levi DM, Yu C. Complete transferof perceptual learning across retinal locations enabled by double training.Curr Biol 18: 1922–1926, 2008.

Yang T, Maunsell J. The effect of perceptual learning on neuronal responsesin monkey visual area V4. J Neurosci 24: 1617–1626, 2004.

Yotsumoto Y, Watanabe T, Sasaki Y. Different dynamics of performanceand brain activation in the time course of perceptual learning. Neuron 57:827–833, 2007.

Zarahn E, Aguirre G, D’esposito M. A trial-based experimental design forfMRI. Neuroimage 6: 122–138, 1997.

Zhang G, Cong L, Song Y, Yu C. ERP P1-N1 changes associated withVernier perceptual learning and its location specificity and transfer. J Vis 13:1–13, 2013.

Zhang J, Meeson A, Welchman AE, Kourtzi Z. Learning alters the tuning offunctional magnetic resonance imaging patterns for visual forms. J Neurosci30: 14127–14133, 2010a.

Zhang T, Xiao LQ, Klein AS, Levi DM, Yu C. Decoupling locationspecificity from perceptual learning of orientation discrimination. Vision Res50: 368–374, 2010b.

Zohary E, Celebrini S, Britten KH, Newsome W. Neuronal plasticity thatunderlies improvement in perceptual performance. Science 263: 1289–1292,1994.



on Septem

ber 2, 2014D

ownloaded from

enhanced attentional gain as a mechanism for cortexsubmitted 9 may 2014; accepted in ﬁnal form 6...

Documents