Report. Practicing Coarse Orientation Discrimination Improves Orientation Signals in Macaque Cortical Area V4

Transcription

1 Current Biology 21, , October 11, 2011 ª2011 Elsevier Ltd All rights reserved DOI /j.cub Practicing Coarse Orientation Discrimination Improves Orientation Signals in Macaque Cortical Area V4 Report Hamed Zivari Adab 1 and Rufin Vogels 1, * 1 Laboratorium voor Neuro-en-Psychofysiologie, Katholieke Universiteit Leuven Medical School, Campus Gasthuisberg, O&N2, Herestraat 49, Bus 1021, 3000 Leuven, Belgium Summary Practice improves the performance in visual tasks, but mechanisms underlying this adult brain plasticity are unclear. Single-cell studies reported no [1], weak [2], or moderate [3, 4] perceptual learning-related changes in macaque visual areas V1 and V4, whereas none were found in middle temporal (MT) [5]. These conflicting results and modeling of human (e.g., [6, 7]) and monkey data [8] suggested that changes in the readout of visual cortical signals underlie perceptual learning, rather than changes in these signals. In the V4 learning studies, monkeys discriminated small differences in orientation, whereas in the MT study, the animals discriminated opponent motion directions. Analogous to the latter study, we trained monkeys to discriminate static orthogonal orientations masked by noise. V4 neurons showed robust increases in their capacity to discriminate the trained orientations during the course of the training. This effect was observed during discrimination and passive fixation but specifically for the trained orientations. The improvement in neural discrimination was due to decreased response variability and an increase of the difference between the mean responses for the two trained orientations. These findings demonstrate that perceptual learning in a coarse discrimination task indeed can change the response properties of a cortical sensory area. Results and Discussion Two monkeys (Macaca mulatta) performed a coarse orientation discrimination task: they had to discriminate two orthogonal, oblique sinusoidal gratings that were masked by noise [9, 10] (Figure 1A). Both the noise and spatial phase varied across trials. In humans, practicing this task produces a location and orientation-specific improvement in performance [11]. Initially, the monkeys were trained with gratings having a signal-to-noise ratio (SNR) of 80% to establish an association between each orientation and a saccadic response. Once performance reached >90% correct, we presented the two gratings with SNRs between 0% and 40%. The behavioral performance improved over the course of several months of practice (Figures 1B and 1C). This improvement was significant (p < 0.01; permutation test) for all nonzero trained SNRs for each monkey, as demonstrated by robust linear regressions of percent correct versus session number. Throughout the training, we recorded single neurons in V4, an area arguably at the same hierarchical level as middle temporal (MT) [12] that shows orientation sensitivity [13] and perceptual learning-related changes in a fine orientation *Correspondence: rufin.vogels@med.kuleuven.be discrimination task [3, 4]. We recorded from a region (4 6mm 2 ) of right dorsal V4 with receptive fields (RF) at small eccentricities. All neurons (n = 177) that responded to at least one stimulus (analysis of variance [ANOVA]; p < 0.05) of the coarse orientation task were included. Based on RF mapping of each neuron, the stimulus was centered on the RF. Therefore, the stimulus position varied slightly between sessions. During the course of the recordings, the mean (6 standard deviation) stimulus eccentricity was 3.1 (61.0 ; n = 105) and 3.0 (60.7 ; n = 72) in monkeys M and P, respectively, and the mean polar angle was 225 (622 ) and 221 (612 ) in monkeys M and P, respectively. We quantified how well each neuron could discriminate between the trained orientations at each SNR using receiver operating characteristic (ROC) analysis [14, 15] on the spiking responses. The area under the ROC curve (AUROC) corresponds to the percent correct classification of the orientation by an ideal observer using the distributions of the number of spikes (see Figure S1 available online). Robust linear regression of the AUROC versus cell number showed significant positive slopes (p < 0.01; permutation test) at SNRs above 15% and 10% in monkeys M and P, respectively (Figures 1D and 1E). This increase in AUROC with training remained significant using partial regression with stimulus eccentricity and polar angle as covariates at SNRs of 20% (partial r = 0.34; p < 0.005; n = 105) and 40% (partial r = 0.25; p < 0.05; n = 105) in monkey M and of 15% (partial r = 0.42; p < ; n = 72) and 40% (partial r = 0.26; p < 0.05; n = 72) in monkey P. To pool the data from both animals, we created an early (E) and late (L) group of neurons consisting of the first 30% and last 30% recorded neurons of each monkey (monkey M: n = 35 neurons per group; P: n = 24). Figure 2A shows the mean psychometric functions obtained during the recordings of the two groups. The percent correct responses were significantly larger in the L compared to the E group (ANOVA; main effect of group; p < ), demonstrating the behavioral learning effect. This group effect was significant in each animal (p < ; Figure S2). The simultaneously obtained neurometric curves (Figure 2B) showed significantly larger AUROC values in the L compared to the E group (ANOVA; main effect of group; p < ). This was significant in each animal (p < 0.01; Figure S2). Although for some neurons neurometric and psychometric performances were equal (Figure S1), the average neuron performed more poorly than the monkey itself, even after the long training (L group average 75% correct neurometric and psychometric thresholds were 24% and 12% SNR, respectively). Note that other studies [15, 16] have also reported average neurometric-psychometric threshold ratios considerably greater than 1 for short stimulus durations. The increase in AUROC with training can result from an increase in the difference between the mean responses to the two trained orientations and/or a decrease of the trial-to-trial response variability. For each neuron, we computed the difference (BT-WT) between the mean net response to the best trained orientation (BT), i.e., the one producing the largest mean net response at 40% SNR, and the mean net response for the worst trained orientation (WT). The mean BT-WT was significantly larger in the L compared to the E group

2 Current Biology Vol 21 No Figure 1. Psychometric and Neurometric Performance while Learning the Coarse Orientation Discrimination Task (A) Task. After 500 ms of fixating upon a red target (size: 0.26 ) on a noise background, a grating patch was superimposed on the noise background for 250 ms during continuous fixation. After stimulus offset, the monkey was required to fixate for 200 ms, after which two targets (size: 0.57 ) were presented at 5 eccentricity. Each of the two orientations (left and right oblique gratings) was associated with one target point. The monkey had to saccade to one of the two target points to indicate his decision (green arrow). Correct choices were rewarded with a small liquid reward. Choices for the 0% signal-to-noise ratio (SNR) stimuli were rewarded with a 0.5 probability. Fixation breaks (square fixation window width of <1.5 ) during the trial sequence or saccades outside the target windows were treated as aborts and discarded for the computation of psychometric and neurometric performances. The panel illustrates a grating patch with 40% SNR. The fixation and peripheral targets are depicted larger than as presented in the experiment for illustration purposes. (B and C) Behavioral performance (percent correct; chance level 50%) in the coarse orientation discrimination task as a function of the number of training sessions (days) for each of the SNRs (0% 40%) is indicated by colored symbols. Filled symbols indicate sessions during which single cells were recorded. (B) shows monkey M and (C) shows monkey P. (D and E) Neurometric performance (area under the receiver operating characteristic [ROC] curve [AUROC]; chance level 50%) of single V4 neurons (see Figure S1 for an example neuron) plotted as a function of cell number recorded during training (filled symbols in B and C) in the coarse orientation discrimination task for different SNRs. Color codes the same as in (B) and (C). Lines are based on robust regression, and slopes significantly larger than 0 (p < 0.01) are indicated by a star in the legend. (D) shows monkey M and (E) shows monkey P. (ANOVA; main effect of group; p < 0.01; Figure 2C), without significant interaction between animal and group (Figure S2). To address the second factor, we computed the Fano factor for each neuron and stimulus with a mean response >0 (Figure 2D). The mean Fano factor, averaged across BT and WT, was lower in the L than in the E group (ANOVA; main effect of group; p < 0.05). This effect on response variance was present in each animal (Figure S2). Regression analyses of log response variance on log mean response (across neurons and orientations) showed a significantly larger slope in the E

3 Learning Coarse Discrimination in Monkey V Figure 2. Comparison of Early and Late Learning Period in Coarse Orientation Discrimination, Averaged across Animals (A) Behavioral performance as a function of SNR. The mean percent of correct responses was significantly larger in the late (L) than in the early (E) group (F[1,114] = 148.3; p < ). (B) Neurometric performance (AUROC) as a function of SNR. AUROC values were significantly larger in the L than in the E period (F[1,114] = 17.6; p < ). (C) Mean difference between the net response to the best trained (BT) and worst trained (WT) orientations as a function of SNR. Best and worst trained orientations were defined for each neuron based on the mean net response at 40% SNR. The mean BT-WT increased with SNR (F[5,570] = 56.9; p < ) and was significantly larger in the L group (F[1,114] = 6.6; p < 0.01). (D) Fano factor as a function of SNR. The Fano factors of the BT and WT orientations were averaged. The Fano factor depended significantly on SNR (F[5,550] = 7.7; p < ), with lower Fano factors for the 40% SNR compared to the lower SNRs. The mean Fano factor was lower in the L than in the E group (F[1,110] = 5.0; p < 0.05). The main effect of E versus L group was also significant when the Fano factors for BT and WT were analyzed separately. (E) Mean net response to the BT orientations as a function of SNR. The difference between E and L was not significant (ns). (F) Mean net response to the WT orientations as a function of SNR. Learning significantly reduced the WT response (F[1,114] = 4.7; p < 0.05). Error bars indicate standard error of the mean (SEM). compared to the L group at each of the SNRs >5% (general linear model; p < 0.01; Figure S2), indicating that the Fano factor decrease reflected a genuine decrease in response variance for SNRs >5%. Regression analyses combining the data of both animals by normalizing cell number within each animal (Supplemental Experimental Procedures; Table S1) confirmed that learning increased differences between the responses to the trained orientations and decreased trial-to-trial variability. A simulation (Supplemental Experimental Procedures; Table S1) showed that the mean change in response variance and BT-WT is sufficient to explain the mean AUROC increase with learning. The baseline activity tended to be reduced by training, but this decrease was not significant (ANOVA; main effect of group; p = 0.38). The same held for the net response to BT (Figure 2E; ANOVA; main effect of group; p = 0.17). Learning affected the response to WT (ANOVA; p < 0.05; Figure 2F; Figure S2), but this effect failed to reach significance using robust regression of WT response versus normalized cell number. Thus, learning did not increase the overall response strength, but instead tended to reduce it. This is similar to what has been observed in area V1 [1, 2] but opposite to the increased responses following fine orientation learning in V4 [3, 4]. The orientation sensitivities of most neurons (n = 133) were assessed by presenting eight randomly interleaved orientations (SNR 80%; range: ) while the monkey was passively fixating. Fine discrimination learning decreases the proportion of V4 neurons preferring the trained orientation [3, 4]. The coarse orientation discrimination learning did not significantly affect the proportion of neurons preferring one of the trained orientations (Figure S3) nor did it affect the response strength to the preferred orientation (Figure S3). Unlike for fine discrimination learning [3, 4], we observed no consistent effect of training on the orientation sensitivity index (SI) (Figure S3): no significant main effect of E (n = 50) versus L (n = 34) period but a significant interaction between monkey and group (p < 0.05; Figure 3C) with a significant increase of SI with training in monkey P only. The absence of any consistent effect of training on SI was not a result of the animal not performing the orientation discrimination task, because during passive fixation we observed a significant increase in AUROC for the trained orientations with an 80% SNR (main effect of period [p < 0.01] [Figure 3A] with no interaction between monkey and group [Figure S3]). As with the orientation discrimination task, the increase in AUROC during passive fixation resulted from an increase in BT-WT (p < 0.05) and a decrease in Fano factor (p < 0.05). To determine whether the training effect was orientation specific, we computed the AUROC for the two untrained orientations differing by 45 from the trained orientations. No effect of training on AUROC was observed for these untrained orientations (Figure 3B; main effect of E versus L period: not significant [ns]; interaction group and monkey: ns). An ANOVA with factors group, monkey and trained versus untrained orientations showed a significant interaction between the trained/ untrained and the E/L variables for the AUROC and Fano factor (p < 0.05), demonstrating the orientation specificity of the training effect directly. Figure 3D shows the mean AUROC as a function of the minimum absolute angle (TO-PO; Figure S3) between the trained orientations (TO) and preferred orientation (PO) of each neuron. To increase power, we pooled the data of the 15% and 20% SNRs. As expected [16], the mean AUROC decreases with increasing TO-PO angle (ANOVA; main effect of TO-PO; p < 0.01). The mean AUROC was greater for the L than for the E groups at each of the TO-PO angles (Figure 3D; main effect of group: p < ). These effects were present in each animal and for the AUROC at the 80% SNR trained orientations during the fixation task (both effects p < 0.05). The

4 Current Biology Vol 21 No Figure 3. Effect of Learning on Properties of V4 Neurons (A) Neurometric performance (AUROC) during passive fixation for the two trained orientations presented with 80% SNR, averaged across the two animals. E and L: early (n = 50) and late (n = 34) period. Learning increased neurometric performance for the trained orientations during passive fixation (F[1,80] = 7.5; p < 0.01). (B) Neurometric performance (AUROC) during passive fixation for two untrained orientations presented with 80% SNR, averaged across the two animals. These stimuli were presented interleaved with the trained orientations, the data for which are shown in (A). No increase in AUROC with learning was present (F[1,80] = 0.0; ns). (C) Mean orientation sensitivity index (SI) in E and L periods for monkeys M and P separately. The larger the SI is, then the more orientation-sensitive the neuron. The main effect of E versus L periods was ns (F[1,80] = 1.5; ns) but the interaction between monkey and group was (F[1,80] = 6.5; p < 0.05). Data were obtained during passive fixation (same neurons as in A and B). (D) Neurometric performance (AUROC) as a function of the minimum absolute angle (TO-PO) between the trained orientations (TO) and preferred orientation (PO) of each neuron, averaged for the E and L groups of both monkeys. When, for example, the trained orientations were 22.5 and and the preferred orientation was 45, then the TO-PO angle is The data of both 15% and 20% SNR were analyzed in a single analysis of variance with factors TO-PO, E versus L group, SNR, and animal. The mean AUROC decreased with increasing TO-PO angle (F[2,63] = 4.9; p < 0.01). The mean AUROC was larger for the L than for the E group (F[1,63] = 14.1; p < 0.001). Error bars indicate SEM. interaction between TO-PO angle and E-L group was not significant in these ANOVAs. To assess with greater power whether the effect of learning on AUROC differed among TO-PO angles, we performed robust linear regression on the AUROC as a function of normalized cell number separately for the three groups of neurons defined by their TO-PO angle (group 1: TO-PO = 0 ; group 2: TO-PO = 22.5 ; group 3: TO-PO = 45 ; pooled across animals; Supplemental Experimental Procedures). The slopes of the regression lines were significantly larger than 0 for SNRs exceeding 5% for group 1 but, although positive at SNRs >10%, failed to reach significance in the other groups (Table S1). The slopes differed significantly among the TO-PO groups for SNRs 10%, 15%, and 20% (general linear model; p < 0.05). Thus, the training effect is larger for neurons for which one of the trained orientations equals the preferred orientation (group 1), which are the most informative neurons for this task [17, 18]. Conversely, learning fine orientation discrimination predominantly affects neurons of which the trained differs from the preferred orientation [2, 4], which carry the most acute orientation signals [15, 17 19]. This agrees with the hypothesis that perceptual learning predominantly affects those neurons that are informative for the task at hand. However, the effects of coarse discrimination learning were not restricted to the most informative neurons. First, combining the 15% and 20% SNR data produced significant AUROC slopes in group 3 (slope = 0.11; p < 0.05) and a marginally significant trend in group 2 (slope = 0.07; p = 0.09). Second, the learning effect for the 80% SNR stimuli was present for neurons with TO-PO >0 (Figure S3). To relate neuronal responses to perceptual decisions trialby-trial, we computed choice probabilities (CP) [20]. Both animals showed mean grand CPs significantly greater than 0.50 (monkey M: mean CP 0.516; p < 0.01 [Signrank test; n = 103]; monkey P: 0.515; p < 0.05 [n = 67]). These CP values are low, but we did not select neurons. Indeed, the mean grand CP of neurons with AUROC >95% at 40% SNR was (p < [n = 36; Figure S3]), which compares well with those of selected MT neurons for motion direction discrimination [16, 20]. These CPs suggest that V4 neurons are linked, but not necessarily causally [21], to the monkey s decision in the coarse orientation discrimination task. Unlike [5], we failed to find an effect of learning on CP (Figure S3). We observed a robust, orientation-specific improvement in the discrimination capacity of V4 neurons during perceptual learning of a coarse orientation discrimination task. Both the orientation-specific nature of the improvement and its presence after stratification of TO-PO angles negate concerns that these effects were caused by factors unrelated to learning (e.g., sampling bias, successive recording-related factors). Unlike some learning effects for more complex stimuli in V1 [22 24], the improved V4 discriminability does not appear to result from changes in attention. First, attended stimuli typically elicit stronger responses, relative to nonattended stimuli in V4 [25 31], and we observed opposite trends in response strengths during training. Second, eight orientations at 80% SNR were presented interleaved in the passive fixation task, making it unlikely that the monkey employed the same attentional resources as in the more difficult orientation discrimination task. The presence of the training effect in both tasks suggests changes in bottom-up-driven responses to the trained orientations, although one cannot exclude top-down influences on the responses in the fixation task because we did not employ a difficult orthogonal task. Indeed, attention and other task-related factors can play a role in perceptual learning [32 34]. Our V4 data concur with changes in early visual cortical activations observed in human functional magnetic resonance imaging (fmri) (e.g [35 38]) and improved contrast sensitivity in cat area 17 [39] with perceptual learning. Our results are also in line with the enhanced discrimination of noisedegraded familiar compared to novel natural images observed in macaque V4 [40], although the mechanisms underlying the learning effects in the two studies might differ. The one or more sources of the discrepancy between our results and those of [5] are unknown. Possible sources are differences in training extent and protocol, but these are difficult to compare

5 Learning Coarse Discrimination in Monkey V given the different stimulus features. One might conjecture that V4 is more plastic than MT for the trained stimulus properties. The direction selectivities of MT neurons are similar to their V1 inputs [41], whereas V4 neurons have broader orientation tunings [4, 13, 42] and more complex spectral tunings [42] than V1 neurons. Perhaps the broad orientation sensitivity leaves more room in V4 for learning-dependent, orientationspecific reweighting of their V1 and/or V2 and local inputs. Such orientation-specific reweighting might explain why the increase in discriminability for the trained orientations was not consistently related to an increase in overall orientation sensitivity (SI) and was orientation specific and present at the onset of the orientation selectivity (response time course analysis; data not shown). The coarse orientation learning affected both the average firing rate and response variance. Part of the response variance might be due to the trial-to-trial variations of the stimulus noise. Hence, the reduction in Fano factor with learning might result from a reduced sensitivity to external noise, but this cannot be the sole cause of learning because it cannot explain its orientation specificity. Possibly, the decreased variance is due to changes of the functional connectivity between the neurons because a portion of response variance reflects trial-totrial covariation of responses within the neuronal network [43, 44]. We observed a trend toward decreased response strength after the lengthy training, in agreement with a fmri study showing that the initially increased activation can decrease and even disappear following long training [45], which supports synaptic downscaling as a mechanism of perceptual learning, as proposed to explain effects of sleep on learning [46]. Note that sleep affects perceptual learning in the coarse orientation discrimination task in humans [11]. A naive Bayes classifier using as input the responses of the early and late groups of neurons (Supplemental Experimental Procedures; Figure S3) showed a 54% decrease in threshold (75% correct) with learning, which compares favorably to the 44% decrease observed behaviorally. This contrasts with the V4 [4] fine discrimination learning data for which the modeled change was considerably smaller than the behavioral one. However, caution is required when comparing model and behavioral performances because possible learning-related changes in correlated activity between neurons, noise in the decision process, or eye movement planning were not taken into account. Note that the increased performance of the classifier with learning reflects both a change of the V4 response properties and of the readout of their responses. Thus, the present study does not exclude the possibility of changes in the sensory signal readout at the decision stage playing some role during perceptual learning. It demonstrates that models of visual perceptual learning need to consider changes in visual cortical signals in order to provide a full account of what becomes modified during the learning. Experimental Procedures Recordings Standard electrophysiological recording techniques were employed. Prelunate dorsal V4 was localized before implantation and verified before and between recordings using 3 Tesla MRI magnetization prepared rapid acquisition gradient echo (resolution 600 mm) in each animal. Animal care and experimental procedures were approved by the ethical committee of the Katholieke Universiteit Leuven Medical School. Subjects eye movements were monitored using infrared eye tracking (EyeLink). Eye position analysis showed no systematic differences between the eye positions for the two trained orientations for both E and L periods. Action potentials were recorded with glass-coated tungsten microelectrodes [4], thresholded online, and sorted offline (Offline Sorter; Plexon). The recording apparatus and electrodes were identical throughout the training. Stimuli The gamma-corrected stimuli were circular patches (2 diameter) containing a 100% contrast (luminance range: cd/m 2 ) sinusoidal grating (2 cycles/ ) spatially masked by noise, superimposed on a noise background that filled the 20 inch display. The SNR was manipulated by random replacement of 100% SNR of the grating pixels (pixel size 0.04 ) by noise. 0% SNR patches were detectable at stimulus onset, reducing spatial uncertainty. The noise was generated from a sinusoidal luminance distribution such that the luminance distributions in the stimulus patch and background were statistically equivalent. Tests RF mapping was performed using handplotting with small checkerboards [4] or gratings when checkerboards failed to produce a response. The trained orientations were 22.5 and in monkey M and 67.5 and in monkey P. Gratings of different SNRs and orientations were presented randomly interleaved. Task difficulty was kept relatively constant throughout the training by adjusting the frequency distributions of the various SNRs. The mean presentation frequencies for the SNR with the smallest frequency were 25 and 27 and for the SNR with the largest frequency were 68 and 62 in monkeys M and P, respectively. Stimulus duration during orientation discrimination and passive fixation was held at 250 ms. Data Analysis Spikes were counted within two windows: [2250 0] for baseline and [50 300] for gross response, 0 being stimulus onset. The animal fixated until at least 200 ms after stimulus offset, precluding the possibility that presaccadic activity (which can occur <100 ms before saccade onset in V4 [47 49]) intruded into spike counts within the response window. AUROC, Fano factor, and CP were computed using gross responses. Other analyses were performed on net response (gross-baseline). To compute AUROC, we determined the preferred (BT) and null (WT) orientations of a neuron by the mean net response at 40% SNR. The Fano factor is the trial-to-trial variance divided by the response averaged across those same trials. sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi SI = P n 2 Ri,sinð2OiÞ + i =1 P n i =1 Ri P n Ri,cosð2OiÞ i =1 R and O are the mean net response and orientation for stimulus i (n = 8), respectively. If the mean net response of a neuron was negative, the absolute minimum mean net response across the eight orientations was added to the responses to compute SI. SI is 0 for a neuron responding equally to all eight orientations and 1 when responding to only one orientation. To compute TO-PO angle, we equaled the preferred orientation with the orientation, out of the eight, with the greatest net response. To compute CPs, we z scored the responses for each SNR <40% with at least one correct and one incorrect choice for each orientation. The preferred and null orientations for the CP analyses were BT and WT, respectively, defined at 40% SNR. The grand CP [20] is the AUROC for the distributions of the z scores, pooled across SNR and orientation, and sorted according to the animal s choice. Although the relationship between session/cell number and psychometric/neurometric performance is nonlinear, we employed robust linear regression (Matlab function robustfit, protecting against outliers) because it is both simple and sufficient to demonstrate the effect of learning (Figure 1). Significance of the slope of the regression line was assessed by permuting session or cell number (n = 1,000). Comparison of the neural properties (Figure 2; Figure 3) in E and L groups was performed using ANOVA with group and monkey and SNR (if applicable) as factors. The interaction between group and monkey was significant only for the SI. We also computed robust regressions with normalized cell number (maximum cell number within each animal = 100) as predictor with the data pooled across monkeys (Supplemental Experimental Procedures). Supplemental Information Supplemental Information includes three figures, one table, and Supplemental Experimental Procedures and can be found with this article online at doi: /j.cub :

6 Current Biology Vol 21 No Acknowledgments We thank Inez Puttemans, Piet Kayenbergh, Gerrit Meulemans, Stijn Verstraeten, Marjan Docx, Wouter Depuydt, and Marc De Paep for assistance, and Bram-Ernst Verhoef, Santosh Mysore, and Steve Raiguel for critical reading of the draft. Supported by Geconcerteerde Onderzoeksactie (GOA/10/019), Interuniversitaire Attractiepool (IUAP P6/29), Programma Financiering (PF 10/008), and Fonds voor Wetenschappelijk Onderzoek Vlaanderen (G ). Received: April 26, 2011 Revised: July 8, 2011 Accepted: August 16, 2011 Published online: September 29, 2011 References 1. Ghose, G.M., Yang, T., and Maunsell, J.H. (2002). Physiological correlates of perceptual learning in monkey V1 and V2. J. Neurophysiol. 87, Schoups, A., Vogels, R., Qian, N., and Orban, G. (2001). Practising orientation identification improves orientation coding in V1 neurons. Nature 412, Yang, T., and Maunsell, J.H. (2004). The effect of perceptual learning on neuronal responses in monkey visual area V4. J. Neurosci. 24, Raiguel, S., Vogels, R., Mysore, S.G., and Orban, G.A. (2006). Learning to see the difference specifically alters the most informative V4 neurons. J. Neurosci. 26, Law, C.T., and Gold, J.I. (2008). Neural correlates of perceptual learning in a sensory-motor, but not a sensory, cortical area. Nat. Neurosci. 11, Petrov, A.A., Dosher, B.A., and Lu, Z.L. (2005). The dynamics of perceptual learning: an incremental reweighting model. Psychol. Rev. 112, Sotiropoulos, G., Seitz, A.R., and Seriès, P. (2011). Perceptual learning in visual hyperacuity: A reweighting model. Vision Res. 51, Law, C.T., and Gold, J.I. (2009). Reinforcement learning can account for associative and perceptual learning on a visual-decision task. Nat. Neurosci. 12, Frankó, E., Seitz, A.R., and Vogels, R. (2010). Dissociable neural effects of long-term stimulus-reward pairing in macaque visual cortex. J. Cogn. Neurosci. 22, Seitz, A.R., Kim, D., and Watanabe, T. (2009). Rewards evoke learning of unconsciously processed visual stimuli in adult humans. Neuron 61, Matarazzo, L., Frankó, E., Maquet, P., and Vogels, R. (2008). Offline processing of memories induced by perceptual visual learning during subsequent wakefulness and sleep: A behavioral study. J. Vis. 8, 7 9, Felleman, D.J., and Van Essen, D.C. (1991). Distributed hierarchical processing in the primate cerebral cortex. Cereb. Cortex 1, Desimone, R., and Schein, S.J. (1987). Visual properties of neurons in area V4 of the macaque: sensitivity to stimulus form. J. Neurophysiol. 57, Newsome, W.T., Britten, K.H., and Movshon, J.A. (1989). Neuronal correlates of a perceptual decision. Nature 341, Vogels, R., and Orban, G.A. (1990). How well do response changes of striate neurons signal differences in orientation: a study in the discriminating monkey. J. Neurosci. 10, Cohen, M.R., and Newsome, W.T. (2009). Estimates of the contribution of single neurons to perception depend on timescale and noise correlation. J. Neurosci. 29, Butts, D.A., and Goldman, M.S. (2006). Tuning curves, neuronal variability, and sensory coding. PLoS Biol. 4, e Jazayeri, M., and Movshon, J.A. (2006). Optimal representation of sensory information by neural populations. Nat. Neurosci. 9, Regan, D., and Beverley, K.I. (1985). Postadaptation orientation discrimination. J. Opt. Soc. Am. A 2, Britten, K.H., Newsome, W.T., Shadlen, M.N., Celebrini, S., and Movshon, J.A. (1996). A relationship between behavioral choice and the visual responses of neurons in macaque MT. Vis. Neurosci. 13, Nienborg, H., and Cumming, B. (2010). Correlations between the activity of sensory neurons and behavior: how much do they tell us about a neuron s causality? Curr. Opin. Neurobiol. 3, Crist, R.E., Li, W., and Gilbert, C.D. (2001). Learning to see: experience and attention in primary visual cortex. Nat. Neurosci. 4, Li, W., Piëch, V., and Gilbert, C.D. (2008). Learning to link visual contours. Neuron 57, Li, W., Piëch, V., and Gilbert, C.D. (2004). Perceptual learning and topdown influences in primary visual cortex. Nat. Neurosci. 7, Moran, J., and Desimone, R. (1985). Selective attention gates visual processing in the extrastriate cortex. Science 229, Reynolds, J.H., and Chelazzi, L. (2004). Attentional modulation of visual processing. Annu. Rev. Neurosci. 27, Maunsell, J.H., and Treue, S. (2006). Feature-based attention in visual cortex. Trends Neurosci. 29, Reynolds, J.H., Pasternak, T., and Desimone, R. (2000). Attention increases sensitivity of V4 neurons. Neuron 26, Hayden, B.Y., and Gallant, J.L. (2005). Time course of attention reveals different mechanisms for spatial and feature-based attention in area V4. Neuron 47, Cohen, M.R., and Maunsell, J.H. (2010). A neuronal population measure of attention predicts behavioral performance on individual trials. J. Neurosci. 30, Ghose, G.M., and Maunsell, J.H. (2008). Spatial summation can explain the attentional modulation of neuronal responses to multiple stimuli in area V4. J. Neurosci. 28, Roelfsema, P.R., and van Ooyen, A. (2005). Attention-gated reinforcement learning of internal representations for classification. Neural Comput. 17, Seitz, A.R., and Dinse, H.R. (2007). A common framework for perceptual learning. Curr. Opin. Neurobiol. 17, Roelfsema, P.R., van Ooyen, A., and Watanabe, T. (2010). Perceptual learning rules based on reinforcers and attention. Trends Cogn. Sci. (Regul. Ed.) 14, Vaina, L.M., Belliveau, J.W., des Roziers, E.B., and Zeffiro, T.A. (1998). Neural systems underlying learning and representation of global motion. Proc. Natl. Acad. Sci. USA 95, Schwartz, S., Maquet, P., and Frith, C. (2002). Neural correlates of perceptual learning: a functional MRI study of visual texture discrimination. Proc. Natl. Acad. Sci. USA 99, Furmanski, C.S., Schluppeck, D., and Engel, S.A. (2004). Learning strengthens the response of primary visual cortex to simple patterns. Curr. Biol. 14, Kourtzi, Z., Betts, L.R., Sarkheil, P., and Welchman, A.E. (2005). Distributed neural plasticity for shape learning in the human visual cortex. PLoS Biol. 3, e Hua, T., Bao, P., Huang, C.B., Wang, Z., Xu, J., Zhou, Y., and Lu, Z.L. (2010). Perceptual learning improves contrast sensitivity of V1 neurons in cats. Curr. Biol. 20, Rainer, G., Lee, H., and Logothetis, N.K. (2004). The effect of learning on the function of monkey extrastriate visual cortex. PLoS Biol. 2, E Born, R.T., and Bradley, D.C. (2005). Structure and function of visual area MT. Annu. Rev. Neurosci. 28, David, S.V., Hayden, B.Y., and Gallant, J.L. (2006). Spectral receptive field properties explain shape selectivity in area V4. J. Neurophysiol. 96, Mitchell, J.F., Sundberg, K.A., and Reynolds, J.H. (2009). Spatial attention decorrelates intrinsic activity fluctuations in macaque area V4. Neuron 63, Churchland, M.M., Yu, B.M., Cunningham, J.P., Sugrue, L.P., Cohen, M.R., Corrado, G.S., Newsome, W.T., Clark, A.M., Hosseini, P., Scott, B.B., et al. (2010). Stimulus onset quenches neural variability: a widespread cortical phenomenon. Nat. Neurosci. 13, Yotsumoto, Y., Watanabe, T., and Sasaki, Y. (2008). Different dynamics of performance and brain activation in the time course of perceptual learning. Neuron 57, Tononi, G., and Cirelli, C. (2003). Sleep and synaptic homeostasis: a hypothesis. Brain Res. Bull. 62, Tolias, A.S., Moore, T., Smirnakis, S.M., Tehovnik, E.J., Siapas, A.G., and Schiller, P.H. (2001). Eye movements modulate visual receptive fields of V4 neurons. Neuron 29, Han, X., Xian, S.X., and Moore, T. (2009). Dynamic sensitivity of area V4 neurons during saccade preparation. Proc. Natl. Acad. Sci. USA 106, Steinmetz, N.A., and Moore, T. (2010). Changes in the response rate and response variability of area V4 neurons during the preparation of saccadic eye movements. J. Neurophysiol. 103,