(PDF) Task relevancy and demand modulate double-training enabled transfer of perceptual learning

Vision Research 61 (2012) 33–38 Contents lists available at ScienceDirect Vision Research journal homepage: www.elsevier.com/locate/visres Task relevancy and demand modulate double-training enabled transfer of perceptual learning Rui Wang a, Jun-Yun Zhang a, Stanley A. Klein b, Dennis M. Levi b, Cong Yu a,⇑ a State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing 100875, China b School of Optometry and Helen Wills Neuroscience Institute, University of California, Berkeley, CA 94720, USA a r t i c l e i n f o a b s t r a c t Article history: Location-specific perceptual learning can be rendered transferrable to a new location with double train- Received 20 April 2011 ing, in which feature training (e.g., contrast) is accompanied by additional location training at the new Received in revised form 28 June 2011 location even with an irrelevant task (e.g. orientation). Here we investigated the impact of relevancy Available online 26 July 2011 (to feature training) and demand of location training tasks on double training enabled learning transfer. We found that location training with an irrelevant task (Gabor vs. letter judgment, or contrast discrim- Keywords: ination) limited transfer of Vernier learning to the trained orientation only. However, performing a rele- Perceptual learning vant suprathreshold orthogonal Vernier task prompted additional transfer to an untrained orthogonal Double training Location specificity orientation. In addition, the amount of learning transfer may depend on the demand of location training Transfer as well as the double training procedure. These results characterize how double training potentiates the functional connections between a learned high-level decision unit and visual inputs from an untrained location to enable transfer of learning across retinal locations. Ó 2011 Elsevier Ltd. All rights reserved. 1. Introduction training at the new location using an irrelevant orientation dis- crimination task (Xiao et al., 2008). Similarly, with a training- Visual perceptual learning improves discrimination of many plus-exposure (TPE) procedure, perceptual learning of orientation basic visual features, such as contrast, orientation, Vernier, and in foveal vision, which is otherwise specific to the trained orienta- texture. A key feature of it is that learning is often specific to the tion, can transfer completely to an orthogonal orientation if an ob- trained retinal location and orientation (Ahissar & Hochstein, server is also passively exposed to the orthogonal orientation while 1997; Crist et al., 1997; Fahle, 1994, 1997; Karni & Sagi, 1991; performing an irrelevant contrast discrimination task (Zhang, Saarinen & Levi, 1995; Schoups, Vogels, & Orban, 1995; Shiu & Zhang, et al., 2010). Pashler, 1992; Yu, Klein, & Levi, 2004). The location and orientation These transfer results suggest that at least in some cases per- specificities place important constraints on various perceptual ceptual learning is more a general learning process and most likely learning models and theories (Adini, Sagi, & Tsodyks, 2002; occurs at a high decision level of information processing beyond Bejjanki et al., 2011; Dosher & Lu, 1998; Law & Gold, 2009; Mollon the retinotopic and orientation selective visual cortex (Xiao et al., & Danilova, 1996; Poggio, Fahle, & Edelman, 1992; Teich & Qian, 2008; Zhang, Zhang, et al., 2010; Zhang, Xiao, et al., 2010). This 2003; Zhaoping, Herzog, & Dayan, 2003), as it is suggested that argument is supported by a recent fMRI study indicating that the modeling the neural mechanisms underlying perceptual learning site of perceptual learning may be located in the human medial must account for these specificities (Tsodyks & Gilbert, 2004). front cortex, which is also the site of perceptual decision making However, in recent studies we demonstrated that location spec- (Kahnt et al., 2011). We thus proposed a rule-based learning theory ificity and orientation specificity can be decoupled from perceptual to explain visual perceptual learning and its specificity and transfer learning in a variety of visual tasks with appropriate training (Zhang, Zhang, et al., 2010). This theory posits that a high-level procedures (Xiao et al., 2008; Zhang, Xiao, et al., 2010; Zhang, decision unit learns the rules for performing a visual task through Zhang, et al., 2010). For example, with a feature-plus-location dou- training. However, the learned rules cannot be applied to a new ble training procedure, perceptual learning of contrast discrimina- location or orientation automatically because the decision unit tion (feature training), which is otherwise location specific, can cannot functionally connect to the visual inputs representing the transfer completely to a new location following additional location new location or orientation with sufficient strength. These inputs are unattended or even suppressed during training when attention ⇑ Corresponding author. Fax: +86 10 5880 6154. is allocated to the trained location or orientation (Gal et al., 2009; E-mail address:

(C. Yu). Sylvester et al., 2009). It is double training and TPE training that 0042-6989/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.visres.2011.07.019 34 R. Wang et al. / Vision Research 61 (2012) 33–38 reactivate these new inputs, so that the functional connections can 2.3. Procedure be strengthened to enable rule application and transfer of learning. The current study manipulated the task relevancy (to feature Vernier and contrast discrimination thresholds were measured training) and the demand of location training and investigates their with a single-interval discrimination staircase procedure. In each impact on the transfer of feature (Vernier) learning from a diagonal trial, the stimulus was presented for 200 ms. For Vernier discrimi- quadrant of the visual field. We used three levels of task relevance nation, an observer judged whether the right Gabor was higher or (passive stimulus exposure, irrelevant, and relevant) and three de- lower than the left Gabor for a horizontal Vernier, or the lower mand levels (passive stimulus exposure, suprathreshold, and near- Gabor was to the left or right of the upper Gabor for a vertical Ver- threshold) with the location training. Here passive stimulus expo- nier. For contrast discrimination, an observer judged which of the sure was least task relevant and least demanding, and was there- two Gabors had a higher contrast. A small fixation cross preceded fore used in both categories as baselines. Vernier learning is each trial by 400 ms and stayed through the trial. Auditory feed- found to be strictly specific to the trained location (Xiao et al., back was given on incorrect responses. Thresholds were estimated 2008), so its transfer to a new location would serve as an excellent using a classical 3-down-1-up staircase rule that resulted in a indicator of the impact of location training under various task rel- 79.4% convergence level. Each staircase consisted of four prelimin- evancy and demand conditions. Our results show that an obser- ary reversals and six experimental reversals (approximately 50 tri- ver’s active participation in a visual task at the new location, als). The step size of the staircase was 0.05 log units. The geometric even a very simple one, rather than passive exposure to the stimuli mean of the experimental reversals was taken as the threshold for is necessary for Vernier learning to transfer. Moreover, task-irrele- each staircase run. vant location training limits learning transfer to only the trained Vernier orientation, but a relevant suprathreshold orthogonal Ver- 2.4. Statistics nier task allows additional learning transfer to the orthogonal ori- entation. In addition, the amount of learning transfer may depend The amount of perceptual learning was quantified as the Per- on the demand level in the location training task as well as the par- cent Improvement (PI), where PI = (ThreshPre ThreshPost)/ ticular double training procedure (simultaneous or sequential). In ThreshPre. We hypothesized that thresholds were lower after train- the context of our rule-based perceptual learning theory, these re- ing in perceptual learning experiments, so that one-tailed hypoth- sults characterize double training potentiating the functional con- esis tests were used. nections between a learned high-level decision unit and new visual inputs from an untrained retinal location, which makes the trans- fer of perceptual learning possible. 3. Results 3.1. Baseline: feature training plus passive exposure to stimuli at the 2. Methods transfer location 2.1. Observers and apparatus We first studied whether passive exposure to stimuli at the transfer location would enable the transfer of Vernier learning. Forty observers (undergraduate students in their early twenties Passive stimulus exposure was least task-relevant and required at Beijing Normal University) with normal or corrected-to-normal the least effort by the observer, so this measure provided baselines vision participated in this study. All were new to psychophysical for the impacts of task relevancy and demand of location training experiments and were unaware of the purpose of the study. In- on learning transfer. Sixteen observers practiced Vernier discrimi- formed consent was obtained from each observer before data nation at a horizontal or vertical orientation in one of four visual collection. quadrants (ori1_loc1) at 5° retinal eccentricity while an orthogonal The stimuli were generated by a Matlab-based WinVis pro- Vernier in the diagonal quadrant (ori2_loc2) was flashed simulta- gram (Neurometrics Institute, Oakland, CA) and presented on a neously for the same 200 ms duration (Fig. 1A). The offset of the 21-in. Sony G520 color monitor (2048 pixel 1536 pixel, flashed Vernier was randomly set at ±15 arcmin, or approximately 0.19 mm 0.19 mm per pixel, 75 Hz frame rate). The mean lumi- ±2.5 times the mean pre-training Vernier threshold. Because nance was 50 cd/m2. Luminance of the monitor was linearized by nearly all attention was allocated to Vernier discrimination at an 8-bit look-up table. Viewing was monocular with one eye cov- the trained location, the flashed Vernier was only passively viewed ered with a translucent plastic pad, and a chin-and-head rest by an observer. Significant learning was evident at the trained helped stabilize the head of the observer. The viewing distance orientation and location (ori1_loc1, Mean Percent Improvement was 1.5 m. Experiments were run in a dimly lit room. (MPI) = 27.3 ± 2.9%, p < 0.001, paired t-test) after six to seven 2-h daily sessions (Fig. 1B and C). However, learning transferred little to the same or orthogonal orientations in the diagonal visual quad- 2.2. Stimuli rant with passive stimulus exposure (ori1_loc2, MPI = 0.5 ± 4.4%, p = 0.55; ori2_loc2, MPI = 1.2 ± 4.0%, p = 0.39), as well as the orthog- The Vernier stimulus consisted of a pair of identical Gabor onal orientation at the trained location (ori2_loc1, MPI = 1.5 ± 5.8%, patches (Gaussian windowed sinusoidal gratings) on a mean lumi- p = 0.40) (Fig. 1B and C). These results indicate that mere passive nance screen background, which was centered in a visual quadrant exposure to the stimuli at the transfer location cannot replace ac- at 5° retinal eccentricity (Fig. 1A). The two Gabors had the same spa- tive location training to override the specificity of Vernier learning tial frequency (3 cpd), fixed phase (90°), standard deviation (0.29°), revealed in our previous study with identical stimuli (Xiao et al., contrast (0.47), orientation (either vertical or horizontal), and a cen- 2008). ter-to-center distance of 4k. To form a specific Vernier offset, the po- sition of each Gabor shifted half the Vernier offset away in opposite 3.2. Double training: feature training plus location training with an directions perpendicular to the Gabor orientation. irrelevant suprathreshold task The same Vernier stimulus was also used for contrast discrimi- nation training with the Vernier offset jittered at ±15 arcmin. The Among the 16 observers in the passive stimulus exposure contrasts of the two Gabors were set at 0.47 and 0.47 + DC. experiment (Fig. 1), one did not learn (Vernier performance R. Wang et al. / Vision Research 61 (2012) 33–38 35 A Vernier B 9 ori1_loc2 C 40 ori1 training 8 ori2_loc2 loc1 Mean % improvement 7 ori2_loc1 30 Ver (arcmin) 6 20 ori1 ori2 ori2 5 loc2 loc2 loc1 10 ori1_loc1 4 Passive N=16 0 exposure 3 -10 1 2 3 4 5 6 7 Training Transfer Session Fig. 1. Transfer of Vernier learning to a diagonal quadrant where an orthogonal Vernier was passively exposed. (A) Stimuli. Observers practiced Vernier discrimination at ori1_loc1 while passively exposed to an orthogonal Vernier flashed simultaneously in the diagonal quadrant (ori2_loc2). (B) The mean session-by-session threshold changes in the trained condition and pre- and post-training thresholds at the transfer conditions. (C) The MPIs of Vernier performance in the trained condition (left bar) and untrained transfer conditions (right three bars). improved by <3%) and two showed significant transfer (the trans- identical to the 30.3 ± 2.9% MPI in Fig. 2B. Vernier learning again fer index (TI), defined as the ratio of transfer/training performance transferred to the same trained orientation ori1_loc2 only improvements, was >0.5). The remaining thirteen showed little or (MPI = 18.2 ± 6.6%, p = 0.015), but not to orthogonal ori2_loc2 no improvement in performance in the transfer conditions (mean (MPI = 5.1 ± 4.9%, p = 0.17) and ori2_loc1 (MPI = 4.3 ± 5.1%, TI = 0.15 ± 0.14). We were able to call eleven of these thirteen p = 0.21) (Fig. 2D and E). These results confirmed that the transfer observers back and split them into two groups, each performing of learning was specific to the trained orientation under the cur- a suprathreshold task either irrelevant or relevant to feature learn- rent double training conditions. However, the transfer was clearly ing. The first group of five observers in the current experiment partial with lower MPI at ori1_loc2 than at ori1_loc2 (p = 0.014) judged whether a pair of Gabors (Fig. 2A, same as the flashing and similar pre-training thresholds. Gabors in Fig. 1A, presented on 80% of the trials,) or an uppercase letter E (20% of the trials) appeared in the diagonal transfer quad- 3.3. Double training: feature training plus location training with a rant for four sessions. Here the observers were forced to perform relevant suprathreshold task an irrelevant, non-demanding suprathreshold task (mean accu- racy = 99.7%) at the transfer location. For these five observers, the The second group of six observers who did not show much previous Vernier training produced significant improvement at or- transfer after passive stimulus exposure (Fig. 1) judged an orthog- i1_loc1 (MPI = 30.3 ± 2.9%, p < 0.001, Fig. 2B and C), but learning did onal Vernier with the offset set at five times the threshold after not transfer to untrained ori1_loc2 (MPI = 9.7 ± 8.0%, p = 0.85), or- passive stimulus exposure at the transfer location for four sessions i2_loc2 (MPI = 1.9 ± 8.1%, p = 0.59), and ori2_loc1 at the trained (Fig. 3A). Since the offset was well above threshold, the task was location (MPI = 9.5 ± 9.4%, p = 0.81) (Fig. 2B and C). After the not demanding (mean accuracy = 99.1%), but this time the task new Gabors vs. E judgments, Vernier performance was improved was relevant to feature (Vernier) learning. For these six observers, at ori1_loc2 (MPI = 14.5 ± 4.0%, p = 0.012), but not at orthogonal previous Vernier training produced significant improvement at or- ori2_loc2 (MPI = 2.8 ± 4.7%, p = 0.71) and ori2_loc1 (MPI = 9.6 ± i1_loc1 (MPI = 24.1 ± 3.8%, p < 0.001), and learning did not transfer 10.4%, p = 0.79) (Fig. 2B and C), indicating that location training to ori1_loc2 (MPI = 0.3 ± 3.1%, p = 0.47), ori2_loc2 (MPI = 2.4 ± 2.3%, with an irrelevant suprathreshold task can enable the transfer of p = 0.18), and ori2_loc1 (MPI = 5.6 ± 8.7%, p = 0.73) (Fig. 3B and C). Vernier learning across retinal locations, but the transfer is limited However, after the suprathreshold Vernier judgments at ori2_ to the trained orientation. On the other hand, the lower MPI at or- loc2, Vernier learning transferred not only to ori1_loc2 (MPI = i1_loc2 (14.5% vs. 30.3%) appeared to indicate partial transfer of 15.5 ± 3.1%, p = 0.002), but also to the orthogonal ori2_loc2 (MPI learning (p = 0.047 for MPIs at ori1_loc1 vs. ori1_loc2). However, = 20.4 ± 4.6%, p = 0.003) and ori2_loc1 (MPI = 13.2 ± 5.2%, p = it was unclear whether the partial transfer truly resulted from 0.026) (Fig. 3B and C). Moreover, after this two-phase double location training with an irrelevant suprathreshold task, or from training there was no significant difference in performance a performance ceiling effect at the transfer location (the post-train- improvement among the trained condition (MPI = 18.6 ± 5.0%, ing thresholds were 4.34 ± 0.47 arcmin at ori1_loc1 and 4.50 ± 0.16 p = 0.007) and three transfer conditions (p = 0.30, repeated arcmin at ori1_loc2, p = 0.37) combined with the lower pre-train- measures ANOVA). ing thresholds at ori1_loc2 (Fig. 2B). In the latter case the transfer We again had seven new observers repeat the experiment in a of learning could actually be complete. simultaneous procedure. They practiced Vernier discrimination at In our previous study (Xiao et al., 2008) double training enabled ori1_loc1 and judged suprathreshold orthogonal Vernier (five complete transfer of learning regardless of whether feature and times the mean pre-training threshold) at ori2_loc2 in alternating location training (with an irrelevant but demanding near-thresh- blocks of trials in the same session for five sessions. Consistent old task) was performed simultaneously or sequentially. However, with the sequential double training data, Vernier learning at ori1_- the above results would suggest that demanding location training loc1 (MPI = 38.9 ± 4.3%, p < 0.001, which is higher than the magni- might be unnecessary if the transfer of learning is indeed complete. tude in the sequential condition described above), transferred to Would the above findings of orientation specificity in learning not only ori1_loc2 (MPI = 17.8 ± 5.8%, p = 0.011), but also orthogo- transfer and potentially complete learning transfer be replicable nal ori2_loc2 (MPI = 20.9 ± 4.9%, p = 0.003) and ori2_loc1 (MPI = 18.7 in a simultaneous double training procedure? We had eight new ± 8.4%, p = 0.034) (Fig. 3D and E). However, as in Fig. 2D, the simul- observers practice Vernier discrimination at ori1_loc1 and judge taneous procedure only produced partial transfer (p = 0.015 for the the same Gabors or E in the diagonal quadrant in alternating blocks contrast between the training MPI and three transfer MPIs, re- of trials in a same session for five sessions. Training improved Ver- peated measures ANOVA). Taking Figs. 2 and 3 results together, nier performance at ori1_loc1 by 29.6 ± 4.0% (p < 0.001), nearly there appears to be a trend that location training with a 36 R. Wang et al. / Vision Research 61 (2012) 33–38 Phase 1 Phase 2 B 9 ori1_loc2 C ori1 40 ori2_loc2 loc1 Mean % improvement 8 Gabors/E? ori2_loc1 Ver (arcmin) 7 30 ori1 Accuracy 6 loc2 20 5 ori1 ori2 ori2 ori2 ori2 10 loc2 loc2 loc1 loc2 loc1 A 4 ori1_loc1 N=5 0 3 2 4 6 8 10 12 -10 Training Transfer Transfer Session D 9 1.00 E 50 Mean % improvement 8 ori1 Gabors/E? Ver (arcmin) 7 40 loc1 0.95 Accuracy ori1 6 30 loc2 0.90 5 20 ori2 ori2 loc2 loc1 4 N=8 0.85 10 3 0.80 0 1 2 3 4 5 6 7 Training Transfer Session Fig. 2. Transfer of Vernier learning to a diagonal quadrant trained with an irrelevant suprathreshold task. (A) Stimuli. Observers judged whether the stimuli were Gabors or a letter E. (B) Sequential procedure: Vernier performance in various training and transfer conditions before and after initial training at ori1_loc1 (replotted from Fig. 1) and after the Gabors-or-E judgments at loc2. (C) The MPIs in the trained condition and untrained transfer conditions after two experimental phases. (D) Simultaneous procedure: Vernier training at ori1_loc1 and Gabors-or-E judgments at loc2 in alternating blocks in the same session. (E) The MPIs in the trained condition (left bar) and untrained transfer conditions (right three bars) after the simultaneous procedure. Phase 1 Phase 2 A B 9 ori1_loc2 C ori1 Mean % improvement 8 ori2_loc2 ori2 Suprathreshold loc1 ori2_loc1 loc2 Ver (arcmin) 7 Vernier ori1 ori2 Accuracy loc2 loc1 6 5 ori1 ori2 ori2 loc2 loc2 loc1 ori1_loc1 4 N=6 Suprathreshold 3 Vernier 2 4 6 8 10 12 Training Transfer Transfer Session D 9 E ori1 F 9 Mean % improvement 8 loc1 8 Ver (arcmin) 7 7 Ver (arcmin) Accuracy Accuracy 6 ori1 ori2 ori2 6 loc2 loc2 loc1 5 5 4 4 N=7 N=5 3 3 Training Transfer 1 2 3 4 5 6 Session Session Fig. 3. Transfer of Vernier learning to a diagonal quadrant trained with a relevant suprathreshold Vernier task. (A) Stimuli. The suprathreshold Vernier at ori2_loc2 was five times the threshold. (B) Sequential procedure: Vernier performance before and after initial training at ori1_loc1, and after suprathreshold Vernier judgments at ori2_loc2. (C) The MPIs in the trained condition and untrained transfer conditions after two experimental phases. (D) Simultaneous procedure: Vernier discrimination at ori1_loc1 and suprathreshold Vernier judgments at ori2_loc2 in alternating blocks in the same session. (E) The MPIs in the trained condition (left bar) and untrained transfer conditions (right three bars) after the simultaneous procedure. (F) Control: Pre- and post-training thresholds in three transfer conditions while only the suprathreshold Vernier task was trained. suprathreshold task may enable complete transfer in a sequential In addition, we ran a control experiment in which five new double-training procedure, but only partial transfer in a simulta- observers performed only the same suprathreshold Vernier task neous procedure. (five times the mean pre-training threshold) at ori2_loc2 for four R. Wang et al. / Vision Research 61 (2012) 33–38 37 △Ver_ori1_loc2 △Con A B △Ver_ori1_loc2 C ori2 Vernier 9 △Ver Mean % improvement △Ver_ori2_loc1 △Ver loc2 8 ori1 ori1 △Con_ori2_loc2 loc2 Ver (arcmin) 7 loc1 6 △Ver △Ver Con 5 ori2 ori2 loc2 loc1 4 △Ver_ori1_loc1 N=9 Contrast 3 Training Transfer Session Fig. 4. Transfer of Vernier learning to a diagonal quadrant trained with an irrelevant near-threshold contrast discrimination task. (A) Stimuli for Vernier and contrast double training. For contrast discrimination the observers reported which Gabor had a higher contrast. (B) The mean session-by-session Vernier and contrast threshold changes in the trained conditions and pre- and post-training Vernier thresholds in the transfer conditions. (C) The MPIs in the trained (left two bars) and untrained transfer conditions (right three bars). sessions. This suprathreshold task had no impact on the Vernier latter shows that foveal learning of orientation discrimination thresholds at ori1_loc2, ori2_loc2, and ori2_loc1 (Fig. 3F), exclud- and contrast discrimination can completely transfer to an orthog- ing the possibility that the transfers described above were caused onal orientation that the observers are passively exposed to in an by the suprathreshold Vernier task alone. irrelevant task. This discrepancy is not caused by task differences because the TPE training also enabled nearly full transfer of foveal 3.4. Double training: feature training plus location training with an Vernier learning to an orthogonal orientation (our unpublished irrelevant but demanding near-threshold task data). One possibility is that transfer of peripheral learning in the current study to an orthogonal orientation in the diagonal quad- We know from our previous study that learning can transfer rant involves one extra step of functional connection, in contrast completely to a new location if feature training is accompanied to foveal learning that connects to an orthogonal orientation at with irrelevant but demanding near-threshold training at the the same location. Thus extra processing is required to activate new location (Xiao et al., 2008). Here we examined whether this the orthogonal inputs necessary for strengthening the extra func- complete transfer was still specific to the trained orientation. Nine tional connections and enabling learning transfer. new observers practiced Vernier discrimination at ori1_loc1 and In addition, the relationship between the amount of learning near-threshold contrast discrimination simultaneously using the transfer and the task demand is less straightforward. Learning ap- same Vernier stimulus (see Section 2) at ori2_loc2 in alternating pears to transfer more with the task demand in simultaneous dou- blocks for five sessions. Training improved Vernier threshold at or- ble training procedures. But the transfer may already be complete i1_loc1 (MPI = 29.7 ± 2.4%, p < 0.001) and contrast threshold at or- in sequential procedures even with much less demanding supra- i2_loc2 (MPI = 35.9 ± 4.1%, p < 0.001). Vernier discrimination at threshold location training tasks. The causes of this difference re- ori1_loc2 also improved (MPI = 31.2 ± 3.0%, p < 0.001), as much as quire further investigation. According to our rule-based learning that at trained ori1_loc1 (MPIs at ori1_loc1 vs. ori1_loc2, p = theory, location specificity results from inattention to, or even sup- 0.26), showing complete learning transfer. However, learning did pression of visual inputs originating from the untrained location not transfer much to orthogonal ori2_loc2 (MPI = 5.0 ± 9.4%, (Zhang, Zhang, et al., 2010). Location training reactivates these p = 0.30) and ori2_loc1 (MPI = 6.8 ± 4.8%, p = 0.10), showing the new inputs to connect the learned high-level decision unit to these same orientation specificity in Fig. 2. inputs and enable transfer. It is possible that during simultaneous double training the reactivation by a suprathreshold location 4. Discussion training task is not strong enough to fully remove the suppression, but in a sequential procedure the suppression weakens as time This study revealed several facts regarding double training that passes and the reactivation is now able to remove the remaining were previously unknown. Active location training is indeed neces- suppression. sary to enable learning transfer across retinal locations. However, if Recently Sagi (2011) proposed an overfitting theory to explain location training is task-irrelevant, as in our previous study (Xiao perceptual learning. This theory posits that when statistically mod- et al., 2008), learning transfer may be limited to the trained feature eling the stimulus representation, an observer through practice orientation only. However, a relevant suprathreshold task at the could ‘‘overfit’’ the spurious and accidental variations of the stim- untrained orientation is sufficient to expand transfer to an un- ulus inputs resulting from local noise specific to the trained retinal trained orientation, even at the trained location. The latter is espe- location and orientation (Mollon & Danilova, 1996), which leads to cially interesting because it shows that orientation specificity can perceptual learning of the task at hand. But perceptual learning be decoupled from perceptual learning by a primer at a different does not transfer to a new location or orientation where the local retinal location. It is also worth noting that in the current study noise changes. Sagi (2011) suggests that double training enables learning transferred to a diagonal quadrant in the untrained visual transfer of learning because ‘‘the perceptual modeling process makes hemifield, rather than to the other quadrant in the same hemifield use of features shared by all locations and tasks the observer is exposed (Xiao et al., 2008). The inter-hemispheric transfer of learning pro- to during training’’, but ‘‘overfitting predicts much less learning with vides additional evidence that perceptual learning occurs in non- two-location as compared with one location training’’. This prediction retinotopic high-level brain areas. is inconsistent with data shown in Fig. 4 as well as in our previous On the other hand, the orientation specificity of learning trans- experiments (Xiao et al., 2008) that double training often produces fer with task-irrelevant location training appears to be at odds with as much learning as in single training and complete transfer of our previous TPE training data (Zhang, Zhang, et al., 2010). The learning across retinal locations. 38 R. Wang et al. / Vision Research 61 (2012) 33–38 Acknowledgments Law, C. T., & Gold, J. I. (2009). Reinforcement learning can account for associative and perceptual learning on a visual-decision task. Nature Neuroscience, 12(5), 655–663. This research is supported by the Natural Science Foundation of Mollon, J. D., & Danilova, M. V. (1996). Three remarks on perceptual learning. Spatial China Grants 30725018 (CY) and 31000459 (JYZ) and the United Vision, 10(1), 51–58. Poggio, T., Fahle, M., & Edelman, S. (1992). Fast perceptual learning in visual States National Institute of Health Grants R01-04776 (SAK) and hyperacuity. Science, 256(5059), 1018–1021. R01-01728 (DML). Saarinen, J., & Levi, D. M. (1995). Perceptual learning in vernier acuity: What is learned? Vision Research, 35(4), 519–527. Sagi, D. (2011). Perceptual learning in vision research. Vision Research, 51(13), 1552–1566. Schoups, A., Vogels, R., & Orban, G. A. (1995). Human perceptual learning in References identifying the oblique orientation: Retinotopy, orientation specificity and monocularity. Journal of Physiology, 483(Pt 3), 797–810. Adini, Y., Sagi, D., & Tsodyks, M. (2002). Context-enabled learning in the human Shiu, L. P., & Pashler, H. (1992). Improvement in line orientation discrimination is visual system. Nature, 415, 790–793. retinally local but dependent on cognitive set. Perception & Psychophysics, 52(5), Ahissar, M., & Hochstein, S. (1997). Task difficulty and the specificity of perceptual 582–588. learning. Nature, 387(6631), 401–406. Sylvester, C. M., Shulman, G. L., Jack, A. I., & Corbetta, M. (2009). Anticipatory and Bejjanki, V. R., Beck, J. M., Lu, Z. L., & Pouget, A. (2011). Perceptual learning as stimulus-evoked blood oxygenation level-dependent modulations related to improved probabilistic inference in early sensory areas. Nature Neuroscience. spatial attention reflect a common additive signal. Journal of Neuroscience, doi:10.1038/nn.2796. 29(34), 10671–10682. Crist, R. E., Kapadia, M. K., Westheimer, G., & Gilbert, C. D. (1997). Perceptual Teich, A. F., & Qian, N. (2003). Learning and adaptation in a recurrent model of V1 learning of spatial localization: Specificity for orientation, position, and context. orientation selectivity. Journal of Neurophysiology, 89(4), 2086–2100. Journal of Neurophysiology, 78(6), 2889–2894. Tsodyks, M., & Gilbert, C. (2004). Neural networks and perceptual learning. Nature, Dosher, B. A., & Lu, Z. L. (1998). Perceptual learning reflects external noise filtering 431(7010), 775–781. and internal noise reduction through channel reweighting. Proceedings of the Xiao, L. Q., Zhang, J. Y., Wang, R., Klein, S. A., Levi, D. M., & Yu, C. (2008). Complete National Academy of Sciences of the United States of America, 95(23), transfer of perceptual learning across retinal locations enabled by double 13988–13993. training. Current Biology, 18(24), 1922–1926. Fahle, M. (1994). Human pattern recognition: Parallel processing and perceptual Yu, C., Klein, S. A., & Levi, D. M. (2004). Perceptual learning in contrast learning. Perception, 23(4), 411–427. discrimination and the (minimal) role of context. Journal of Vision, 4(3), Fahle, M. (1997). Specificity of learning curvature, orientation, and vernier 169–182. discriminations. Vision Research, 37, 1885–1895. Zhang, T., Xiao, L. Q., Klein, S. A., Levi, D. M., & Yu, C. (2010). Decoupling location Gal, V., Kozak, L. R., Kobor, I., Banko, E. M., Serences, J. T., & Vidnyanszky, Z. (2009). specificity from perceptual learning of orientation discrimination. Vision Learning to filter out visual distractors. European Journal of Neuroscience, 29(8), Research, 50(4), 368–374. 1723–1731. Zhang, J. Y., Zhang, G. L., Xiao, L. Q., Klein, S. A., Levi, D. M., & Yu, C. (2010). Rule- Kahnt, T., Grueschow, M., Speck, O., & Haynes, J. D. (2011). Perceptual learning and based learning explains visual perceptual learning and its specificity and decision-making in human medial frontal cortex. Neuron, 70(3), 549–559. transfer. Journal of Neuroscience, 30(37), 12323–12328. Karni, A., & Sagi, D. (1991). Where practice makes perfect in texture discrimination: Zhaoping, L., Herzog, M. H., & Dayan, P. (2003). Nonlinear ideal observation and Evidence for primary visual cortex plasticity. Proceedings of the National recurrent preprocessing in perceptual learning. Network, 14(2), 233–247. Academy of Sciences of the United States of America, 88(11), 4966–4970.