1

Superior identification of familiar visual patterns a year after learning Zahra Hussain1, Allison B. Sekuler2,3 & Patrick J. Bennett2,3 1. School of Psychology, University of Nottingham, Nottingham, England, NG72RD 2. Department of Psychology, Neuroscience & Behaviour, McMaster University, Hamilton, ON, Canada, L8S 4K1 3. Center for Vision Research, York University, Toronto, ON Canada These authors contributed equally to this work. Practice improves visual performance on simple tasks in which stimuli vary along one dimension (e.g., Fiorentini & Berardi, 1981). Such learning frequently is stimulus-specific and enduring, and has been associated with plasticity in striate corext (e.g., Furmanski, Schluppeck & Engel, 2004). It is unclear if similar lasting effects occur for naturalistic patterns that vary on multiple dimensions. We measured perceptual learning in identification tasks that used faces and textures, stimuli that engage multiple stages in visual processing. Performance improved significantly across two consecutive days of practice. More importantly, the effects of practice were remarkably stable across time: improvements were maintained approximately one year later, and the relative difficulty of identifying individual stimuli, as well as individual differences were essentially constant across sessions. Finally, the effects of practice were largely stimulus specific. Our results suggest that the characteristics of perceptual learning are similar across a spectrum of stimulus complexities.

2

Real objects comprise multiple features, and tasks like identification generally require observers to utilize information carried by more than one stimulus dimension. As with simple tasks, performance in visual tasks using complex stimuli improves with practice (Kolers 1976; Sireteanu & Rettenbach, 1995; Gold, Bennett & Sekuler, 1999; Hussain, Bennett & Sekuler, 2009a, 2009b; Lobley & Walsh, 1998; Tseng, Gobell & Sperling, 2004; Tanaka, Curran & Sheinberg, 2005), indicating that there is plasticity at later stages in visual processing that combine the outputs of earlier areas. Do these effects of perceptual learning in tasks using complex stimuli persist for long periods of time? In certain cases the improvements do not last (Lobley & Walsh, 1998), or lasting improvements are not specific to the stimuli or tasks used during training (Sireteanu & Rettenbach, 1995; Tanaka et al, 2005), suggesting that there is greater long-term retention of more general perceptual operations. Such learning may form the basis of visual expertise in categorizing patterns and objects, which generalizes to novel instances of familiar categories (Tanaka et al, 2005). However, long-lasting itemspecific effects have been reported in implicit memory for complex items such as words and pictures (Sloman, Hayman, Ohta, Law & Tulving, 1988; Cave, 1997; deSchepper & Triesman, 1996), and in the motor domain for a serial reaction-time task (Romano, Howard & Howard, 2010). These long-term effects obtained across paradigms suggest that the visual system might also show resilient and specific learning of complex input, as suggested by an early report of content-specific retention of improvements in reading mirror-transformed text (Kolers, 1976).

Performance in a 10-AFC face and texture identification task improves substantially with practice (Gold et al, 1999): a noisy stimulus is presented briefly at one of several

3

contrasts, followed by a choice of ten noiseless exemplars, one of which is the correct response (Figure 1). Improvements in identification accuracy, both with faces and textures, are largely stimulus-specific: there is little transfer to novel items that share the spatial characteristics of the trained set, to the trained items rotated by 180 degrees (Hussain et al, 2009a, 2009b), and, in the case of textures, to contrast-reversed versions of the trained items (Hussain et al, 2009a). Here, we ask whether this type of stimulusspecificity endures for long periods. Unlike previous studies of item-specific memory of objects and words (Sloman et al, 1998; Cave, 1997), or of basic attributes of simple visual stimuli (Fiorentini & Berardi, 1981; Karni & Sagi, 1993; Ball & Sekuler, 1982; Adini, Sagi & Tsodyks, 2004), the present task addresses learning of complex stimuli at the perceptual level, both for items that are highly familiar (faces), and unfamiliar patterns with no semantic content (textures). Long lasting, stimulus-specific learning of face - and texture identification, if found, would suggest that the characteristics of perceptual learning are similar across a range of stimulus complexities, and also resemble learning in higher-level cognition. METHOD Nine subjects performed a face-identification task, and six subjects performed a texture-identification task on two occasions: an initial learning and test phase (Days 1 & 2) in experiments that were conducted with a larger sample of observers, and a followup test 10-18 months later. On average, subjects in the face- and texture identification task performed the follow-up session respectively 13 months (SD = 3.4) and 15 months (SD =1.6) after the initial test session. The apparatus, stimuli and procedure have been described in detail previously (Gold et al, 1999; Hussain et al, 2009a, 2009b). Briefly, faces were cropped to show

4

only internal features and were equated in terms of their global amplitude spectrum. The textures were band-limited noise patterns. The stimuli subtended 4.8 x 4.8 degrees of visual angle from the viewing distance of 114 cm, and were presented in low, medium, or high levels of static two-dimensional Gaussian noise. Stimulus contrast was varied at one of seven levels (per noise condition), using the method of constant stimuli. On each of Days 1 and 2, observers performed 840 trials per day, with one stimulus set within a given stimulus class (faces or textures). The follow-up session consisted of 420 trials with the trained stimulus set, and 420 trials with a novel stimulus set for the same stimulus class. The same and novel sets of stimuli were presented in separate blocks of trials and the order was roughly counterbalanced across subjects. For analyses, the 840 trials on Days 1 and 2 were divided into four bins of 210 sequential trials (Trial bins 1-8), and the 420 trials for each of the stimulus sets during follow-up were divided into two bins of 210 trials (Bins A and B). For each bin, the proportion of correct responses was calculated after collapsing across all levels of stimulus contrasts and noise.

RESULTS Performance with faces and textures was assessed separately. For each type of stimulus, learning during the initial session was assessed with three planned comparisons: Bin 1 vs. Bin 4, Bin 5 vs. Bin 8, and the average of Day 1 vs. the average of Day 2. Retention of learning was assessed with eight planned comparisons: familiar stimuli in Bin A vs. Bins 1, 4 and 8; novel stimuli in Bin A vs. Bins 1, 4, and 8; and familiar vs. novel stimuli in Bins A and B. Hence, there were 11 comparisons for each type of stimulus. Bins 1, 4, and 8 were chosen because they represent pre-learning baseline (Bin 1), postlearning performance at the end of Day 1 (Bin 4), and post-learning performance at the

5

end of both days (Bin 8). For each set of 11 comparisons, Holm’s sequential Bonferroni test (Kirk, 1995) was used to maintain a familywise Type I error rate of 0.05. Significant p values are indicated by asterisks.

Face identification accuracy showed significant within-session improvement on Day 1, as evidenced by a 20% increase in accuracy from Bin 1 to Bin 4 (t(8) = 9.54, p < .00001*; Figure 2 top). There also was significant within-session improvement on Day 2, with performance improving by 10% from Bin 5 to Bin 8 (t(8) = 4.54, p < .01*). Overall accuracy averaged across bins was 18% higher on Day 2 than on Day 1, (t(8) = 11.36, p < .00001*). Approximately one year later, accuracy in the first trial bin (Bin A) in the same-face condition was 18% higher than in Bin 1 (t(8) = 7.883, p < .0001*), did not differ from accuracy in Bin 4 (t(8) = 1.054, p = .322), and was 15% lower than accuracy in Bin 8 (t(8) = 7.17, p < .0001*). Therefore, performance in the follow-up session in the same-face condition was equivalent to that achieved at the end of the first, but not the second, session. In the novel-face condition, accuracy in Bin A was no different than accuracy in Bin 1 (t(8) = 1.8, p = .10), 15% lower than accuracy in Bin 4 (t(8) = 4.05, p = .0036*), and 29% lower than accuracy in Bin 8 (t(8) = 8.17, p < .0001*). Finally, accuracy in the same-face condition was higher than accuracy in the novel-face condition in Bin A (difference = 14%; t(8) = 3.57, p = .007*) and Bin B (difference = 20%; t(8) = 4.81, p = .001*). These analyses confirm that response accuracy in the follow-up session, held approximately one year after training, was higher for faces that were seen in the original training sessions. Texture identification accuracy (Figure 2 bottom), improved significantly on Day 1 (25% from Bin 1 to Bin 4, t(5) = 5.070, p = .003*), but the improvement on Day 2 was

6

not significant (16% from Bin 5 to Bin 8, t(5) = 2.96, p = .03). Averaged across bins, accuracy on Day 2 was 21% higher than on Day 1 (t(5) = 8.18, p = .0004*). Approximately one year later, accuracy in Bin A in the same-texture condition was 35% higher than in Bin 1 (t(5) = 15.70, p < .0001*), but did not differ significantly from accuracy in Bin 4 (t(5) = 2.68, p = .04) or Bin 8 (t(5) = 1.54, p = .18). In the noveltexture condition, accuracy in Bin A was 18% better than in Bin 1 (t(5) = 6.41, p = .001*), no different than in Bin 4 (t(5) = 1.87, p = .11), and 24% lower than in Bin 8 (t(5) = 4.88, p = .004*). Performance in the same-texture condition was significantly better than in the novel-texture condition in Bin A (16% difference; t(5) = 6.44, p = .0013*) but not in Bin B (10% difference; t(5) = 3.29, p = .02), indicating that a significant proportion of the improvement was stimulus-specific.

We conducted several tests to compare retention measured with faces and textures. The Holm’s sequential Bonferroni test (Kirk, 1995) was again used to maintain a familywise error rate of 0.05. First, we examined if the amount of transfer to novel items differed for faces and textures by comparing the difference between Bin A and Bin 1in the novel-texture and novel-face conditions. There was a significant difference in the amount of transfer to novel items obtained with faces (4%) and textures (18%) (t(10.873) = 3.79, p = .003*). However, the advantage of same over novel items did not differ between faces and textures in Bin A (faces (19%) vs. textures (11%); t(12.637) = .55, p = .58) or in Bin B (faces (19%) vs. textures (11%), t(12.998) = 1.69, p = .11). Thus, although there was more generalization to novel items in the texture identification task, the same-stimulus advantage did not differ between faces and textures.

7

Despite the large variability in performance across observers, the rank order of individuals was stable across sessions in both tasks (Figure 3). For subjects who were tested with faces, the Spearman rank-order correlation was significant for all pairs of test sessions (rho ! 0.96, p < .001). For subjects who were tested with textures, the rank-order correlations were all positive, but, perhaps due to the small sample size (n = 6), only the correlation between Day 1 and the follow-up session was significant (Days 1 & 2: rho = 0.60, p = .24; Day 1 & follow-up: rho= .94, p = .015; Day 2 & follow-up: rho = .71, p = .13). To examine whether the effects of practice reflected improved performance for all items or just a few items, we calculated the mean accuracy for individual faces and textures during each test session (Figure 4). Each symbol in Figure 4a and 4b represents response accuracy for a single face averaged across subjects. There were clear differences in the initial difficulty in identification across individual items. However, regardless of initial difficulty, accuracy relative to Day 1 was higher for 29 out of 30 faces on Day 2 (Figure 4a), and 27 out of 30 faces during follow-up (Figure 4b). The accuracies for individual faces measured during different sessions were highly correlated (all r's ! .90, p < .001), implying that learning did not change the relative discriminability of the faces. Figures 4c and 4d show the effects of practice on accuracy for individual textures. Again, response accuracy was higher for every texture on Day 2 than on Day 1 (Figure 4c) and for 19 out of 20 textures during follow-up (Figure 4d). As was found with faces, identification accuracy for each texture remained relatively stable across all

8

sessions (all r's ! .71, p < .001), which suggests that learning (or forgetting) did not produce qualitative changes in the way the items were represented.

DISCUSSION Perceptual learning of face- and texture identification was remarkably stable and stimulus-specific 10-18 months after practice. Some of the long-term benefits of learning generalized to novel stimuli, but the degree of stimulus specificity was similar to what has been obtained in experiments using a retention interval of just one day (Hussain et al, 2009a, 2009b). Individual differences in response accuracy were stable across testing sessions, consistent with research showing that practice does not eliminate, but rather crystallizes, initial performance differences for certain skilled tasks (deSchepper & Triesman, 1996; Ackerman, 2007). Finally, despite significant variation across stimuli in the initial level identification accuracy, practice increased performance nearly uniformly for all items, indicating that the relative discriminability of the stimuli was unaltered after learning.

The textures used in our experiment were unfamiliar stimuli that lacked the spatial regularities and semantic content that exist in faces. Nevertheless, accuracy a year later was much higher for familiar than novel textures, indicating that familiarity with the object class prior to training was not essential for long-term retention. Analogous learning in the auditory system has been reported recently for noise stimuli, with the effects of previous exposure lasting up to 3 weeks (Agus, Thorpe & Pressnitzer, 2010). The enduring stimulus specificity of learning found with faces also is notable given the

9

exposure to other faces that presumably occurred for all participants in natural contexts during the intervening period. Evidently, exposure to other faces did not diminish the specific effects of perceptual learning up to a year later. The primary goal of this study, however, was not to compare the extent and duration of learning across stimulus classes, so one cannot conclude, for example, that unfamiliar or uncommon stimuli lead to longer lasting learning than familiar or common stimuli. Future studies could assess this issue more explicitly by controlling factors such as stimulus similarity and initial task difficulty directly. Long-lasting, stimulus-specific effects of learning on simple visual discriminations have been attributed to changes early in the visual pathway, particularly to the primary visual area (Furmanski et al, 2005; Hua et al, 2010; Yotsumoto, Watanabe & Sasaki, 2008). However, it is likely that performance in our tasks depends in part on a more distributed network of neurons, including those in inferotemporal cortex (IT), which show stimulus-selective activation by stimuli such as faces and other complex patterns (Logothetis, Pauls & Poggio, 1995; Leopold, Bondar & Giese, 2006), and the selectivity of which can be altered by practice (Li & DiCarlo, 2008). Other work has shown a role for IT neurons in the acquisition of complex visual skills (Poldrack, Desmond, Glover & Gabrieli, 1998), in the formation of visual memory (Desimone, 1996) and in longterm visual priming (Meister et al, 2005), which suggests common neural substrates for long-lasting perceptual learning and visual memory. Learning even in simple conditions is thought to activate a broad network that encompasses task context among other stimulus parameters (Yotsumoto et al, 2008; Logothetis et al, 1995; Meister et al, 2005; Xiao et al, 2008). Therefore, as with the

10

short-term learning effects, it is plausible that long-lasting effects of practice are accompanied by changes in a broad cortical network that includes several levels of the visual hierarchy. Although the underlying mechanisms may differ, the long-lasting and stimulus-specific perceptual learning reported here resembles durable effects found with in implicit memory (Sloman et al, 1988; Cave, 1997; deSchepper & Triesman, 1996; Romano et al, 2010), and certain types of sensory adaptation (Jones & Holding, 1975).

References Ackerman, P. L. (2007). New developments in understanding skilled performance. Current Directions in Psychological Sciences. 16(5), 235-239.

Adini, Y., Sagi, D., & Tsodyks, M. (2002). Context-enabled learning in the human visual system. Nature, 415, 790-793.

Agus, T. R., Thorpe, S. J., & Pressnitzer, D. (2010) Rapid formation of robust auditory memories: Insights from noise. Neuron, 66, 610-618.

Ball, K., & Sekuler, R. (1982). A specific and enduring improvement in visual motion discrimination. Science, 218, 697-698.

Cave, C. B. (1997). Very long-lasting priming in picture naming. Psychological Science, 8(4), 322-325. (1997).

deSchepper, B., & Triesman, A. (1996). Visual memory for novel shapes: Implicit

11

coding without attention. Journal of Experimental Psychology: Learning, Memory & Cognition. 22(1), 22-47.

Desimone, R. (1996). Neural mechanisms for visual memory and their role in attention. Proceedings of the National Academy of Sciences-USA, 93, 13494-13499.

Fiorentini, A., & Berardi, N. (1981). Learning in waveform discrimination: specificity for orientation and spatial frequency. Vision Research, 27, 1149-1158.

Furmanski, C. S., Schluppeck, D., & Engel, S. A. (2004). Learning strengthens the response of primary visual cortex to simple patterns. Current Biology, 14, 573576.

Gold, J., Bennett, P. J., & Sekuler, A. B. (1999). Signal but not noise changes in perceptual learning. Nature, 402, 176-178.

Hua, T. M., Bao, P. L., Huang, C. B., Wang, Z. H., Xu, J. W., Zhou, Y. F., & Lu, Z. L. (2010). Perceptual learning improves contrast sensitivity of V1 neurons in cats. Current Biology, 20(10), 887-894.

Hussain, Z., Sekuler, A. B., & Bennett, P. J. (2009a). Contrast-reversal abolishes Perceptual learning. Journal of Vision. 9(4), 1-8.

Hussain, Z., Sekuler, A. B., & Bennett, P. J. (2009b). Perceptual learning modifies

12

inversion effects for faces and textures, Vision Research, 49, 2273-2284.

Jones, P. D., & Holding, D. H. (1975). Extremely long-term persistence of the McCollough Effect. Journal of Experimental Psychology: Human Perception and Performance, 1(4), 322-327.

Karni, A., & Sagi, D. (1993). The time course of learning a visual skill. Nature, 365, 250- 252.

Kirk, R. E. (1995). Experimental design: Procedures for the behavioural sciences (3rd Ed.). Pacific Grove, California: Brooks/Cole.

Kolers, P. (1976). Pattern-analyzing memory. Science, 191, 1280-1281.

Leopold, D. A., Bondar, I. V., & Giese, M. A. (2006). Norm-based face encoding by single neurons in the monkey inferotemporal cortex. Nature, 442(3), 572-575.

Li, N., & DiCarlo, J. (2008). Unsupervised natural experience rapidly alters invariant object representation in visual cortex. Science, 321, 1502-1506.

Lobley, K., & Walsh, W. (1998). Perceptual learning in visual conjunction search. Perception, 27, 1245-1255.

Logothetis, N., Pauls, J., & Poggio, T. (1995). Shape representation in the inferior

13

temporal cortex of monkeys. Current Biology, 5, 552-563.

Meister, I. G., Weidemann, J., Foltys, H., Brand, H., Wilmes, K., Krings, T., Thron, A., Topper, R & Boroojerdi, B. (2005). The neural correlate of very-long-term picture priming. European Journal of Neuroscience. 21, 1101-1106.

Poldrack, R. A., Desmond, J. E., Glover, G. H. & Gabrieli, J. D. E. (1998). The neural basis of visual skill learning: An fMRI study of mirror reading. Cerebral Cortex, 8, 1-10.

Romano, J. C., Howard Jr. J. H., & Howard, D. V. (2010). One-year retention of general and sequence-specific skills in a probabilistic, serial reaction time task. Memory, 18(4), 427-441.

Sireteanu, R., & Rettenbach, R. (1995). Perceptual learning in visual search: Fast, enduring but non-specific. Vision Research, 35, 2037-2043.

Sloman, S. A., Hayman, C. A., Ohta, N., Law, J. & Tulving, E. (1988). Forgetting in primed fragment completion. Journal of Experimental Psychology: Learning, Memory & Cognition. 14(2), 223-239.

Tanaka, J. W., Curran, T & Sheinberg, D. L. (2005). The training and transfer of realworld perceptual expertise. Psychological Science, 16(2),145-151.

14

Tseng, C., Gobell, J. L. & Sperling, G. (2004). Long-lasting sensitization to a given colour after visual search. Nature, 428, 657-660.

Xiao, L. Q., Zhang, J. Y., Wang, R., Klein, S. A., Levi, D. M., & Yu, C. (2008). Complete transfer of perceptual learning across retinal locations enabled by double training. Current Biology, 18(24). 1922-1926.

Yotsumoto, Y., Watanabe, T., & Sasaki, Y. (2008). Different dynamics of performance and brain activation in the time course of perceptual learning. Neuron, 57, 827833.

15

Acknowledgements

We thank Donna Waxman for her assistance in running subjects. This research was funded by NSERC and the Canada Research Chair program. The authors declare no competing interests.

Correspondence and requests for materials should be addressed to Zahra Hussain (e-mail: [email protected]).

Figure captions Figure 1. Examples of the stimuli and a schematic representation of the trial sequence.

16

Figure 2. Proportion correct on Day 1, Day 2, and approximately 1 year (10-18 months) later for 9 subjects in the face identification task (top), and 6 subjects in the texture identification task (bottom). Each session comprised 840 trials, therefore each trial bin represents 210 trials. Performance a year later was measured both with the same items shown on Days 1 and 2 (filled symbols), and with novel items that subjects had not seen before (open symbols). Figure 3: Performance of individual subjects on the face and texture identification tasks on Day 1, Day 2, and during the follow-up session about one year later. Each symbol represents a subject. Different subjects performed the face- and texture identification tasks. Figure 4. Scatter plots showing accuracy for three sets of ten faces used in separate experiments in the face identification task (a, b), and two sets of ten textures used in the texture identification task. (c, d). Each point is based on the average of 2-6 subjects. Area above the solid line indicates improvements. Dashed line indicates the least squares fit. Face identification: each symbol shows accuracy averaged across four (black circles), three (stars), and two (open circles) subjects. a, b) Day 1 vs. Day 2. c, d) Day 1 vs. Follow-up. All correlations are positive and significant. Learning is retained a year later relative to performance on Day 1 (c,d).

fixate Fixate 100ms 100ms

ISI ISI 100ms 100ms

ee titm im

ISI ISI 100ms 100ms

Stimulus 200ms 200ms stimulus

thumbnails Thumbnails until untilresponse response

Day 1

Faces

Day 2

1 yr later

0.8

0.6

0.4

same novel 0.2

Textures 0.6

0.4

0.2

same novel 1

2

3

4

5

Bin

6

7

8

A

B

0.8

0.6

0.4

0.2

Face identification

Day 1

Day 2

1 yr later

Texture identification

Day 1

Day 2

1 yr later

a

b 1.0

Faces

0.8

Accuracy 1 yr later

Day 2 accuracy

1.0

0.6

0.4

0.2

Faces

0.8

0.6

0.4

0.2 r < 0.90, p < .0001 0.2

0.4

0.6

0.8

r < 0.92, p < .0001

1.0

0.2

Day 1 accuracy 1.0

0.6

0.8

1.0

Day 1 accuracy

d 1.0

Textures

Accuracy 1 yr later

Day 2 accuracy

c

0.4

0.8

0.6

0.4

Textures

0.8

0.6

0.4

0.2

0.2 r < 0.89, p < .0001 0.2

0.4

0.6

0.8

Day 1 accuracy

1.0

r < 0.71, p < .0001 0.2

0.4

0.6

Day 1 accuracy

0.8

1.0