Perceptual learning in Vision Research

Vision Research 51 (2011) 1552–1566

Contents lists available at ScienceDirect

Vision Research journal homepage: www.elsevier.com/locate/visres

Review

Perceptual learning in Vision Research Dov Sagi ⇑ The Weizmann Institute of Science, Rehovot 76100, Israel

a r t i c l e

i n f o

Article history: Received 24 April 2010 Received in revised form 15 October 2010 Available online 23 October 2010 Keywords: Associative learning Memory consolidation Neural networks Neural rehabilitation Neuronal plasticity Sensory adaptation Learning specificity Statistical modeling Overfitting Visual cortex

a b s t r a c t Reports published in Vision Research during the late years of the 20th century described surprising effects of long-term sensitivity improvement with some basic visual tasks as a result of training. These improvements, found in adult human observers, were highly specific to simple visual features, such as location in the visual field, spatial-frequency, local and global orientation, and in some cases even the eye of origin. The results were interpreted as arising from the plasticity of sensory brain regions that display those features of specificity within their constituting neuronal subpopulations. A new view of the visual cortex has emerged, according to which a degree of plasticity is retained at adult age, allowing flexibility in acquiring new visual skills when the need arises. Although this ‘‘sensory plasticity’’ interpretation is often questioned, it is commonly believed that learning has access to detailed low-level visual representations residing within the visual cortex. More recent studies during the last decade revealed the conditions needed for learning and the conditions under which learning can be generalized across stimuli and tasks. The results are consistent with an account of perceptual learning according to which visual processing is remodeled by the brain, utilizing sensory information acquired during task performance. The stability of the visual system is viewed as an adaptation to a stable environment and instances of perceptual learning as a reaction of the brain to abrupt changes in the environment. Training on a restricted stimulus set may lead to perceptual overfitting and over-specificity. The systemic methodology developed for perceptual learning, and the accumulated knowledge, allows us to explore issues related to learning and memory in general, such as learning rules, reinforcement, memory consolidation, and neural rehabilitation. A persistent open question is the neuro-anatomical substrate underlying these learning effects. Ó 2010 Elsevier Ltd. All rights reserved.

1. Historical background It is well known that performance on perceptual tasks is affected by practice, a fact acknowledged by many authors in Vision Research since its first volume was published. However, the neuronal mechanisms underlying these effects were not considered in these few early writings, probably due to the prevailing direct view of perception. According to this view, it was assumed that perceptual learning involves an improved selection of the information available in the world that is relevant for the task (Gibson, 1969). Thus, these earlier studies frequently mentioned that practice trials were given to the experimental observers to familiarize them with the task, but only a small number of reports documented the resulting effects. Ekman and Lindman (1962) measured fluctuations in the perception of light intensity (i.e. internal noise; Thurstone, 1927) and found that the reported fluctuations were markedly affected by practice. Anstis (1970), describing results from apparent motion experiments, mentioned that ‘‘In addition, some kind of perceptual learning took place. Practice at observing made the phi more readily visible, and experienced observers could ⇑ Fax: +972 (8) 934 4131. E-mail address: [email protected] 0042-6989/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.visres.2010.10.019

see the phi some time before naive observers. This parallels the build-up reported by Julesz in perceiving stereo in the random-dot patterns.’’ Indeed, a successful fusion of random-dot stereograms (Julesz, 1971) requires some practice, often to improve the control of vergence eye movement. Ramachandran & Braddick (1973) replaced the dots by small oriented line segments and found the learning effect to be orientation selective, suggesting that ‘‘the stereoscopic skill that has been acquired may be specific to those orientation analysers that were stimulated during the training period.’’ This was probably the first suggested link between perceptual learning and specific neuronal mechanisms within the visual cortex. The late 1970s and early 1980s were characterized by a paradigm shift toward neuronal-based accounts of learning, in line with the newly emerging understanding of the visual cortex (Barlow, 1972; Campbell & Robson, 1968; Hubel & Wiesel, 1962). In a 1977 Vision Research paper, Karen De Valois (1977), who studied the effects of adapting gratings, reported that practice had large effects on contrast sensitivity. According to her results, there was a 10-fold increase in contrast sensitivity with almost daily practice extending over a 1-year period. Most interestingly, the spatial-frequency bandwidth of the adaptation effect was reduced with practice and the amplitude of adaptation increased. Two possible explanations were provided: (1) ‘‘. . .with increasing practice, the

D. Sagi / Vision Research 51 (2011) 1552–1566

subject simply becomes more efficient at restricting the sample pool to these cells which are most sensitive to the [spatial] frequency being observed’’, or (2) ‘‘. . .change over time in the sensitivity profiles of the detectors involved, or perhaps the establishment of new connections between existing units’’. These earlier studies employed measures of sensitivity subjected to decisional bias, such as contrast adjustment, leaving open the possibility of effects due to changes in the subjective criterion. In a 1983 Vision Research paper, Fendick and Westheimer (1983), using objective measures of sensitivity, reported improvement in stereo acuity, but they did not consider any specific mechanism, probably keeping in line with an earlier report by McKee and Westheimer (1978) on practice effects in hyper-acuity, attributing improvement to ‘‘a kind of fine tuning of whatever neural mechanism is responsible for sensing these differences’’. These studies validated that perceptual learning involves improvement in sensitivity (d0 ) of basic (low level) visual tasks, rather than a change in bias. Moreover, acuity thresholds at the hyper-acuity level, achieved in these experiments with a variety of visual tasks (Fig. 1b), is thought to require access to spatial information at a high-resolution present only within neuronal representations residing in the primary visual cortex (Poggio, Fahle, & Edelman, 1992). The first article in Vision Research, dedicated to perceptual learning, was published in 1981 by Fiorentini and Berardi. These experiments, employing an objective 2-Alternative-Forced-Choice (2AFC) method, showed a fast learning effect, developed over 200 trials, related to phase discrimination in compound gratings (Fig. 1a). Fiorentini and Berardi (1981) found that the learning effects were selective for stimulus orientation, spatial-frequency, and retinal location, and that transfer of learning occurred only between stimuli considered to be within the bandwidth of a single spatial filter in early vision (Blakemore & Campbell, 1969). Their results, however, did not show learning with pure sine wave gratings, using a task of spatial-frequency discrimination. Thus, they thought that learning involves modification at a level that integrates the output of orientation-selective neurons in the visual cortex, but still maintains specificity for low-level features. The finding that the learning effect was retained after a few days, and partially after 7 months, made the result of particular interest, hinting at some type of plasticity in the visual system at some intermediate stage of processing where the outputs of early visual filters are combined. Fiorentini and Berardi (1981) concluded that ‘‘whatever the neural networks underlying learning process, it is impossible to escape from the conclusion that it has to occur at a relatively early stage in the sequence of transformations leading to form perception’’. In another experiment published in Vision Research the same year by Melanie Mayer (1983), observers practiced contrast detection with gratings, 3000 trials over 3 weeks; this resulted in reduced orientation anisotropy after practice. The observers improved with diagonal orientations, but not with horizontal and vertical gratings, exhibiting orientation selectivity. Ball and Sekuler (1987) were the next to publish a study of perceptual learning in Vision Research, reporting long-term improvement in motion discrimination that was direction specific and that was characterized by only a partial transfer between eyes. They specifically indicated that area MT was a possible cortical site for the learning effect that took place. These few studies, published in Vision Research in the 1980s, some of which were accompanied by shorter reports in Nature (Fiorentini & Berardi, 1980) and Science (Ball & Sekuler, 1982), laid the foundation for the current studies of perceptual learning. Clearly, these studies were stimulated by the emerging new understanding of early vision as consisting of specific serial transformations implemented via neuronal units having well-defined features. However, unlike the prevailing dogma stating a finite period of sensory development (critical periods), the discovered

1553

practice effects pointed to effective plasticity in the mature visual system, and suggested that sensory plasticity in vision extends much beyond the assumed critical period into adulthood. The basic questions, still without satisfactory answers, were put forward: which specific brain sites are modified by learning, and more specifically, does learning involve rewiring of neurons in early visual areas, or can it all be explained by improved efficiency in the readout of unchanged early neuronal representations? Later on many more reports showed that learning takes place in a large variety of basic visual tasks, some of which were found to be highly specific. The most salient issue discussed by these works, in attempting to understand the levels of processing affected by learning, concerns the mapping to physiology, that is, where learning actually takes place in the brain. Two important heuristics were found useful regarding this issue: (1) different tasks rely on different visual areas. Thus, to the extent that mapping exists between a brain region and a task, learning corresponds to plasticity at that specific cortical area subserving the probed behavior; and (2) neurons in the visual system respond to a limited set of stimuli; thus, specificity of learning is expected to reflect the specificities found in the neural responses underlying learning. New findings within the neurosciences, based on methodologies developed in recent years, opened the way to exploring the mechanisms underlying learning, such as consolidating the temporary internal response to a stimulus into long-lasting functional changes, and the dependency of learning on non-sensory aspects, such as the actual task used, response feedback, stimulus uncertainties, and attention. The rich outcome, the result of carefully developed scientific methodologies, allows us to portray a consistent view of perceptual learning with the aim of generalizing other domains of learning where reliable experimental results are more difficult to obtain. This vast body of research is reviewed in an excellent book by Fahle and Poggio (2002), in several review papers covering a large range of topics related to learning, plasticity and sensory development (Goldstone, 1998; Holtmaat & Svoboda, 2009; Huxlin, 2008; Morishita & Hensch, 2008; Sale, Berardi, & Maffei, 2009; Sasaki, Nanez, & Watanabe, 2010; Tsodyks & Gilbert, 2004; Wandell & Smirnakis, 2009), and in special issues published in recent years, including Vision Research (Lu, Yu, Watanabe, Sagi, & Levi, 2009; Lu, Yu, Watanabe, Sagi, & Levi, 2010), Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences (McGraw, Webb, & Moore, 2009), Topics in Cognitive Science (Jacobs, 2010), and Learning & Perception (Sagi, Kovács, & Racsmany, 2009). The present review is aimed at providing an integrated view of the most recent issues discussed in the literature.

2. Limitations – what cannot be learned Perceptual learning is seen in many visual tasks (Fahle & Poggio, 2002), and almost on every occasion where an observer’s vision is challenged. But first I would like to note some basic tasks in which no learning takes place. As mentioned above, Fiorentini and Berardi (1981) reported no learning in a task where observers discriminated between two similar spatial-frequency gratings. This is rather surprising, given our knowledge of early visual processes, according to which an initial filtering stage exists consisting of spatial-frequency selective filters (Daugman, 1985; Wilson & Bergen, 1979) that obey the concept of ‘‘labeled lines’’ (Watson & Robson, 1981). These filters are affected by intensive exposure to high contrast, exhibiting short-term effects on a time scale of a few seconds, termed ‘‘contrast adaptation’’ (Greenlee, Georgeson, Magnussen, & Harris, 1991). The effects of adaptation are selective for spatialfrequency and orientation and are used to define filter properties (Blakemore & Campbell, 1969; Pantle & Sekuler, 1968). Spatialfrequency discrimination is affected by contrast adaptation,

1554

D. Sagi / Vision Research 51 (2011) 1552–1566

a

b d

c

e a

b

g

f

B

A

h

(b)

Fig. 1. Some visual stimuli used in studies of perceptual learning. (a) Two sine wave gratings having spatial frequencies f and 3f, with differing contrasts of the higher frequency between the two gratings (Fiorentini & Berardi, 1981). (b) Stimuli used in spatial acuity tasks (Poggio et al., 1992). (c) Task-irrelevant learning (TIPL): processing of the peripheral motion stimulus is improved while observers perform a letter identification task at fixation (Watanabe et al., 2001). (d) Texture and face stimuli used in identification tasks (Husk, Bennett, & Sekuler, 2007). (e) Contrast detection in external noise (Dosher & Lu, 1999). (f) Gabor stimuli used in contrast discrimination tasks, with (top) and without (bottom) spatial context (Adini et al., 2002). (g) Texture stimuli with a target frame in A, and a mask in B (Karni & Sagi, 1991). (h) Stimuli used in lateralfacilitation experiments (Polat & Sagi, 1994b) and in training amblyopic observers (Polat et al., 2004). (i) Stimuli used to measure contour integration with contour saliency decreasing from A to C (Li et al., 2008b).

D. Sagi / Vision Research 51 (2011) 1552–1566

1555

i

Fig. 1 (continued)

though not at the adapting frequency, but rather at frequencies removed by 1 octave from it (Regan & Beverley, 1985). Given that the filter bandwidth is two octaves, such a shift is expected from a system that optimally weighs filter responses to compute a frequency discriminator (Regan & Beverley, 1985). Thus, the efficiency of discrimination is expected to improve with practice if perceptual learning is capable of filter reweighting (e.g. (Dosher & Lu, 1999)). Practicing contrast detection with Gabor patches or with grating stimuli yields only slight improvements (Dosher & Lu, 2005; Sowden, Rose, & Davies, 2002) or none (Mayer, 1983; Dorais & Sagi, 1997; Maehara & Goryo, 2007). However, the possibility exists that improvement in such basic visual skills is very rapid, with the number of trials being too small to permit measurements. Or perhaps, on the other hand, such improvement would require extensive long-term learning, as indicated by results showing improved contrast sensitivity and integration time in video gamers (Li, Polat, Makous, & Bavelier, 2009). Similarly, the ability to discriminate between two contrast levels of otherwise identical grating patches does not improve with practice (Adini, Sagi, & Tsodyks, 2002; Dorais & Sagi, 1997; Maehara & Goryo, 2007). Landolt C acuity and two-line resolution thresholds, unlike Vernier, do not improve much with practice, if at all, possibly because these discriminations can be reduced to a task involving detection of an intensity change (Westheimer, 2001). Westheimer argued that ‘‘where there is no learning the processing is of a more primitive kind, more robust and nearer to sensory origins’’. These results indicate that our visual system shows expertise in some basic tasks but they cannot be further developed. Such expertise may vary across individuals; thus, some learning is possible with observers having particularly high initial thresholds. Fahle and Henke-Fahle (1996) reported large inter-observer variability in the learning of a hyper-acuity task, which was accounted for by the initial variability among observers. Initially observers had very different displacement thresholds (a 3-fold range) but ended up with very similar thresholds, equaling those that were initially better. Similar effects were seen in the contrast-detection data of Dosher and Lu (2005). This observation highlights the difficulty in obtaining reliable measurements at an initial performance level, but it is also a source of optimism for learning, suggesting possible applications for sensory rehabilitation (see below: critical period). Some visual tasks, however, cannot be learned when stimulus uncertainty is introduced to the task, but these tasks can be learned when uncertainty concerning the stimulus is reduced or, even better, eliminated. Swift and Smith (1983) reported a dramatic decrease in contrast detection thresholds for targets presented in fixed additive noise (a stimulus constructed by the addition of several spatial-frequency gratings), with the dependency of contrast-discrimination thresholds on noise contrast changing from Weber’s law to power law during training. However, learning was not observed when the noise was renewed with each trial. Stimulus uncertainty was also shown to affect learning

of contrast discrimination (as in the experiments mentioned above). More specifically, learning was possible when the stimulus set was reduced to a narrow range of contrasts (Adini, Wilkonsky, Haspel, Tsodyks, & Sagi, 2004; Yu, Klein, & Levi, 2004), or when different contrasts were grouped (Zhang et al., 2008), but not under conditions of complete contrast uncertainty (Adini et al., 2004; Yu et al., 2004). Perhaps practicing with a narrow range of stimuli, or a single stimulus, permits learning unique features associated with a specific task, which are not useful when the predictability of the stimulus is reduced (Adini et al., 2004). This issue will be further considered below as ‘‘perceptual modeling and overfitting’’. Yet another limitation arises from intrinsic limitations in the visual system, which do not allow for adequate processing of some encountered stimuli. One example is the detection of a conjunction target in a search task. The search time for such a task increases linearly with the number of non-target display elements (serial search), unlike the detection of a target defined by a unique feature that is independent of the number of non-targets (parallel search) (Treisman & Gelade, 1980). In Sireteanu & Rettenbach’s study (2000), observers practiced a search task for 11 days, with the target defined by conjunction of color and orientation. Surprisingly, they found little improvement in search efficiency, and the serial behavior was preserved. Thus, although some pairs of features are conjointly encoded by the visual system, such as spatial-frequency and orientation (Sagi, 1988), others, such as color and orientation, are not, and their conjunctions cannot be learned at the early processing level where the machinery devoted to parallel searching seems to reside.

3. Low-level components There is an ongoing intensive debate concerning the plasticity of the primary visual cortex in adults (Wandell & Smirnakis, 2009), and its role in perceptual learning. Within the framework of psychophysics, stimulus specificity is viewed as the main indicator of the level of processing at which learning takes place (certainly inconclusive, this issue will be further discussed in other sections). This inference relies on current modeling of the visual system, and on the processing stages assumed. It is generally assumed that an initial stage of processing within the visual system, the low-level stage, analyzes, in parallel over the retinal image, basic features such as local luminance contrast on different spatial scales, orientation, color, motion and eye-disparity, whereas higher levels of processing analyze a selective region of interest to perform object recognition, that is, higher level processes read out the output of the early representations (Dosher & Lu, 1999). Such a feedforward architecture, with many processing layers stacked hierarchically, has strong theoretical appeal and was shown to be capable of predicting human performance on rapid categorization tasks (Serre, Oliva, & Poggio, 2007). Within this conceptual

1556

D. Sagi / Vision Research 51 (2011) 1552–1566

framework, cortical area V1 is seen ‘‘as the dominant cortical relay station distributing visual sensory input to the rest of the neocortex’’, with higher levels in the neocortex generating an interpretation of the sensory data (Wandell & Smirnakis, 2009). Whereas during the early days of perceptual learning, the processing stages were viewed as hierarchically arranged with low-level processes fedforward into higher ones, later research revealed difficulties in identifying low-level processes that are local, context independent, and free of higher-level modulation. Experimental results provide strong evidence for long-range interactions between locally oriented detectors (Polat & Sagi, 1993) with brain correlates found within the primary visual cortex (Gilbert, Li, & Piëch, 2009; Tajima et al., 2010). Furthermore, strong task-dependent modulation of V1 activity (implying top-down process) was found (Li, Piëch, & Gilbert, 2008b, Offen, Schluppeck, & Heeger, 2009). Learning in visual tasks involving target identification, such as random patterns (Nazir & O’Regan, 1990), faces, and textures (Fig. 1d; (Hussain, Sekuler, & Bennett, 2009a)) shows specificity to retinal location (Nazir & O’Regan, 1990), orientation, and even to the particular exemplars (Hussain et al., 2009a). Within this state of affairs, the high degree of specificity of perceptual leaning with regard to stimulus parameters poses a challenge to theories of vision and to experiments attempting to localize learning within a specific stage of processing. Another path to take while attempting to isolate processing levels is by considering tasks that are limited by constraints on visual processing, which are known to have well-defined functional–anatomical correlates, that is, to consider the mapping between brain anatomy and visual function where it exists. A well-studied example concerns object recognition, with imaging studies showing a strong correlation between recognition performance and fMRI signals in the lateral occipital complex (LOC), with both performance and BOLD responses increasing with training (Grill-Spector, Kushnir, Hendler, & Malach, 2000). In the perception of motion, global analysis of movement in the visual field is thought to involve area MT. Lu, Qian, and Liu (2004) constructed a motion stimulus that was designed to overload area MT and found minimal learning under such conditions (Thompson & Liu, 2006). However, for most behavioral studies of perceptual learning such a hypothesized mapping does not suffice. It was already noted that performance at hyper-acuity levels, reached after extensive training (Fendick & Westheimer, 1983), requires access to low level neuronal representations where the details necessary for achieving this level of performance with an arbitrary task are encoded. However, theory shows (Poggio et al., 1992) that learning in such cases can be explained by a process that integrates the outputs of these low-level representations, in a way specifically configured for the special task. Indeed, experimental results show task specificity in learning of hyper-acuity (Fahle & Morgan, 1996), arguing against a generalized improvement in spatial resolution of early representation as the main cause of the learning effects. Psychophysical results point to an important difference between two task classes already mentioned regarding the absence of learning in search tasks. In one class of tasks the target and its location are well defined and are known to the observer; the relevant information for task performance can be found in a limited part of the display, thus allowing the observers to focus on that part with all resources available to them (e.g., attention and detailed pattern recognition). Most tasks used in studying perceptual learning are of this type, including orientation discrimination of a line element, vernier, bisection, and contrast discrimination. In another class of tasks, the target is presented among many distracters, and observers have to find a target placed in a background composed of many non-target distracters. Thus, they are forced to simultaneously handle many display elements using distributed processing. An example of such a task is the texture segmentation task (Fig. 1g). Here the visual field is covered with texture elements

(texels) according to some well-defined statistics, with different regions in the visual field having different statistical properties. Results obtained during the 1980s indicated that some texel combinations yield ‘‘effortless’’ segmentation, with performance being independent of the number of texels (i.e. parallel), whereas other combinations result in reduced performance with an increasing number of texels (Bergen & Julesz, 1983). Effortless segmentation was observed when texture targets differed from the background by some basic low-level feature, such as orientation. The important property of this task is the need to search for a target across many (often >100) display elements, which, with a limited presentation time, is doomed to fail when attempting to search through the display, texel by texel. It became increasingly apparent that searching is carried out by a parallel neuronal network with its neurons having receptive field properties similar to those observed in the primary visual cortex (Knierim & van Essen, 1992; Li, 2002; Rubenstein & Sagi, 1990; Tajima et al., 2010). With this distinction in mind, between distributed (parallel) and focused (serial) tasks, the finding of perceptual learning in orientation-based search (Ahissar & Hochstein, 1993), texture segmentation (Fig. 1g; (Karni & Sagi, 1991)), and contour integration (Fig. 1i; (Gilbert et al., 2009)) tasks strongly implies that experience modifies low-level visual processes, attention free (Braun & Sagi, 1991), expected to be found in the primary visual cortex. These studies demonstrated specificity to target location and orientation, as found with other tasks, with the additional partial specificity for the trained eye (Ahissar & Hochstein, 1996; Karni & Sagi, 1991). Eye specificity was found only at a later stage of learning after the first training session, and it was characterized by its slow time course and by a higher level of specificity to image features (Karni & Sagi, 1993). Similar results were reported for the focused type of experiments, mostly location and orientation specificity, and in some cases, also eye specificity such as with the high-accuracy Vernier discrimination (Fig. 1b; (Fahle, 2004)). Overall, the emerging pattern of results supports neuronal plasticity at a low-level stage of visual processing, though not only, and possibly at all cortical sites involved in the task.

4. Neuronal correlates Neuronal correlates of perceptual learning in humans were studied using fMR-Imaging and EEG methods. fMRI methods are used to identify a mapping between brain anatomy and function, that is, which brain areas change their activity level as a result of learning. Of particular interest here are the effects of practice on activity in the primary visual cortex. Schwartz, Maquet, and Frith (2002) had observers practicing the texturediscrimination task with only one eye open. In accordance with the monocularity of texture learning (Karni & Sagi, 1991), BOLD signals from V1 were found to be stronger when the observers viewed the texture stimuli with their trained eye as compared with their untrained eye, thus supporting the V1 hypothesis of texture learning. Lewis, Baldassarre, Committeri, Romani, and Corbetta (2009), using fMRI methods, found a learning-dependent increase in the evoked BOLD activity in the visual cortex, combined with an increase in resting BOLD functional connectivity between the trained region in the visual cortex and frontal–parietal regions which was correlated with learning. Schwarzkopf, Zhang, and Kourtzi (2009) identified a learning network that includes occipitotemporal and frontoparietal areas, with occipital BOLD activity better correlating with performance gain owing to learning whereas frontoparietal responses do not. Walker, Stickgold, Jolesz, and Yoo (2005) found increased BOLD activity in the primary visual cortex (V1) when the texture training was followed by sleep relative to no-sleep. Yotsumoto,

D. Sagi / Vision Research 51 (2011) 1552–1566

Watanabe, and Sasaki (2008) found that these training-based increases in BOLD activity disappeared after a few days of training, supporting the normalization hypothesis discussed in the context of memory consolidation (see below) and sleep (Censor & Sagi, 2009; Tononi & Cirelli, 2003). This hypothesis was further supported by the finding of increased BOLD activity in V1 during the slow-wave-sleep (SWS) stage, which was correlated with texture learning (Yotsumoto, Sasaki et al., 2009). Furmanski, Schluppeck, and Engel (2004) found increased fMRI signals in V1 after practicing a grating task for 30 days. Overall, these studies support a localized increase in processing efficiency in the primary visual cortex as a result of practice. Such a change in efficiency can be either due to functional plasticity at the primary visual cortex area or due to changes in inputs this area receives from other brain regions that may have changed with practice. Differential effects between lower and higher cortical regions along the visual processing stream, such as were found by Schwarzkopf et al. (2009), may provide us with a better understanding of the relationship between learning and changes in BOLD activity. The more recently introduced Diffusion-TensorImaging (DTI) method can be used to estimate anatomical changes associated with skill learning, corresponding to changes in white and gray matter density (Sagi, Tavor, Sasson, Pasternak, & Assaf, 2009; Scholz, Klein, Behrens, & Johansen-Berg, 2009). Yotsumoto et al. (2010) provide evidence supporting such structural changes in visual area V1 associated with texture learning. Yet another method to study plasticity of visual areas is by training patients with cortical damage; however efficient training methods remain to be developed (Huxlin, 2008). The V1 plasticity hypothesis is further supported by the time course of the learning-dependent brain activity, measurable with EEG methods. Both within session (Casco, Campana, Grieco, & Fuggetta, 2004; Skrandies & Fahle, 1994), and between-session (Pourtois, Rauss, Vuilleumier, & Schwartz, 2008) learning were found to modulate EEG responses. Early EEG components (C1) corresponding to V1 activity, starting at 40 ms after stimulus onset, were found to be modulated by perceptual learning in a texturediscrimination task (Pourtois et al., 2008). Since top-down processes are seen in V1 only 100 ms after stimulus onset (Li, Piëch, & Gilbert, 2004), the early influences (<85 ms) observed by Pourtois et al. (2008) suggest that learning induces local changes within V1. Censor, Bonneh, Arieli, and Sagi (2009) examined the relationship between human visual performance and visual event-related potentials (ERPs) using the backward-masked texture segmentation task. Their results showed practice-dependent temporal-interactions between early components (N1) of the ERPs corresponding to the target and the mask, recorded over occipital electrodes. These temporal interactions correlated with reduced performance on the task and could be used to predict observers’ thresholds. It seems that practice reduces temporal interactions between successive stimuli, possibly by increasing the efficiency, or speeding up, of target processing within early visual areas. Electrophysiological studies point to neural correlates of perceptual learning in the primary visual cortex of the monkey for certain tasks. Experiments using the orientation discrimination task show little improvement (Schoups, Vogels, Qian, & Orban, 2001) or none (Ghose, Yang, & Maunsell, 2002) in neuronal responses that can support the observed improved performance. Vogels (2010) showed an increase in the response slope of V1 neurons at the trained orientation and the retinal location, using data from Schoups et al. (2001). Although these experiments, employing tasks corresponding to localized stimuli, fail to display large effects of learning in V1, experiments probing contextual effects, such as in contour integration, do display strong effects (Gilbert, Sigman, & Crist, 2001). Interestingly, the corresponding neuronal enhancements are not seen when the trained task is not performed or when

1557

the monkey is anesthetized (Li et al., 2004), implying that the expression of these learning effects depend on task dependent processes, possibly managed by some higher levels of processing. In this regard, Law and Gold (2008) showed strong correlations between learning effects and neuronal response in area LIP, but not in MT, in monkeys performing a motion discrimination task. These results were interpreted as modification in the readout of MT neurons by LIP neurons. Such a result shows that learning mechanisms have access to low-level perceptual representations within the visual cortex. New technologies offer exciting tools for studying cortical plasticity. In the monkey, molecular tools are used to measure gene expression associated with cortical reorganization (Chen, Yamahachi, & Gilbert, 2010). Advances in axon labeling by viruses and in vivo two-photon microscopy are used to investigate axon branching and bouton dynamics in the primary visual cortex of adult Macaque monkeys. There is increasing evidence that experience-dependent plasticity of specific circuits in the somatosensory and visual cortex involves cell type-specific structural plasticity: some boutons and dendritic spines appear and disappear, accompanied by synapse formation and elimination, respectively (Holtmaat & Svoboda, 2009). Stettler, Yamahachi, Li, Denk, and Gilbert (2006) found that overall axonal branching patterns were stable but that a subset of small branches appeared and disappeared every week. Synaptic bouton losses and gains were both 7% of the total population per week, with no net change in the overall density. These results suggest constant plasticity in adult V1 in the absence of external cause. It is possible that these continuous changes are shaped by visual experience to produce functional plasticity within the visual system.

5. Network architecture Several theoretical and experimental paradigms were developed to study possible implementations of the learning process within neuronal networks, broadly divided into the feedforward (reweighting) and feedback (recurrent) types. It is possible to have a neuronal implementation that models the different brain states before, during, and after learning without a commitment to the learning process, but at the end, a specific implementation is expected to be capable of developing new neuronal states driven by well-defined forces, both internal (activity-dependent rewiring) and external (feedback). Much of the available data in perceptual learning can be modeled using a feedforward design without recurrent interactions (Dosher & Lu, 1999; Eckstein, Abbey, Pham, & Shimozaki, 2004; Poggio et al., 1992). Such a neuronal architecture assumes a cascade of processing stages, starting with an input layer and ending up with a decision unit. The decision unit integrates weighted inputs from neurons at a lower layer, and produces a response, with output levels corresponding to the possible behavioral responses in the task. Learning in such models can be implemented using a teaching signal that can be applied to modify the input weights to the decision unit using associative learning rules (Dosher & Lu, 1999), or one can reconfigure intermediate layers (Poggio et al., 1992). Stimulus and task specificity are modeled by having new learning modules (related to the decision or pre-decision stages) configured to yield optimal performance on new tasks. Feedback concerning decision errors is important for learning in such networks, though not absolutely necessary (Herzog & Fahle, 1997; Herzog & Fahle, 1999; Poggio et al., 1992). Alternatively, it can be internally generated through evaluation based on obvious discriminations (Fahle & Edelman, 1993; Herzog & Fahle, 1998). Petrov, Dosher, and Lu (2006) implemented learning in their model of reweighting by introducing a feedback-dependent bias to the decision unit. This bias, which depends on information received

1558

D. Sagi / Vision Research 51 (2011) 1552–1566

by the observer only after a response is made, requires the inputs to persist until the decision is reached, for a temporal duration of a few seconds. Polat and Sagi (1994b); Fig. 1h suggested a form of associative learning on such a time scale when studying lateral interactions, where some evidence for long-term persistence of sub-threshold effects was found (Tanaka & Sagi, 1998a; Tanaka & Sagi, 1998b). Learning often depends on factors external to the learned task, such as task-irrelevant stimuli that surround the target (spatial context) or stimuli preceding the task, possibly on the already existing memories (temporal context). These effects are easier to understand within the theoretical framework of recurrent (feedback) networks. Recurrent networks do not require external feedback to learn. Learning in these networks is often implemented using associative rules that modify the connection strength between pairs of neurons according to activity correlation. Cortical anatomy reveals that long-range horizontal interactions exist in all regions, including the visual cortex (Gilbert & Wiesel, 1983; Rockland & Lund, 1983), modulated by perceptual learning (Gilbert et al., 2009). In human psychophysics, similar experiencedependent long-range interactions were found between laterally displaced Gabor patches (Polat & Sagi, 1993; Polat & Sagi, 1994a). Polat and Sagi (1994b) found an increase in the range of lateral interactions, which was explained by the increased efficacy of existing connections, enabling lateral propagation of activity across multiple connections. This account is supported by a finding showing improved contour integration with practice, e.g., learning in tasks involving the detection of contours generated from multiple segments embedded in noise (Kovács & Julesz, 1993; Kovács, Kozma, Feher, & Benedek, 1999). Zhaoping Li (Zhaoping, 2009), applying a V1-based model of visual segmentation (Li, 2002), used lateral inhibitory interactions to model learning in texture tasks (Ahissar & Hochstein, 1996; Karni & Sagi, 1991) to successfully model a variety of results found in the literature. It was suggested that learning in texture-discrimination tasks involves strengthening of inhibitory interactions between adjacent cortical units responding to the oriented texture elements, in agreement with the experimental result showing specificity of learning to background orientation (Karni & Sagi, 1991). Such a model shows that behavior is dependent on spatial parameters such as line length and density (Sagi, 1990; Sagi & Julesz, 1987). A V1-like model, assuming cortical columns consisting of interconnected excitatory and inhibitory subpopulations (Wilson & Cowan, 1972), is supported by contextual effects found in contrast discrimination (Adini et al., 2002). The experimental results of contrast discrimination with localized grating patches (Gabor signals) indicate stable contrast-discrimination thresholds (see above in Limitations). However, the addition of fixed-contrast Gabor flankers (Fig. 1f, bottom) enabled learning, resulting in reduced discrimination thresholds. Importantly, this learning effect was preserved when the contextual flankers were removed. Adini et al. (2002) suggested that performance on contrast discrimination depends on a balanced activity between the excitatory and inhibitory subpopulations. This balance, maintained by activitydependent synaptic connections obeying hebbian and antihebbian learning rules, is preserved when the network is locally stimulated but is disturbed when lateral inputs are added. Thus, contextual perturbations lead to learning within local networks. In this model, lateral interactions are not necessarily modified through weight changes in lateral connections but rather by modifications of excitatory and inhibitory weights within a column. Recent reports suggest that plasticity in the visual cortex during the critical period (see below) is controlled by the excitatoryinhibitory balance (Morishita & Hensch, 2008), with reduced intra-cortical inhibition allowing for adult cortical plasticity (Harauzov et al., 2010; Sale et al., 2007).

6. Critical period The concept of a critical period states that some brain functions are shaped during early life and cannot be later modified. Properties of cortical neurons, like ocular dominance, were shown to be affected by experience only during a critical period of several months early in the postnatal life of kittens (Hubel & Wiesel, 1970; Wiesel & Hubel, 1963). Apparently, there are multiple time scales of functional maturation, with some functions maturing relatively early in life, such as stereo vision within a few (3–4) months after birth (Braddick et al., 1980), but some, such as contour integration, take many (>15) years to reach the adult level of performance (Kovács, 2000). An open question is whether perceptual learning is continuous with visual development. One way to tackle this question is by considering cases of abnormal visual development. Sinha and coworkers (Bouvrie & Sinha, 2007; Ostrovsky, Andalman, & Sinha, 2006) studied a case of a woman born blind owing to congenital cataracts that were removed at the age of 12. Her vision was tested 20 years later, showing relatively low acuity but very good image segmentation and recognition. These results suggest that the visual system retains its plasticity after many years of deprivation. A well-known case of abnormal visual development is amblyopia. Amblyopia develops as a result of abnormal binocular experience during the critical period. Whereas the absence of visual inputs during development may leave the visual cortex untouched, as in day zero (Mitchell & Sengpiel, 2009), or these inputs may be replaced by other functions, such as verbal memory (Amedi, Raz, Pianka, Malach, & Zohary, 2003), abnormal visual experience may introduce developmental forces that suppress some functions, such as increased intra-cortical inhibition observed with rats that underwent monocular deprivation during the critical period (Maffei, Nataraj, Nelson, & Turrigiano, 2006). Recent studies show that adult amblyopic rats restore ocular dominance plasticity and can regain visual acuity after a period of complete visual deprivation (He, Ray, Dennis, & Quinlan, 2007) or after being exposed to an enriched environment (Sale et al., 2007). Intra-cortical inhibition was suggested to be a limiting factor in those cases of adult cortical plasticity (Harauzov et al., 2010). In humans, perceptual learning was found to be effective in amblyopic eyes and to improve visual acuity (Levi & Li, 2009a; Levi & Li, 2009b; Levi & Polat, 1996; Levi, Polat, & Hu, 1997; Pennefather, Chandna, Kovács, Polat, & Norcia, 1999; Polat, 2009). The broader bandwidth of learning, with learning effects better generalized to other visual stimuli, was considered as an indication of greater visual plasticity in amblyopes (Astle, Webb, & McGraw, 2010; Huang, Zhou, & Lu, 2008). Polat, Ma-Naim, Belkin, and Sagi (2004) had a large group of amblyopes practicing contrast tasks of detection and discrimination with flanked Gabor patches as stimuli (Fig. 1h), and they found a significant improvement in the trained task. Most surprisingly, the trained amblyopes had their visual acuity improved by a factor of two after training, thus suggesting an effective treatment for amblyopia (Polat et al., 2004). Such a treatment may require a better understanding of the neuronal deficits underlying amblyopia. Polat et al. (2004) based their treatment on the assumption that lateral interactions within the primary visual cortex are at fault, an assumption supported by data showing the absence of collinear lateral facilitation (Polat & Sagi, 1993; Polat & Sagi, 1994a) in amblyopes (Polat, Sagi, & Norcia, 1997). Furthermore, lateral interactions measured through the amlyopic eye were characterized by excessive inhibition, which was much reduced through training (Polat et al., 2004). Not in contradiction with the former, amblyopes may also benefit from improved detection templates matched to the recovering eye (Li, Klein, & Levi, 2008; Li, Levi, & Klein, 2004). Perceptual learning combined with pharmacological treatment, possibly to reduce inhibitory effects,

D. Sagi / Vision Research 51 (2011) 1552–1566

suggested by the animal studies mentioned above (Harauzov et al., 2010; Morishita & Hensch, 2008), or with transcranial brain stimulation (TMS: (Thompson, Mansouri, Koski, & Hess, 2008)), offers a very promising method for overcoming developmental disorders.

7. Task requirements Low-level functional plasticity requires some behavioral control and thus learning is expected to depend on the task used and to exhibit task specificity. It is well established that the development of visual function in kittens during their ‘‘critical period’’ is very much affected by visually guided behavior (Held & Hein, 1963), though such a result does not rule out passive recognition learning (Christou & Bulthoff, 1999). Ahissar and Hochstein (1993) showed that learning in a pop-out detection task takes place when observers are trained with the detection task but to a lesser extent when trained on another task using the same stimuli. A similar result was found by Karni and Sagi (1995), using two uncorrelated texture targets with only one being task relevant: performance on the task-irrelevant target did not improve despite the target being presented as frequently as the trained one. More recent studies (Seitz & Watanabe, 2003) suggest that correlations between the task-relevant and task-irrelevant targets, even if created by different visual features, enable the learning of the task-irrelevant target (see below). These findings do not rule out low-level learning but rather imply that neuronal modification depends on a gating signal that is task dependent. Ahissar and Hochstein (1996), Ahissar and Hochstein (1997) postulated a ‘‘reverse hierarchy’’ of learning, according to which learning within different processing levels is guided by task requirements, some of which are non-visual, which are defined at higher levels of processing. Cortical networks are very complex; when a stimulus is presented, it activates thousands of neurons that are highly connected, mostly by local connections but also with long-range connections. Thus, stimulus-driven synaptic modification is not expected to improve performance on an arbitrary task unless modulated by task demands. For example, some tasks, such as vernier, may require better spatial localization of receptive fields, whereas other tasks, such as orientation discrimination, may require better orientation tuning, which may benefit from spatial integration. Moreover, even when the same task of spatial acuity is used, learning is not transferred to stimuli slightly different from the trained one (Fahle & Morgan, 1996). Although behavioral relevance is essential for storing a new experience in the brain, suggesting the existence of an extra-retinal control mechanism to gate functional plasticity, the underlying mechanism is not known. It was suggested that attention mechanisms control learning by means of selection (Ahissar & Hochstein, 1993), or competition combined with a reward-dependent reinforcement mechanism (Roelfsema, van Ooyen, & Watanabe, 2010). Attention mechanisms can be expedited by training and, as a consequence, performance can be improved for a wide range of tasks (Green & Bavelier, 2003). Some stimuli escape the task requirement and leave a mark without being noticed. Watanabe, Nanez, and Sasaki (2001) demonstrated that improved detection with a motion detection task was possible when the observer was repeatedly presented with motion stimuli without any task assigned to these stimuli during learning. In these experiments the observer’s task was to identify letters presented at the center of the visual field while the peripheral field was presented with dots moving at a low coherence level (Fig. 1c), effectively sub-threshold. With the random dot stimuli used, the estimated motion direction depends on spatial integration. Thus, it is possible to generate stimuli with different motion directions on different scales. In particular, integration of local motion vectors can produce global motion in a direction that is not

1559

present on a local scale. Watanabe et al. (2002) showed that during passive learning, termed TIPL (task-irrelevant perceptual learning), performance improves with motion directions defined on a local, rather than a global scale, implying low-level neural correlates, which are thought to reside in the primary visual cortex. Further experimentation revealed that TIPL occurs only when the target is sub-threshold (Tsushima, Seitz, & Watanabe, 2008). Most importantly, TIPL occurs for a nonrelevant peripheral target when presented simultaneously with a correctly identified task-relevant target (Seitz & Watanabe, 2003), suggesting that a global reward signal is triggered by successful performance on the apparently unrelated, letter identification task. Sasaki et al. (2010) suggest that it is the reward signal, rather than visual attention, that reinforces learning both in task-relevant and task-irrelevant learning. Within this context, TIPL represents a case where attention fails, with attention viewed as a mechanism for selection between salient stimuli competing for access to ‘‘awareness’’ (see a detailed description in (Roelfsema et al., 2010)). This theory assumes that successful processing of sensory information and control of plasticity depends on feedback from frontal regions and a global error correction mechanism.

8. Memory consolidation and reconsolidation After the stimulus is detected, attended to, and has responded and been rewarded, a memory trace needs to be formed and consolidated for the gained experience to be efficiently used in the future. Behavioral markers of consolidation time include resistance to interference and delayed between-session performance gain. It is possible to interfere with consolidation by having a second task practiced during the consolidation time. Here it is assumed that memory traces generated by the first task are affected by the new training, and are possibly changed if the second task imposes different processing requirements on brain networks shared by the two tasks. Evidence for task-dependent interference during visual consolidation was recently observed with the 3-dot alignment task (Seitz et al., 2005) and with the texture-discrimination task (Yotsumoto, Chang, Watanabe, & Sasaki, 2009), indicating a time scale of 1 h. Seitz et al. (2005) had observers practice the alignment task with the middle dot shifted to one side in nonaligned trials (Fig. 1b), resulting in a 10% improvement in acuity after a few days. However, this improvement was not observed under conditions where the observers had the opposite task (a nonaligned stimulus with the dot shifted to the other side) practiced during the same session, in different blocks of trials, or within 1 h after practicing the original task. This effect was shown to be location and orientation specific, supporting interference at a stimulus encoding level rather than at a decision or response level. Evidence from motor learning suggests an interference time window of up to 4–6 h (Brashers-Krug, Shadmehr, & Bizzi, 1996). Experiments with the visual texture-discrimination task show a latent period with performance improvement detected only when the time between training sessions exceeds 4–6 h (Karni & Sagi, 1993), supporting a time scale of a few hours for visual consolidation. The literature on memory consolidation suggests a wide range of consolidation timelines, ranging from minutes to years (Dudai, 2004). As a general rule, the mixing of visual stimuli and tasks in the same temporal proximity tends to reduce learning. However, there is distinction between mixing different stimuli within the same block of trials, the ‘roving’ method, and mixing where training on one task is completed before the other task begins. The latter is traditionally used in studying consolidation, whereas the former is used in studying modes of learning. The two mixing methods may produce different results, as the roving method, probably because the introduced stimulus uncertainty

1560

D. Sagi / Vision Research 51 (2011) 1552–1566

leads to learning that is generalized across the trained tasks (see below: perceptual modeling and overfitting). Interference in consolidation can be obtained by manipulations that indirectly affect the neuronal process involved in consolidation. Manipulations of sleep after training were found useful in identifying two processes involved in consolidating perceptual learning: normalization (or stabilization) and enhancement. Karni, Tanne, Rubenstein, Askenasy, and Sagi (1994) found that deprivation from Rapid-Eye-Movement (REM) sleep interferes with learning of texture discrimination if training is carried out shortly before sleep, suggesting that specific sleep stages have a functional role in consolidating perceptual learning, possibly in memory consolidation in general. Stickgold, James, and Hobson (2000), Stickgold, Whidbee, Schirmer, Patel, and Hobson (2000), using the same texture task (Fig. 1g), found that learning effects correlate with sleep duration, depending on both REM and Slow-Wave-Sleep (SWS) stages. Importantly, different sleep stages alternate several times during sleep, showing stronger correlations with SWS during the first quarter of sleep, whereas correlations with REM were stronger during the last quarter of sleep. These results suggest that a twostage process of memory consolidation occurs during sleep. Mednick, Nakayama, and Stickgold (2003) showed that improvement also takes place after a brief nap (60–90 min) containing both SWS and REM stages but not when only SWS was recorded. A short nap containing only SWS sleep was found useful in preventing performance deterioration that otherwise develops with repeated task performance during a day of training (Mednick et al., 2002) or within a training session (Mednick, Arman, & Boynton, 2005; Ofen, Moran, & Sagi, 2007). Improved performance can be obtained without sleep, within a training session (Aberg, Tartaglia, & Herzog, 2009; Fahle, Edelman, & Poggio, 1995; Hussain, Sekuler, & Bennett, 2008; Hussain, Sekuler, & Bennett, 2009b; Karni & Sagi, 1993), or between training sessions (Karni & Sagi, 1993). Gervan and Kovács (2010), using a contour integration task (Kovács & Julesz, 1993), found between-session learning (offline) to be relatively ineffective during the day but significant after a night of sleep. Censor, Karni, and Sagi (2006), using the texture-discrimination task, showed that the dependency on sleep is affected by the number of trials within a training session. A relatively small number of trials (200) produced equal learning effects with and without sleep, whereas learning with an increased number of trials (400) required sleep. A further increase in the number of trials blocked learning regardless of sleep conditions (Censor et al., 2006), a saturation effect possibly analogous to the interference effects mentioned above (Yotsumoto, Chang et al., 2009). Overall, these results suggest that sleep plays a role in protecting against interference and over-training (normalization of synaptic weights to avoid local saturation, without which further training may cause interference), and in strengthening the desired memories (enhancement). It is reasonable to assume that normalization is carried out in the SWS stage and that enhancement is carried out in the REM stage. Such a role of sleep in learning is consistent with the above-mentioned results: (1) effects of learning without sleep with short or non-demanding training sessions where saturation is avoided, (2) removal of saturation effects after a short sleep limited to SWS, (3) a gain in performance obtained after longer sleep periods including REM, and (4) during sleep, an initial correlation of performance gain with SWS followed by correlation with REM. Such a description is consistent with a broader view of the memory function of sleep, suggesting that sleep optimizes the consolidation of newly acquired information in memory, depending on the specific conditions of learning and the timing of sleep (Diekelmann & Born, 2010). The suggested division of consolidation into two stages, normalization and enhancement, can be mapped into stages of visual processing defined by their spatial invariance. Censor and Sagi (2008), using the texture-discrimination task,

found that a short training session of 200 trials, followed by sleep, protects against performance deterioration observed in longer sessions (Censor et al., 2006). This result indicates that a short training episode, with its learning result efficiently consolidated, enables interference-free performance, thus avoiding saturation due to over-learning. Interestingly, this ‘‘protection’’ effect was found to be transferred across space, that is, an initial short training with the target presented to some retinal locations protects against interference at other, previously untrained, locations (given current knowledge, see below in ‘‘perceptual modeling and overfitting’’, this transfer is consistent with transfer properties of perceptual learning after a single practice session; e.g. (Jeter, Dosher, Liu, & Lu, 2010) and when having positional uncertainty (Harris & Sagi, 2010). Most importantly, this benefit of short training was not observed when the trained location was previously exposed to interference through extended training (Censor & Sagi, 2009). These results indicate that the deterioration effects owing to interference and saturation are locally preserved whereas the benefits of short training are possibly stored at a higher level of processing and are generalized across space. These implied processing stages may correspond to the normalization and enhancement phases that exist during sleep. Accordingly, interference and saturation are expected to reflect properties of early sensory cortices, whereas enhancement may correspond to processes involved in recognition, possibly in creating a template (Dosher & Lu, 1999; Li et al., 2004) whose output can be used for decision (Censor & Sagi, 2009). Such an account is consistent with findings of increased localized SWS activity in the primary visual cortex after texture training, in correlation with the improved performance (Yotsumoto et al., 2009) and in the motor cortex after practicing a visuomotor task (Huber, Ghilardi, Massimini, & Tononi, 2004). Furthermore, Yotsumoto et al. (2008), using fMR-Imaging methods, found that the BOLD response in area V1 is reduced for a trained target in comparison with an untrained target, pointing to more efficient processing, possibly due to the normalization process suggested above. The overall product of learning must involve both local and global processing, and both normalization and enhancement. Reduced effective brain connectivity during SWS was proposed (Esser, Hill, & Tononi, 2009), suggesting that the SWS stage is better suited for local normalization. Although consolidated memories are thought to be protected against interference, it was also suggested that memories change with time, and that they are susceptible to interference upon retrieval (Bartlett, 1932). Such a mechanism, termed reconsolidation (Dudai, 2004), can be useful for keeping memories adjusted to current behavioral needs. In a recognition experiment carried out by Preminger, Blumenfeld, Sagi, and Tsodyks (2009), Preminger, Sagi, and Tsodyks (2007), the observers’ task was to report some previously familiarized faces (‘‘friends’’) when presented mixed in a series of trials with other faces (‘‘non-friends’’). Without the observers’ knowledge, one ‘‘friend’’ was gradually modified through a morphing process into another unfamiliar face, and indeed, observers confused the morphed faces with the familiar one as long as the two faces did not differ greatly. However, after a few days of practice, initially non-friend morphs were reported as friends, and the complete sequence was confused. Furthermore, perceptual similarities were also affected by this training, showing high similarity judgments for the two morph ends, initially of low perceptual similarity (Preminger, Blumenfeld, Sagi, & Tsodyks, 2009; Preminger, Sagi, & Tsodyks, 2007). Importantly, this merging of face memories was observed only when the morphed faces were orderly presented. Blumenfeld, Preminger, Sagi, and Tsodyks (2006) modeled memory of faces as attractors in an Attractor Neuronal Network ANN: (Hopfield, 1982), with morph instances implemented as activity patterns of varying degrees of correlations. This implementation predicts a single memory for

D. Sagi / Vision Research 51 (2011) 1552–1566

all correlated morph patterns, regardless of the order of presentation. To account for the order effect, Blumenfeld et al. (2006) suggested that learning is proportional to the distance between the experienced stimulus and its closest memory attractor. Thus, exposure to a stimulus that is only slightly different from an existing memory has little impact on network connectivity. This learning rule enables continuous recognition of slowly changing objects, such as with aging friends and family members, and provides long-term adaptation to drifting environmental parameters. Such mechanisms, as well as temporal associations between different object views or between visual cues may contribute to object appearance and recognition invariance (Haijiang, Saunders, Stone, & Backus, 2006; Wallis, Backus, Langer, Huebner, & Bulthoff, 2009).

9. Perceptual modeling and overfitting When performing a visual task, the observer’s goal is to construct a model of the sampled data that can efficiently guide decisions. The problem can be viewed as a statistical modeling problem, and amounts to describing a set of noisy samples by a small number of parameters that, by a formal model, can be fitted to the data and can be used to predict future behavior. Theories of perception assume that models constructed by the visual system are based on actual experience, an adaptation state achieved during early development; thus, perceptual learning is not expected as long as the environment’s statistical properties are fixed. The effects of perceptual learning can be viewed as a perturbation of this adaptive state by exposure to a newly defined environment, such as in the case of adaptation to rotation of the visual world (Kohler, 1962). This amounts to re-fitting of the model used by the brain, and it may possibly, when the stimuli and task used are very narrowly defined, as in most perceptual learning experiments, lead to over-specificity or to ‘‘overfitting’’. Perceptual overfitting is supported by some rather surprising recent experimental results showing that session length and the mixing of different stimuli and tasks affect learning and its specificity. The account, presented next, is based on the known fact that sensory systems are not invariant in space and time, and that internal activity is affected by a variety of conditions, deterministic or not, such as local adaptation to light and contrast. Hence, repeated stimuli do not evoke the same pattern of neuronal activity during each presentation, even within an experimental session, and, owing to retinal nonuniformity, translation of a stimulus is expected to produce different patterns of activity. My goal here is not to develop a detailed theory of perceptual learning that can fit the data, but rather, to consider a basic principle in statistical modeling, namely, the degree of fitting, i.e., how detailed the model needs to be to be useful. Stimuli and tasks used in perceptual learning are very simple, and for a good reason. The goal is to probe neural substrates underlying the basic visual process in the visual cortex, and to avoid highly specific processes involved in recognizing more complex objects. However, even within the primary visual cortex these simple stimuli, of low dimensionality, are processed by a complex network of many neurons connected by thousands of synapses; thus the neurons operate in high dimensional space. This situation may easily lead to perceptual overfitting. In the case of overfitting, the learning process models spurious properties of stimulus representation, possibly accidental noisy variations of the means that are found most useful for the task at hand. These highly complex fitted features may or may not be explicitly available to the observer. Such a detailed modeling of the stimulus is doomed to fail the task when the stimulus is later presented under a slightly different condition, or it may be added with new noisy events, or under different adapting conditions. Consider, for example, the task of contrast discrimination (Fig. 1f, top). In this task an observer

1561

discriminates between two contrast levels, C and C + DC, and the relevant performance measure is the magnitude of DC required for 75% correct discrimination. The two contrast levels are coded by neurons within the visual cortex, which increase their response as a function of contrast. It is therefore reasonable to assume that the contrast dependence of this internal response, and the trial-totrial variance of these responses, define the limits of performance. Thus, for the observer it is sufficient to construct a model that considers the relative efficiencies of the different neurons available in contributing to the differential signal, based on experience gained during training (Dosher & Lu, 1999; Eckstein et al., 2004; Jacobs, 2009; Liu & Weinshall, 2000; Petrov, Dosher, & Lu, 2005). However, a particular set of weights optimized for one contrast level may fail for another due to nonlinearities in the transduction process. In other words, with practice, the observer may find some special details regarding the neuronal response distribution, which carry information about the specific contrast that is practiced, but these details are useless with other contrast levels. Two predictions follow: (1) contrast discrimination learning is selective for contrast if training is restricted to a single contrast level; and (2) contrast discrimination learning is much reduced, possibly eliminated, when different contrast levels are interleaved during training. Both predictions are in agreement with the available experimental results (Aberg & Herzog, 2009; Adini et al., 2004; Yu et al., 2004). Interestingly, the initial performance level in contrast discrimination is not affected by the mixing of different contrast levels (contrast uncertainty), as if the visual system is optimized for contrast discrimination over the complete range of available contrasts (Adini et al., 2004). Perceptual overfitting can explain specificity in perceptual learning as well as reported failures in learning. Accordingly, given a simple task, the brain learns the peculiarities of the specific stimulus and the specifics of its neuronal mapping. The functional anatomy of the visual system indicates that simple visual features are coded by different networks in the visual cortex, such as in the case of location, orientation, and the eye. These networks may differ in the details of their construction, and an optimal discriminator adapted for one displayed feature may not generalize to another one. Mollon and Danilova (1996) suggested that such brain (and eye) non-uniformity can explain learning specificity without invoking plasticity at the actual low-level network: ‘‘. . .what the subject may be learning about are the local idiosyncracies of his retinal image, of his receptor mosaic’’. Consider a readout mechanism by which the optimal low-level activity pattern for the task is learned. Each stimulus change requires a readjustment of the readout process to find the specific activity pattern most useful for the task (Petrov et al., 2005). An example of overfitting is the use of accidental properties not invariant in space or time. Overfitting can be eliminated by using a richer stimulus set. Indeed, recent studies show that learning specificity is reduced under some experimental conditions. Zhang, Xiao, Klein, Levi, and Yu (2010), using an orientation discrimination task with a Gabor patch as a stimulus, found location specificity of orientation learning in agreement with previous results by Schoups, Vogel, and Orban (Schoups, Vogels, & Orban, 1995). In other words, the effects of training at one location are not transferred to a second location. However, Zhang et al. (2010) found that a brief pretest (200 trials) at the second location before practicing the first is sufficient to produce a substantial amount of transfer. Thus, this seemingly naïve pretest has a profound effect on learning. Similarly, transfer of texture learning across eyes was found when the eye of transfer was tested before training (Schoups & Orban, 1996), whereas eye specificity was found without such a pretest (Karni & Sagi, 1991) (though stimuli differed in some important details). These results indicate that during the initial phase of learning, perhaps 200 trials, the learning process captures some general properties of the task (performing

1562

D. Sagi / Vision Research 51 (2011) 1552–1566

a low-order approximation) without falling into the trap offered by the peculiarities of the local network activated by the stimulus. Indeed, better generalization of learning during the first session of learning was found with texture learning, indicating transfer between eyes (Karni & Sagi, 1993) and across orientations with masked contrast-discrimination (Jeter et al., 2010). Censor and Sagi (2009), using the texture-discrimination task, found reduced performance deterioration at untrained retinal locations after practicing a single short session, suggesting global effects of learning (see also above ‘‘Memory consolidation and reconsolidation’’). Learning obtained during this initial phase is sufficiently general to be useful for carrying out a similar task performed in a range of different experimental conditions. Overfitting is expected with fine discriminations but not with coarse discriminations, in agreement with the higher specificity found for ‘‘difficult’’ tasks relative to ‘‘easy’’ tasks (Ahissar & Hochstein, 1997). Jeter, Dosher, Petrov, and Lu (2009) showed that the precision of the task affects specificity of learning – the learning effect was transferred to new locations and orientations when tested with coarse orientation discrimination but not so much when tested with fine orientation discrimination. Liu and Weinshall (2000) found that a brief test with an easy discrimination task can facilitate transfer of motion discrimination learning. It seems that with coarse discrimination the fitting process ignores the fine details required for the finer discrimination tasks and as a result, overfitting is avoided. Of course, the result of this generalized learning should be preserved and consolidated in order to be protected from interference and to be useful in the future (see the relevant section on consolidation). Of particular importance are results pointing to a breakdown of locality in perceptual learning, since, as noted above, most perceptual tasks, if not all, show location specific learning, including tasks considered to involve high level recognition processes (e.g., Nazir & O’Regan, 1990). Xiao et al. (2008) developed a novel double-training paradigm where two different tasks (Gabor contrast discrimination and orientation discrimination) were practiced at two different locations, either simultaneously or at a different time. Their results showed significant transfer of the performance gained on each task to the location trained with the other, apparently irrelevant, task. This result is in apparent contradiction with the locality of perceptual learning, but here can be explained by the perceptual modeling process which is making use of features shared by all locations and tasks the observer is exposed to during training (including pretests!). Hila Harris and I have recently asked whether learning can be generalized to locations previously not tested, nor trained. We used a typical texture task (Karni & Sagi, 1991), but instead of using a single target location, we introduced location uncertainty, having the target positioned in one of two possible locations (different quadrants in the visual field, as in Censor & Sagi, 2008; Censor & Sagi, 2009). Learning curves with the two-locations training overlapped with standard learning curves obtained with single-location training, however, while the single-location training resulted in location specific learning, to our great surprise there was a complete transfer to untrained locations for observers trained with two locations (Harris & Sagi, 2010)! This result clearly shows that even learning through extended practice (four daily sessions), leading to high accuracy performance, can be generalized across retinal locations. The result supports the ‘‘overfitting’’ hypothesis according to which the modeling process, unable to keep track of the two randomly stimulated locations, produces a common fit to both locations, thus avoiding the ‘‘peculiarities’’ of each individual location, which in turn allows for generalization. However, overfitting predicts much less learning with two-location as compared with one location training, since the latter allows adjustments to the peculiarities of the single location (Mollon & Danilova, 1996).

Perceptual learning may show specificity even after a short training phase (Karni & Sagi, 1993; Rubin, Nakayama, & Shapley, 1997). Though a challenge to the account presented here, overfitting on a short time scale may lead to interesting effects, such as a failure of learning with long testing sessions. Within a testing session, stimulus encoding may vary owing to adaptation or neuronal noise. Thus, a learning mechanism that improperly integrates information obtained over time may fail to generalize to situations encountered later in the training session, hence failing with additional trials. Such performance deterioration was observed with texture discrimination (Censor et al., 2006; Mednick et al., 2005). The deterioration effect was found to be local, specific to the trained target location. Most importantly, a brief learning period, followed by efficient consolidation, was found to prevent this deterioration (Censor & Sagi, 2008), i.e., a relatively short training session followed by sleep prevents deterioration in subsequent training sessions. This effect was found to transfer to other target locations (Censor & Sagi, 2009). Thus, very much like the doubletraining effects with orientation and location (Zhang et al., 2010), previously discussed, a brief training period prevents overfitting and enables generalization to different states of sensory adaptation. This is most probably achieved by establishing an efficient set of weights for the task, or a stimulus-based template. Lu, Liu, and Dosher (2010) describe a large set of experimental results that are explained by stimulus-template enhancement. Of particular interest here is the finding that practicing with noise-free stimuli improves performance on noisy stimuli whereas practicing at a high noise level leads to improved performance only at that specific noise level. According to the present account, practicing with high noise leads to overfitting to the specific stimuli encountered, whereas without noise an efficient stimulus template is configured to provide a generalized filter that is noise independent. If all perceptual learning is overfitting, how much of the performance gain can be explained by it? Consider an extreme case where the visual system is optimized for the task under standard conditions so that any additional learning is not expected to improve performance, unless the training set is reduced to a small size so that overfitting can take place. Under such conditions, learning is expected to disappear when different stimuli are mixed during the learning process. Indeed, some experimental paradigms produce interesting results when training is carried out with trials of different stimuli mixed within a series of trials. Adini et al. (2004) and Yu et al. (2004), using contrast discrimination-based tasks, found reduced learning effects (Adini et al., 2004) or none (Yu et al., 2004) when different contrast levels were randomly mixed in a single series of trials, as compared with blocked contrast (a fixed pedestal). Otto, Herzog, Fahle, and Zhaoping (2006) found reduced learning with mixed stimuli using the bisection task, though Parkosadze, Otto, Malania, Kezeli, and Herzog (2008) demonstrated that significant learning occurs with very long training (18,000 trials). Tartaglia et al. (Tartaglia, Aberg, & Herzog, 2009) found that interference in mixed training depends on stimulus overlap. Using a rich stimulus set may reduce the risk of overfitting and thus reduce learning, possibly eliminating learning. However, this is not always the case. Ahissar and Hochstein (2000), using an orientation ‘‘pop-out’’ detection task, found strong learning effects with a target presented in one of many possible retinal locations, an effect not very different from the effects found with a single retinal location. Similarly, learning effects with one eye were found to equal those found with two eyes (Ahissar & Hochstein, 1996; Karni & Sagi, 1991). Yotsumoto, Chang et al. (2009), using the texture-discrimination task, found similar learning effects when stimulus properties were fixed as compared with a random mix of different background and target orientations. These results prove that extended learning, and its specificity, when a target is presented at a single location, or to a single eye, or with a single

D. Sagi / Vision Research 51 (2011) 1552–1566

orientation cannot be explained by overfitting alone. It is still possible that some learning effects under mixed conditions are due to overfitting when the different conditions are clearly marked, such as with explicit tagging (Zhang et al., 2008) or parameter grouping (Kuai, Zhang, Klein, Levi, & Yu, 2005), though the efficiency of these methods was found to dependent on the prticulat stimuli used (Aberg & Herzog, 2009). Does overfitting, where exists, imply any specific brain implementation? I do not think so. In fact, it assumes a single process that samples the sensory information at a very high resolution; it can be implemented within networks of early vision or as a highlevel shiftable readout of these low-level networks. This account of the phenomenology described above does not require any cortical hierarchy of receptive field sizes, upward or in reverse (Ahissar & Hochstein, 1996; Nahum, Nelken, & Ahissar, 2010), but the presence of such a cortical hierarchy is not in conflict with it. One interesting possibility is that the cortical hierarchy serves as a regularization process to reduce overfitting, maybe by providing global constraints on local computations, or by imposing prior knowledge on the learning process (Censor & Sagi, 2009). However, the reverse is also possible, by implementing regularization in lowlevel networks as constraints on learning, with higher levels introducing complex, possibly accidental, features. The latter option predicts, maybe against intuition, less overfitting in learning situations where attention is not available to the trained task (TIPL: (Sasaki et al., 2010)), or when it is not required for task performance. Indeed, results with texture discrimination (Karni & Sagi, 1991; Yotsumoto, Chang et al., 2009) and pop-out (Ahissar, Laiwand, & Hochstein, 2001) tasks, mentioned above, do not conform to the overfitting hypothesis. Thus, it is possible that only when the task allows for a complex or a semantic high-level process to be involved, such as when a constant target is presented or when very high accuracy is required, does overfitting take place.

10. Summary Maybe the most important conclusion that can be drawn from the perceptual learning studies of recent years is that understanding the mechanisms of learning within perceptual systems is a tractable problem. Importantly, perceptual learning offers a structured methodology to advance the understanding of learning and memory. The research described here shows that experience with visual stimuli results in long-term changes in the perception of these stimuli, pointing to experience-dependent plasticity in the visual system. More recent experiments demonstrate the dependence of learning on temporal and spatial context, allowing for a better identification of the brain networks underlying learning and of the rules governing learning. An increasing number of studies show that under some conditions the learning outcome generalizes to untrained stimuli and tasks. Better generalization is seen with shorter training sessions, with coarse discriminations, and in training with two or more stimuli, whereas extensive training with a fixed stimulus may lead to over-specificity. These properties of perceptual learning, generalization, and over-specificity can be explained by the statistical properties of the learning process. Stimulus uncertainty (‘‘roving’’) is suggested as a tool to minimize perceptual overfitting. Processes following the acquisition (encoding) stage are thought to operate on the resulting memory traces to normalize network connectivity in order to avoid local synaptic saturation, to eliminate spurious memories (overfitting), and to stabilize the resulting outcome against future interference. Specific sleep stages are thought to underlie the different stages of the consolidation process. Of particular interest are populations with abnormal visual development, such as amblyopia, where practice with a limited

1563

set of stimuli can lead to improved vision. These latter results suggest that sensory development extends beyond the traditionally thought critical period. Extra-sensory functions, such as attention and reinforcement, were considered critical for perceptual learning. The high performance level achieved with training, such as in hyper-acuity and with information-rich stimuli (e.g. textures), implies that the learning process has access to details of sensory representation represented only in low-level cortical networks. Learning within these networks were suggested to follow rules of associative learning. Studies employing single-cell recording methods in the monkey as well as fMR-Imaging methods in humans show practice-dependent modulation of neuronal activity with the primary visual cortex. However, whether these low-level networks are actually modified is a source of controversy. Most interestingly, it was reported that cortical networks, including sensory areas exhibit spontaneous plasticity in the absence of stimulations, suggesting that functional plasticity, i.e. learning, is driven by external inputs controlling the ongoing cortical plasticity. Acknowledgments This work was supported by the Basic Research Foundation administered by the Israel Academy of Sciences and Humanities. I also thank Nitzan Censor, Hila Harris, Alex Petrov, Yaniv Sagi, and the two reviewers for helpful comments on an earlier version of this paper. References Aberg, K. C., & Herzog, M. H. (2009). Interleaving bisection stimuli – randomly or in sequence – does not disrupt perceptual learning, it just makes it more difficult. Vision Research, 49(21), 2591–2598. Aberg, K. C., Tartaglia, E. M., & Herzog, M. H. (2009). Perceptual learning with Chevrons requires a minimal number of trials, transfers to untrained directions, but does not require sleep. Vision Research, 49(16), 2087–2094. Adini, Y., Sagi, D., & Tsodyks, M. (2002). Context-enabled learning in the human visual system. Nature, 415(6873), 790–793. Adini, Y., Wilkonsky, A., Haspel, R., Tsodyks, M., & Sagi, D. (2004). Perceptual learning in contrast discrimination: The effect of contrast uncertainty. Journal of Vision, 4(12), 993–1005. Ahissar, M., & Hochstein, S. (1993). Attentional control of early perceptual learning. Proceedings of the National Academy of Sciences of the United States of America, 90(12), 5718–5722. Ahissar, M., & Hochstein, S. (1996). Learning pop-out detection: Specificities to stimulus characteristics. Vision Research, 36(21), 3487–3500. Ahissar, M., & Hochstein, S. (1997). Task difficulty and the specificity of perceptual learning. Nature, 387(6631), 401–406. Ahissar, M., & Hochstein, S. (2000). The spread of attention and learning in feature search: Effects of target distribution and task difficulty. Vision Research, 40(10– 12), 1349–1364. Ahissar, M., Laiwand, R., & Hochstein, S. (2001). Attentional demands following perceptual skill training. Psychological Science, 12(1), 56–62. Amedi, A., Raz, N., Pianka, P., Malach, R., & Zohary, E. (2003). Early ‘visual’ cortex activation correlates with superior verbal memory performance in the blind. Nature Neuroscience, 6(7), 758–766. Anstis, S. M. (1970). Phi movement as a subtraction process. Vision Research, 10(12), 1411–1430. Astle, A. T., Webb, B. S., & McGraw, P. V. (2010). Spatial frequency discrimination learning in normal and developmentally impaired human vision. Vision Research, 50(23), 2445–2454. Ball, K., & Sekuler, R. (1982). A specific and enduring improvement in visual motion discrimination. Science, 218(4573), 697–698. Ball, K., & Sekuler, R. (1987). Direction-specific improvement in motion discrimination. Vision Research, 27(6), 953–965. Barlow, H. B. (1972). Single units and sensation: A neuron doctrine for perceptual psychology? Perception, 1(4), 371–394. Bartlett, F. C. (1932). Remembering: A study in experimental and social psychology. Cambridge: University Press. Bergen, J. R., & Julesz, B. (1983). Parallel versus serial processing in rapid pattern discrimination. Nature, 303(5919), 696–698. Blakemore, C., & Campbell, F. W. (1969). On the existence of neurones in the human visual system selectively sensitive to the orientation and size of retinal images. The Journal of physiology, 203(1), 237–260. Blumenfeld, B., Preminger, S., Sagi, D., & Tsodyks, M. (2006). Dynamics of memory representations in networks with novelty-facilitated synaptic plasticity. Neuron, 52(2), 383–394.

1564

D. Sagi / Vision Research 51 (2011) 1552–1566

Bouvrie, J. V., & Sinha, P. (2007). Visual object concept discovery: Observations in congenitally blind children, and a computational approach. Neurocomputing, 70(13–15), 2218–2233. Braddick, O., Atkinson, J., Julesz, B., Kropfl, W., Bodis-Wollner, I., & Raab, E. (1980). Cortical binocularity in infants. Nature, 288(5789), 363–365. Brashers-Krug, T., Shadmehr, R., & Bizzi, E. (1996). Consolidation in human motor memory. Nature, 382(6588), 252–255. Braun, J., & Sagi, D. (1991). Texture-based tasks are little affected by second tasks requiring peripheral or central attentive fixation. Perception, 20(4), 483–500. Campbell, F. W., & Robson, J. G. (1968). Application of Fourier analysis to the visibility of gratings. Journal of Physiology, 197(3), 551–566. Casco, C., Campana, G., Grieco, A., & Fuggetta, G. (2004). Perceptual learning modulates electrophysiological and psychophysical response to visual texture segmentation in humans. Neuroscience Letters, 371(1), 18–23. Censor, N., Bonneh, Y., Arieli, A., & Sagi, D. (2009). Early-vision brain responses which predict human visual segmentation and learning. Journal of Visions, 9(4). 12 11–19. Censor, N., Karni, A., & Sagi, D. (2006). A link between perceptual learning, adaptation and sleep. Vision Research, 46(23), 4071–4074. Censor, N., & Sagi, D. (2008). Benefits of efficient consolidation: Short training enables long-term resistance to perceptual adaptation induced by intensive testing. Vision Research, 48(7), 970–977. Censor, N., & Sagi, D. (2009). Global resistance to local perceptual adaptation in texture discrimination. Vision Research, 49(21), 2550–2556. Chen, J., Yamahachi, H., & Gilbert, C. D. (2010). Experience-dependent gene expression in adult visual cortex. Cerebral Cortex, 20(3), 650–660. Christou, C. G., & Bulthoff, H. H. (1999). View dependence in scene recognition after active learning. Memory and Cognition, 27(6), 996–1007. Daugman, J. G. (1985). Uncertainty relation for resolution in space, spatial frequency, and orientation optimized by two-dimensional visual cortical filters. Journal of the Optical Society of America A – Optics and Imagescience, 2(7), 1160–1169. De Valois, K. K. (1977). Spatial frequency adaptation can enhance contrast sensitivity. Vision Research, 17(9), 1057–1065. Diekelmann, S., & Born, J. (2010). The memory function of sleep. Nature Reviews Neuroscience, 11(2), 114–126. Dorais, A., & Sagi, D. (1997). Contrast masking effects change with practice. Vision Research, 37(13), 1725–1733. Dosher, B. A., & Lu, Z. L. (1999). Mechanisms of perceptual learning. Vision Research, 39(19), 3197–3221. Dosher, B. A., & Lu, Z. L. (2005). Perceptual learning in clear displays optimizes perceptual expertise: Learning the limiting process. Proceedings of the National Academy of Sciences of the United States of America, 102(14), 5286–5290. Dudai, Y. (2004). The neurobiology of consolidations, or, how stable is the engram? Annual Review of Psychology, 55, 51–86. Eckstein, M. P., Abbey, C. K., Pham, B. T., & Shimozaki, S. S. (2004). Perceptual learning through optimization of attentional weighting: Human versus optimal Bayesian learner. Journal of Visions, 4(12), 1006–1019. Ekman, G., & Lindman, R. (1962). Measurement of the underlying process in perceptual fluctuations. Vision Research, 2(7–8), 253–260. Esser, S. K., Hill, S., & Tononi, G. (2009). Breakdown of effective connectivity during slow wave sleep: Investigating the mechanism underlying a cortical gate using large-scale modeling. Journal of Neurophysiology, 102(4), 2096–2111. Fahle, M. (2004). Perceptual learning: A case for early selection. Journal of Visions, 4(10), 879–890. Fahle, M., & Poggio, T. (2002). Perceptual learning. Cambridge, Masssachusetts: MIT Press. Fahle, M., & Edelman, S. (1993). Long-term learning in vernier acuity: Effects of stimulus orientation, range and of feedback. Vision Research, 33(3), 397–412. Fahle, M., Edelman, S., & Poggio, T. (1995). Fast perceptual learning in hyperacuity. Vision Research, 35(21), 3003–3013. Fahle, M., & Henke-Fahle, S. (1996). Interobserver variance in perceptual performance and learning. Investigative Ophthalmology and Visual Science, 37(5), 869–877. Fahle, M., & Morgan, M. (1996). No transfer of perceptual learning between similar stimuli in the same retinal position. Current Biology, 6(3), 292–297. Fendick, M., & Westheimer, G. (1983). Effects of practice and the separation of test targets on foveal and peripheral stereoacuity. Vision Research, 23(2), 145–150. Fiorentini, A., & Berardi, N. (1980). Perceptual learning specific for orientation and spatial frequency. Nature, 287(5777), 43–44. Fiorentini, A., & Berardi, N. (1981). Learning in grating waveform discrimination: Specificity for orientation and spatial frequency. Vision Research, 21(7), 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(7), 573–578. Gervan, P., & Kovács, I. (2010). Two phases of offline learning in contour integration. Journal of Vision, 10(6), 24. Ghose, G. M., Yang, T., & Maunsell, J. H. (2002). Physiological correlates of perceptual learning in monkey V1 and V2. Journal of Neurophysiology, 87(4), 1867–1888. Gibson, E. J. (1969). Principles of perceptual learning. New York: Appleton-CenturyCrofts. Gilbert, C. D., Li, W., & Piëch, V. (2009). Perceptual learning and adult cortical plasticity. The Journal of Physiology, 587(Pt 12), 2743–2751. Gilbert, C. D., Sigman, M., & Crist, R. E. (2001). The neural basis of perceptual learning. Neuron, 31(5), 681–697.

Gilbert, C. D., & Wiesel, T. N. (1983). Clustered intrinsic connections in cat visual cortex. Journal of Neuroscience, 3(5), 1116–1133. Goldstone, R. L. (1998). Perceptual learning. Annual Review of Psychology, 49, 585–612. Green, C. S., & Bavelier, D. (2003). Action video game modifies visual selective attention. Nature, 423(6939), 534–537. Greenlee, M. W., Georgeson, M. A., Magnussen, S., & Harris, J. P. (1991). The time course of adaptation to spatial contrast. Vision Research, 31(2), 223–236. Grill-Spector, K., Kushnir, T., Hendler, T., & Malach, R. (2000). The dynamics of object-selective activation correlate with recognition performance in humans. Nature Neuroscience, 3(8), 837–843. Haijiang, Q., Saunders, J. A., Stone, R. W., & Backus, B. T. (2006). Demonstration of cue recruitment: Change in visual appearance by means of Pavlovian conditioning. Proceedings of the National Academy of Sciences of the United States of America, 103(2), 483–488. Harauzov, A., Spolidoro, M., DiCristo, G., De Pasquale, R., Cancedda, L., Pizzorusso, T., et al. (2010). Reducing intracortical inhibition in the adult visual cortex promotes ocular dominance plasticity. Journal of Neuroscience, 30(1), 361–371. Harris, H., & Sagi, D. (2010). Specificity in perceptual learning: Is it all overfitting? in preparation. He, H. Y., Ray, B., Dennis, K., & Quinlan, E. M. (2007). Experience-dependent recovery of vision following chronic deprivation amblyopia. Nature Neuroscience, 10(9), 1134–1136. Held, R., & Hein, A. (1963). Movement-produced stimulation in the development of visually guided behavior. Journal of Comparative and Physiological Psychology, 56, 872–876. Herzog, M. H., & Fahle, M. (1997). The role of feedback in learning a vernier discrimination task. Vision Research, 37(15), 2133–2141. Herzog, M. H., & Fahle, M. (1998). Modeling perceptual learning: Difficulties and how they can be overcome. Biological Cybernetics, 78(2), 107–117. Herzog, M. H., & Fahle, M. (1999). Effects of biased feedback on learning and deciding in a vernier discrimination task. Vision Research, 39(25), 4232–4243. Holtmaat, A., & Svoboda, K. (2009). Experience-dependent structural synaptic plasticity in the mammalian brain. Nature Reviews Neuroscience, 10(9), 647–658. Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences of the United States of America, 79(8), 2554–2558. Huang, C. B., Zhou, Y. F., & Lu, Z. L. (2008). Broad bandwidth of perceptual learning in the visual system of adults with anisometropic amplyopia. Proceedings of the National Academy of Sciences of the United States of America, 105(10), 4068–4073. Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. Journal of Physiology (London), 160, 106–154. Hubel, D. H., & Wiesel, T. N. (1970). The period of susceptibility to the physiological effects of unilateral eye closure in kittens. Journal of Physiology, 206(2), 419–436. Huber, R., Ghilardi, M. F., Massimini, M., & Tononi, G. (2004). Local sleep and learning. Nature, 430(6995), 78–81. Husk, J. S., Bennett, P. J., & Sekuler, A. B. (2007). Inverting houses and textures: Investigating the characteristics of learned inversion effects. Vision Research, 47(27), 3350–3359. Hussain, Z., Sekuler, A. B., & Bennett, P. J. (2008). Robust perceptual learning of faces in the absence of sleep. Vision Research, 48(28), 2785–2792. Hussain, Z., Sekuler, A. B., & Bennett, P. J. (2009a). Contrast-reversal abolishes perceptual learning. Journal of Vision, 9(4). 20, 21–28. Hussain, Z., Sekuler, A. B., & Bennett, P. J. (2009b). How much practice is needed to produce perceptual learning? Vision Research, 49(21), 2624–2634. Huxlin, K. R. (2008). Perceptual plasticity in damaged adult visual systems. Vision Research, 48(20), 2154–2166. Jacobs, R. A. (2009). Adaptive precision pooling of model neuron activities predicts the efficiency of human visual learning. Journal of Vision, 9(4), 22 21–15. Jacobs, R. A. (2010). Integrated approaches to perceptual learning. Topics in Cognitive Science, 2(2), 182–188. Jeter, P. E., Dosher, B. A., Liu, S. H., & Lu, Z. L. (2010). Specificity of perceptual learning increases with increased training. Vision Research, 50(19), 1928–1940. Jeter, P. E., Dosher, B. A., Petrov, A., & Lu, Z. L. (2009). Task precision at transfer determines specificity of perceptual learning. Journal of Vision, 9(3), 1–13. Julesz, B. (1971). Foundations of cyclopean perception. Chicago: The University of Chicago Press. Karni, A., & Sagi, D. (1991). Where practice makes perfect in texture discrimination: Evidence for primary visual cortex plasticity. Proceedings of the National Academy of Sciences of the United States of America, 88(11), 4966–4970. Karni, A., & Sagi, D. (1993). The time course of learning a visual skill. Nature, 365(6443), 250–252. Karni, A., & Sagi, D. (1995). A memory system in the adult visual cortex. In B. Julesz & I. Kovács (Eds.), Maturational windows and adult cortical plasticity. SFI Studies in the Sciences of Complexity. Vol. XXIV. Karni, A., Tanne, D., Rubenstein, B. S., Askenasy, J. J., & Sagi, D. (1994). Dependence on REM sleep of overnight improvement of a perceptual skill [see comments]. Science, 265(5172), 679–682. Knierim, J. J., & van Essen, D. C. (1992). Neuronal responses to static texture patterns in area V1 of the alert macaque monkey. Journal of Neurophysiology, 67(4), 961–980. Kohler, I. (1962). Experiments with goggles. Scientific American, 206, 62–72. Kovács, I. (2000). Human development of perceptual organization. Vision Research, 40(10–12), 1301–1310.

D. Sagi / Vision Research 51 (2011) 1552–1566 Kovács, I., & Julesz, B. (1993). A closed curve is much more than an incomplete one: Effect of closure in figure-ground segmentation. Proceedings of the National Academy of Sciences of the United States of America, 90(16), 7495–7497. Kovács, I., Kozma, P., Feher, A., & Benedek, G. (1999). Late maturation of visual spatial integration in humans. Proceedings of the National Academy of Sciences of the United States of America, 96(21), 12204–12209. Kuai, S. G., Zhang, J. Y., Klein, S. A., Levi, D. M., & Yu, C. (2005). The essential role of stimulus temporal patterning in enabling perceptual learning. Nature Neuroscience, 8(11), 1497–1499. Law, C. T., & Gold, J. I. (2008). Neural correlates of perceptual learning in a sensorymotor, but not a sensory, cortical area. Nature Neuroscience, 11(4), 505–513. Levi, D. M., & Li, R. W. (2009a). Improving the performance of the amblyopic visual system. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 364(1515), 399–407. Levi, D. M., & Li, R. W. (2009b). Perceptual learning as a potential treatment for amblyopia: A mini-review. Vision Research, 49(21), 2535–2549. Levi, D. M., & Polat, U. (1996). Neural plasticity in adults with amblyopia. Proceedings of the National Academy of Sciences of the United States of America, 93(13), 6830–6834. Levi, D. M., Polat, U., & Hu, Y. S. (1997). Improvement in vernier acuity in adults with amblyopia. Investigative Ophthalmology and Visual Science, 38(8), 1493–1510. Lewis, C. M., Baldassarre, A., Committeri, G., Romani, G. L., & Corbetta, M. (2009). Learning sculpts the spontaneous activity of the resting human brain. Proceedings of the National Academy of Sciences of the United States of America, 106(41), 17558–17563. Li, Z. (2002). A saliency map in primary visual cortex. Trends in Cognitive Sciences, 6(1), 9–16. Li, R. W., Klein, S. A., & Levi, D. M. (2008). Prolonged perceptual learning of positional acuity in adult amblyopia: Perceptual template retuning dynamics. Journal of Neuroscience, 28(52), 14223–14229. Li, R. W., Levi, D. M., & Klein, S. A. (2004). Perceptual learning improves efficiency by re-tuning the decision ‘template’ for position discrimination. Nature Neuroscience, 7(2), 178–183. Li, W., Piëch, V., & Gilbert, C. D. (2004). Perceptual learning and top-down influences in primary visual cortex. Nature Neuroscience, 7(6), 651–657. Li, W., Piëch, V., & Gilbert, C. D. (2008). Learning to link visual contours. Neuron, 57(3), 442–451. Li, R., Polat, U., Makous, W., & Bavelier, D. (2009). Enhancing the contrast sensitivity function through action video game training. Nature Neuroscience, 12(5), 549–551. Liu, Z., & Weinshall, D. (2000). Mechanisms of generalization in perceptual learning. Vision Research, 40(1), 97–109. Lu, Z. L., Liu, J., & Dosher, B. A. (2010). Modeling mechanisms of perceptual learning with augmented Hebbian re-weighting. Vision Research, 50(4), 375–390. Lu, H., Qian, N., & Liu, Z. (2004). Learning motion discrimination with suppressed MT. Vision Research, 44(15), 1817–1825. Lu, Z. L., Yu, C., Watanabe, T., Sagi, D., & Levi, D. (2009). Perceptual learning: Functions, mechanisms, and applications. Vision Research, 49(21), 2531–2534. Lu, Z. L., Yu, C., Watanabe, T., Sagi, D., & Levi, D. (2010). Perceptual learning: Functions, mechanisms, and applications. Vision Research, 50(4), 365–367. Maehara, G., & Goryo, K. (2007). Perceptual learning in monocular pattern masking: Experiments and explanations by the twin summation gain control model of contrast processing. Perception & Psychophysics, 69(6), 1009–1021. Maffei, A., Nataraj, K., Nelson, S. B., & Turrigiano, G. G. (2006). Potentiation of cortical inhibition by visual deprivation. Nature, 443(7107), 81–84. Mayer, M. J. (1983). Practice improves adults’ sensitivity to diagonals. Vision Research, 23(5), 547–550. McGraw, P. V., Webb, B. S., & Moore, D. R. (2009). Introduction. Sensory learning: from neural mechanisms to rehabilitation. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 364(1515), 279–283. McKee, S. P., & Westheimer, G. (1978). Improvement in vernier acuity with practice. Percept Psychophys, 24(3), 258–262. Mednick, S. C., Arman, A. C., & Boynton, G. M. (2005). The time course and specificity of perceptual deterioration. Proceedings of the National Academy of Sciences of the United States of America, 102(10), 3881–3885. Mednick, S. C., Nakayama, K., Cantero, J. L., Atienza, M., Levin, A. A., Pathak, N., et al. (2002). The restorative effect of naps on perceptual deterioration. Nature Neuroscience, 5(7), 677–681. Mednick, S., Nakayama, K., & Stickgold, R. (2003). Sleep-dependent learning: A nap is as good as a night. Nature Neuroscience, 6(7), 697–698. Mitchell, D. E., & Sengpiel, F. (2009). Neural mechanisms of recovery following early visual deprivation. Philosophical Transactions of the Royal Society B: Biological Sciences, 364(1515), 383–398. Mollon, J. D., & Danilova, M. V. (1996). Three remarks on perceptual learning. Spatial Vision, 10(1), 51–58. Morishita, H., & Hensch, T. K. (2008). Critical period revisited: Impact on vision. Current Opinion in Neurobiology, 18(1), 101–107. Nahum, M., Nelken, I., & Ahissar, M. (2010). Stimulus uncertainty and perceptual learning: Similar principles govern auditory and visual learning. Vision Research, 50(4), 391–401. Nazir, T. A., & O’Regan, J. K. (1990). Some results on translation invariance in the human visual system. Spatial Vision, 5(2), 81–100. Ofen, N., Moran, A., & Sagi, D. (2007). Effects of trial repetition in texture discrimination. Vision Research, 47(8), 1094–1102.

1565

Offen, S., Schluppeck, D., & Heeger, D. J. (2009). The role of early visual cortex in visual short-term memory and visual attention. Vision Research, 49(10), 1352–1362. Ostrovsky, Y., Andalman, A., & Sinha, P. (2006). Vision following extended congenital blindness. Psychological Science, 17(12), 1009–1014. Otto, T. U., Herzog, M. H., Fahle, M., & Zhaoping, L. (2006). Perceptual learning with spatial uncertainties. Vision Research, 46(19), 3223–3233. Pantle, A., & Sekuler, R. (1968). Size-detecting mechanisms in human vision. Science, 162(858), 1146–1148. Parkosadze, K., Otto, T. U., Malania, M., Kezeli, A., & Herzog, M. H. (2008). Perceptual learning of bisection stimuli under roving: Slow and largely specific. Journal of Vision, 8(1), 5 1–8. Pennefather, P. M., Chandna, A., Kovács, I., Polat, U., & Norcia, A. M. (1999). Contour detection threshold: Repeatability and learning with ‘contour cards’. Spatial Vision, 12(3), 257–266. Petrov, A. A., Dosher, B. A., & Lu, Z. L. (2005). The dynamics of perceptual learning: An incremental reweighting model. Psychological Review, 112(4), 715–743. Petrov, A. A., Dosher, B. A., & Lu, Z. L. (2006). Perceptual learning without feedback in non-stationary contexts: Data and model. Vision Research, 46(19), 3177–3197. Poggio, T., Fahle, M., & Edelman, S. (1992). Fast perceptual learning in visual hyperacuity. Science, 256(5059), 1018–1021. Polat, U. (2009). Making perceptual learning practical to improve visual functions. Vision Research, 49(21), 2566–2573. Polat, U., Ma-Naim, T., Belkin, M., & Sagi, D. (2004). Improving vision in adult amblyopia by perceptual learning. Proceedings of the National Academy of Sciences of the United States of America, 101(17), 6692–6697. Polat, U., & Sagi, D. (1993). Lateral interactions between spatial channels: Suppression and facilitation revealed by lateral masking experiments. Vision Research, 33(7), 993–999. Polat, U., & Sagi, D. (1994a). The architecture of perceptual spatial interactions. Vision Research, 34(1), 73–78. Polat, U., & Sagi, D. (1994b). Spatial interactions in human vision: From near to far via experience- dependent cascades of connections. Proceedings of the National Academy of Sciences of the United States of America, 91(4), 1206–1209. Polat, U., Sagi, D., & Norcia, A. M. (1997). Abnormal long-range spatial interactions in amblyopia. Vision Research, 37(6), 737–744. Pourtois, G., Rauss, K. S., Vuilleumier, P., & Schwartz, S. (2008). Effects of perceptual learning on primary visual cortex activity in humans. Vision Research, 48(1), 55–62. Preminger, S., Blumenfeld, B., Sagi, D., & Tsodyks, M. (2009). Mapping dynamic memories of gradually changing objects. Proceedings of the National Academy of Sciences of the United States of America, 106(13), 5371–5376. Preminger, S., Sagi, D., & Tsodyks, M. (2007). The effects of perceptual history on memory of visual objects. Vision Research, 47(7), 965–973. Ramachandran, V. S., & Braddick, O. (1973). Orientation-specific learning in stereopsis. Perception, 2(3), 371–376. Regan, D., & Beverley, K. I. (1985). Postadaptation orientation discrimination. Journal of the Optical Society of America A – Optics and Imagescience, 2(2), 147–155. Rockland, K. S., & Lund, J. S. (1983). Intrinsic laminar lattice connections in primate visual cortex. Journal of Comparative Neurology, 216(3), 303–318. Roelfsema, P. R., van Ooyen, A., & Watanabe, T. (2010). Perceptual learning rules based on reinforcers and attention. Trends in Cognitive Sciences, 14(2), 64–71. Rubenstein, B. S., & Sagi, D. (1990). Spatial variability as a limiting factor in texturediscrimination tasks: Implications for performance asymmetries. Journal of the Optical Society of America A, 7(9), 1632–1643. Rubin, N., Nakayama, K., & Shapley, R. (1997). Abrupt learning and retinal size specificity in illusory-contour perception. Current Biology, 7(7), 461–467. Sagi, D. (1988). The combination of spatial frequency and orientation is effortlessly perceived. Perception and Psychophysics, 43(6), 601–603. Sagi, D. (1990). Detection of an orientation singularity in Gabor textures: Effect of signal density and spatial-frequency. Vision Research, 30(9), 1377–1388. Sagi, Y., Tavor, I., Sasson, E., Pasternak, O., & Assaf (2009). Learning induced structural plasticity in humans using diffusion MRI. NeuroImage, 47, S39–S41. Sagi, D., & Julesz, B. (1987). Short-range limitation on detection of feature differences. Spatial Vision, 2(1), 39–49. Sagi, D., Kovács, I., & Racsmany, M. (2009). Preface. Learning & Perception, 1(1), 1–2. Sale, A., Berardi, N., & Maffei, L. (2009). Enrich the environment to empower the brain. Trends in Neurosciences, 32(4), 233–239. Sale, A., Maya Vetencourt, J. F., Medini, P., Cenni, M. C., Baroncelli, L., De Pasquale, R., et al. (2007). Environmental enrichment in adulthood promotes amblyopia recovery through a reduction of intracortical inhibition. Nature Neuroscience, 10(6), 679–681. Sasaki, Y., Nanez, J. E., & Watanabe, T. (2010). Advances in visual perceptual learning and plasticity. Nature Reviews Neuroscience, 11(1), 53–60. Scholz, J., Klein, M. C., Behrens, T. E., & Johansen-Berg, H. (2009). Training induces changes in white-matter architecture. Nature Neuroscience, 12(11), 1370–1371. Schoups, A. A., & Orban, G. A. (1996). Interocular transfer in perceptual learning of a pop-out discrimination task. Proceedings of the National Academy of Sciences of the United States of America, 93(14), 7358–7362. Schoups, A. A., Vogels, R., & Orban, G. A. (1995). Human perceptual learning in identifying the oblique orientation: Retinotopy, orientation specificity and monocularity. The Journal of physiology, 483(Pt 3), 797–810. Schoups, A., Vogels, R., Qian, N., & Orban, G. (2001). Practising orientation identification improves orientation coding in V1 neurons. Nature, 412(6846), 549–553. Schwartz, S., Maquet, P., & Frith, C. (2002). Neural correlates of perceptual learning: A functional MRI study of visual texture discrimination. Proceedings of the

1566

D. Sagi / Vision Research 51 (2011) 1552–1566

National Academy of Sciences of the United States of America, 99(26), 17137–17142. Schwarzkopf, D. S., Zhang, J., & Kourtzi, Z. (2009). Flexible learning of natural statistics in the human brain. Journal of Neurophysiology, 102(3), 1854–1867. Seitz, A. R., & Watanabe, T. (2003). Psychophysics: Is subliminal learning really passive? Nature, 422(6927), 36. Seitz, A. R., Yamagishi, N., Werner, B., Goda, N., Kawato, M., & Watanabe, T. (2005). Task-specific disruption of perceptual learning. Proceedings of the National Academy of Sciences of the United States of America, 102(41), 14895–14900. Serre, T., Oliva, A., & Poggio, T. (2007). A feedforward architecture accounts for rapid categorization. Proceedings of the National Academy of Sciences of the United States of America, 104(15), 6424–6429. Sireteanu, R., & Rettenbach, R. (2000). Perceptual learning in visual search generalizes over tasks, locations, and eyes. Vision Research, 40(21), 2925–2949. Skrandies, W., & Fahle, M. (1994). Neurophysiological correlates of perceptual learning in the human brain. Brain Topography, 7(2), 163–168. Sowden, P. T., Rose, D., & Davies, I. R. (2002). Perceptual learning of luminance contrast detection: specific for spatial frequency and retinal location but not orientation. Vision Results, 42, 1249–1258. Stettler, D. D., Yamahachi, H., Li, W., Denk, W., & Gilbert, C. D. (2006). Axons and synaptic boutons are highly dynamic in adult visual cortex. Neuron, 49(6), 877–887. Stickgold, R., James, L., & Hobson, J. A. (2000). Visual discrimination learning requires sleep after training. Nature Neuroscience, 3(12), 1237–1238. Stickgold, R., Whidbee, D., Schirmer, B., Patel, V., & Hobson, J. A. (2000). Visual discrimination task improvement: A multi-step process occurring during sleep. Journal of Cognitive Neuroscience, 12(2), 246–254. Swift, D. J., & Smith, R. A. (1983). Spatial frequency masking and Weber’s Law. Vision Research, 23(5), 495–505. Tajima, S., Watanabe, M., Imai, C., Ueno, K., Asamizuya, T., Sun, P., et al. (2010). Opposing effects of contextual surround in human early visual cortex revealed by functional magnetic resonance imaging with continuously modulated visual stimuli. Journal of Neuroscience, 30(9), 3264–3270. Tanaka, Y., & Sagi, D. (1998a). Long-lasting, long-range detection facilitation. Vision Research, 38, 2591–2599. Tanaka, Y., & Sagi, D. (1998b). A perceptual memory for low-contrast visual signals. Proceedings of the National Academy of Sciences of the United States of America, 95(21), 12729–12733. Tartaglia, E. M., Aberg, K. C., & Herzog, M. H. (2009). Perceptual learning and roving: Stimulus types and overlapping neural populations. Vision Research, 49(11), 1420–1427. Thompson, B., & Liu, Z. (2006). Learning motion discrimination with suppressed and un-suppressed MT. Vision Research, 46(13), 2110–2121. Thompson, B., Mansouri, B., Koski, L., & Hess, R. F. (2008). Brain plasticity in the adult: Modulation of function in amblyopia with rTMS. Current Biology, 18(14), 1067–1071. Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review, 34, 273–286. Tononi, G., & Cirelli, C. (2003). Sleep and synaptic homeostasis: A hypothesis. Brain Research Bulletin, 62(2), 143–150. Treisman, A. M., & Gelade, G. (1980). Feature-integration theory of attention. Cognitive Psychology, 12(1), 97–136.

Tsodyks, M., & Gilbert, C. (2004). Neural networks and perceptual learning. Nature, 431(7010), 775–781. Tsushima, Y., Seitz, A. R., & Watanabe, T. (2008). Task-irrelevant learning occurs only when the irrelevant feature is weak. Current Biology, 18(12), R516–517. Vogels, R. (2010). Mechanisms of perceptual learning in macaque cortex. Topics in Cognitive Science, 2(2), 239–250. Walker, M. P., Stickgold, R., Jolesz, F. A., & Yoo, S. S. (2005). The functional anatomy of sleep-dependent visual skill learning. Cerebral Cortex, 15(11), 1666–1675. Wallis, G., Backus, B. T., Langer, M., Huebner, G., & Bulthoff, H. (2009). Learning illumination- and orientation-invariant representations of objects through temporal association. Journal of Vision, 9(7), 6. Wandell, B. A., & Smirnakis, S. M. (2009). Plasticity and stability of visual field maps in adult primary visual cortex. Nature Reviews Neuroscience, 10(12), 873–884. Watanabe, T., Nanez, J. E., Sr., Koyama, S., Mukai, I., Liederman, J., & Sasaki, Y. (2002). Greater plasticity in lower-level than higher-level visual motion processing in a passive perceptual learning task. Nature Neuroscience, 5(10), 1003–1009. Watanabe, T., Nanez, J. E., & Sasaki, Y. (2001). Perceptual learning without perception. Nature, 413(6858), 844–848. Watson, A. B., & Robson, J. G. (1981). Discrimination at threshold: Labelled detectors in human vision. Vision Research, 21(7), 1115–1122. Westheimer, G. (2001). Is peripheral visual acuity susceptible to perceptual learning in the adult? Vision Research, 41(1), 47–52. Wiesel, T. N., & Hubel, D. H. (1963). Single-cell responses in striate cortex of kittens deprived of vision in one eye. Journal of Neurophysiology, 26, 1003–1017. Wilson, H. R., & Bergen, J. R. (1979). A four mechanism model for threshold spatial vision. Vision Research, 19(1), 19–32. Wilson, H. R., & Cowan, J. D. (1972). Excitatory and inhibitory interactions in localized populations of model neurons. Biophysical Journal, 12(1), 1–24. 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., Chang, L. H., Ni, R., Salat, D., Andersen, G., Watanabe, T., et al. (2010). Perceptual learning and changes in white matter in aged brain revealed by difusion-tensor imaging (DTI). Journal of Vision, 10(7), 912. Yotsumoto, Y., Chang, L. H., Watanabe, T., & Sasaki, Y. (2009). Interference and feature specificity in visual perceptual learning. Vision Research, 49(21), 2611–2623. Yotsumoto, Y., Sasaki, Y., Chan, P., Vasios, C. E., Bonmassar, G., Ito, N., et al. (2009). Location-specific cortical activation changes during sleep after training for perceptual learning. Current Biology, 19(15), 1278–1282. Yotsumoto, Y., Watanabe, T., & Sasaki, Y. (2008). Different dynamics of performance and brain activation in the time course of perceptual learning. Neuron, 57(6), 827–833. Yu, C., Klein, S. A., & Levi, D. M. (2004). Perceptual learning in contrast discrimination and the (minimal) role of context. Journal of Vision, 4(3), 169–182. Zhang, J. Y., Kuai, S. G., Xiao, L. Q., Klein, S. A., Levi, D. M., & Yu, C. (2008). Stimulus coding rules for perceptual learning. PLoS Biology, 6(8), e197. Zhang, T., Xiao, L. Q., Klein, S. A., Levi, D. M., & Yu, C. (2010). Decoupling location specificity from perceptual learning of orientation discrimination. Vision Research, 50(4), 368–374. Zhaoping, L. (2009). Perceptual learning of pop-out and the primary visual cortex. Learning & Perception, 1(1), 135–146.