Perceptual learning

The term perceptual learning refers to the process of long lasting improvement in performing perceptual (visual, auditory, tactile, olfactory or taste) tasks as a function of experience. Such tasks range from simple discriminations along a single dimension (e.g. orientation discrimination in the visual modality, frequency discrimination in the auditory modality) to complex categorizations, which typically involve the integration of several dimensions (e.g. an X-ray image, with or without a tumor). The ability to improve in such tasks is retained throughout life.

Early observations
The fact that with huge amounts of practice individuals can reach impressive perceptual expertise, whether in wine tasting, fabric evaluation or musical preferences, was well acknowledged for centuries, along with the prevalent idiom that "practice makes perfect". Indeed perceptual learning was one of the first fields of systematic behavioral studies. The first documented report, in the middle of the 19th century, was of tactile training aimed to decrease the minimal distance at which individuals can discriminate whether one or two points on their skin were touched. It was found that this distance (JND, Just Noticeable Difference) decreases dramatically with practice, and that this improvement is, at least partially, retained for subsequent days. Moreover, this improvement is at least partially specific to the trained skin area. A particularly dramatic improvement was found for skin positions at which initial discrimination was very crude (e.g. on the back), though training could not bring JND of initially crude areas to that of initially accurate ones (e.g. finger tips).

This improvement raised the question, which is still debated today, to what extent it stems from peripheral modifications compared with improvement in higher-level readout stages. Early interpretations, such as that suggested by William James, attributed it to higher-level categorization mechanisms whereby initially blurred differences are gradually associated with distinctively different labels.

The role of attention
William James (Principles of Psychology, 1890) presented an extreme view asserting that "My experience is what I agree to attend to. Only those items which I notice shape my mind — without selective interest, experience is an utter chaos." His view was extreme, yet its gist was largely supported by subsequent behavioral and physiological studies. Mere exposure does not seem to suffice for acquiring expertise.

Indeed, a relevant signal in a given behavioral condition may be considered as noise in another. For example, when presented with two similar stimuli, one may prefer to increase the difference between their representations and improve the ability to discriminate between them, or to attend the similarities and improve the ability to identify both as belonging to the same category. A particular difference between them will be considered as signal in the first case and as noise in the second case.

However, recent studies suggest that perceptual learning occurs without selective attention. Studies of such task-irrelevant perceptual learning (TIPL) show that degree of TIPL is similar to that found through direct training procedures. TIPL for a stimulus depends on the relationship between that stimulus and important task events or upon stimulus reward contingencies. It has thus been suggested that learning (of task irrelevant stimuli) is contingent upon spatially diffusive learning signals. Similar effects, but upon a shorter time scale, have been found for memory processes and in some cases is called attentional boosting. Thus, when an important (alerting) event occurs, learning may also affect concurrent, non-attended and non-salient stimuli.

The impact of training protocol and the dynamics of learning
Ivan Pavlov discovered conditioning. He found that when a stimulus (e.g. sound) is immediately followed by food for several times, the mere presentation of this stimulus would subsequently elicit saliva in a dog's mouth. He further found that when he used a differential protocol, by consistently presenting food after one stimulus while not presenting food after another stimulus, dogs were quickly conditioned to selectively salivate in response to the rewarded one. He then asked whether this protocol could be used to increase perceptual discrimination, by differentially rewarding two very similar stimuli (e.g. tones with similar frequency). However, he found that differential conditioning was not effective.

Pavlov's studies were followed by many training studies which found that an effective way to increase perceptual resolution is to begin with a large difference along the required dimension and gradually proceed to small differences along this dimension. This easy-to-difficult transfer was termed "transfer along a continuum".

These studies showed that the dynamics of learning depend on the training protocol, rather than on the total amount of practice. Moreover, it seems that the strategy implicitly chosen for learning is highly sensitive to the choice of the first few trials during which the system tries to identify the relevant cues.

Several studies asked whether learning takes place during the practice sessions or in between, for example, during subsequent sleep. The dynamics of learning is hard to evaluate since the directly measured parameter is performance, which is affected by both learning, inducing improvement, and fatigue, which hampers performance. Current studies suggest that sleep contributes to improved and durable learning effects, by further strengthening connections in the absence of continued practice. Both slow-wave and REM (rapid eye movement) stages of sleep may contribute to this process, via not-yet-understood mechanisms.

Recent studies – the relations between behavioral improvement and brain sites of modification
The study of learning in general and perceptual learning in particular were largely abandoned between the 1950s and 1980s, since the scientific community tended to underestimate the impact of learning compared with innate mechanisms. During the 1980s, two independent lines of research gradually modified this view. The trend began with earlier findings of Hubel and Wiesel that perceptual representations at sensory areas of the cortex are substantially modified during a short ("critical") period immediately following birth. Merzenich, Kaas and colleagues showed that though neuroplasticity is diminished, it is not eliminated when the critical period ends. Thus, when the external pattern of stimulation is substantially modified, neuronal representations in lower-level (e.g. primary) sensory areas are also modified. In parallel, a series of behavioral studies found that, at least under some training conditions, improvement in simple discrimination tasks is large and has characteristics of primary (or secondary) sensory areas. For example, visual training on discrimination between similarly oriented short bars presented at a given retinal position with a consistent orientation, improves perceptual resolution at the trained position around the trained orientation. Even a small shift in retinal position yields a significant degradation in performance towards pre-trained levels. Since at early visual areas neuronal responses are position selective whereas at higher visual areas (located more anteriorily) responses are more broadly tuned, the position specificity was interpreted as indicating modifications that involve lower visual areas. Some recent human imaging studies (e.g. fMRI) also find modified BOLD (the signal measured in fMRI) responses in primary sensory areas, which are apparent only for the trained retinal area. Together these studies suggest that training on simple tasks may involve changes at early cortical areas, expressed as changes of basic representations of the external world.

However, these impressive stimulus specificities do not characterize all participants. Nor do they characterize all training paradigms that use simple stimuli. A no-less-prominent characteristic of these studies is the large degree of inter-lab and inter-subject variability in the degree of learning specificity. This variability is hard to reconcile solely on the basis of variability between parameters of primary sensory areas among different individuals. It suggests that even in these simple tasks, different individuals implicitly use different strategies. In line with this interpretation are the findings that slightly different protocols yield significantly different learning rates and different asymptotic performance.

Recent theories
Several recent theories attempted to account for the behavioral findings at different levels, from large scale networks to single synaptic connections. This entry focuses on the large scale. Most interpretations of the stimulus selectivity described above suggest a genuine change in representations at sensory cortical areas, which could be recorded even under passive conditions. Gilbert and colleagues suggested that though neuronal selectivity at sensory areas is modified by training, this modification is only apparent when the task is performed. Dosher & Lu suggested that training increases the weighting of lower-level representations according to their relevance to the task. Thus, modifications occur at the output level, at a task specific readout stage.

The Reverse Hierarchy Theory (RHT), proposed by Ahissar & Hochstein, aims to link between learning dynamics and specificity and the underlying neuronal sites. RHT proposes that naïve performance is based on responses at high-level cortical areas, where crude, categorical level representations of the environment are represented. Hence initial learning stages involve understanding global aspects of the task. Subsequent practice may yield better perceptual resolution as a consequence of accessing lower-level information via the feedback connections going from high to low levels. Accessing the relevant low-level representations requires a backward search during which informative input populations of neurons in the low level are allocated. Hence, subsequent learning and its specificity reflect the resolution of lower levels. RHT thus proposes that initial performance is limited by the high-level resolution whereas post-training performance is limited by the resolution at low levels. Since high-level representations of different individuals differ due to their prior experience, their initial learning patterns may differ. Several imaging studies are in line with this interpretation, finding that initial performance is correlated with average (BOLD) responses at higher-level areas whereas subsequent performance is more correlated with activity at lower-level areas. RHT proposes that modifications at low levels will occur only when the backward search (from high to low levels of processing) is successful. Such success requires that the backward search will "know" which neurons in the lower level are informative. This "knowledge" is gained by training repeatedly on a limited set of stimuli, such that the same lower-level neuronal populations are informative during several trials. Recent studies found that mixing a broad range of stimuli may also yield effective learning if these stimuli are clearly perceived as different, or, are explicitly tagged as different. These findings further support the requirement for top-down guidance in order to obtain effective learning.

Perceptual expertise-enrichment versus differentiation
In some complex perceptual tasks all humans are experts. We are all experts at scene identification, face identification and speech perception. Traditional explanations attribute these expertises to some kind of holistic, somewhat specialized, mechanisms. Perhaps such quick identifications are achieved by more specific and complex perceptual detectors which gradually "chunk" (i.e. unitize) features that tend to concur, making it easier to pull a whole set of information. Whether any concurrence of features can gradually be chunked with practice or chunking can only be obtained with some pre-disposition (e.g. faces, phonological categories) is an open question. Current findings suggest that such expertises are correlated with a significant increase in the cortical volume involved in these processes. Thus, we all have somewhat specialized face areas, which may reveal an innate property, but we also develop somewhat specialized areas for written words as opposed to single letters or strings of letter-like symbols. Moreover, special experts in a given domain have larger cortical areas involved in that domain. Thus, expert musicians have larger auditory areas. These observations are in line with traditional theories of enrichment proposing that improved performance involves an increase in cortical representation. For these expertises, basic categorical identification may be based on enriched and detailed representations, located to some extent at specialized brain areas. Physiological evidence suggests that training for refined discrimination along basic dimensions (e.g. frequency in the auditory modality) also increases the representation of the trained parameters, though in these cases the increase may mainly involve lower-level sensory areas.

Clearly, increased representation areas cannot suffice for learning. The rest of the brain needs to learn to interpret these signals correctly.

The effect of lab training on performance in real-life
An important potential application of perceptual learning is the acquisition of skill for practical purposes. Towards that goal it is important to understand whether training for increased resolution in lab conditions induces a general upgrade which transfers to other environmental contexts, or results from mechanisms which are context specific. In the 1940s, Eleanor J. Gibson made the claim that lab-based training did not improve pilots' skill when landing under difficult visual conditions. Indeed, improving complex skills is typically gained by training under complex simulation conditions rather than one component at a time. Recent lab-based training protocols with complex action computer games have shown that such practice indeed modifies visual skills in a general way, which transfers to new visual contexts. An important characteristic is the functional increase in the size of the effective visual field (within which viewers can identify objects), which is trained in action games and transfers to new settings. Whether learning of simple discriminations, which are trained in separation, transfers to new stimulus contexts (e.g. complex stimulus conditions) is still an open question.

Perceptual learning and other forms of learning
Lab based experiments typically dissociate between the sensory and the motor aspects of perception. Typically, participants are asked to make a response that is not the natural response they are used to make to the stimuli. Furthermore, the term attention is typically associated with a covert motor response. Under natural conditions, performance is better understood in sensory-motor loops, rather than in two separate routes, bottom-up and top-down, which most lab-based studies refer to.

Perceptual and sensory-motor learning probably share mechanisms with other forms of learning, which were traditionally viewed as cognitive. This entry emphasized the important role of selective attention. But other aspects are also shared. For example, recent findings show quickly induced, long-lasting insights in perception. This phenomenon, traditionally thought to characterize only problem-solving tasks, suggests that part of the traditional segregation between perceptual and problem-solving learning results from the use of different experimental paradigms rather than from inherently different underlying mechanisms.