Perceptual training: A tool for both modifying the brain and exploring it

During the past 20 years it became increasingly clear that sensory areas in the brain of adult primates, including humans, retain a large degree of plasticity (reviewed in ref. 1). Consequently, experiences continuously modify our perception of the world even as adults. When perceptual experience is monitored, for example, in a training regime that includes intensive practice with a limited set of stimuli, the neuronal representations of these stimuli are gradually modified (e.g., ref. 2). But the nature of the changes that are induced by practice and the degree of modifiability of different sensory sites are still not well understood. The study of Wright and Fitzgerald (3), reported in this issue of PNAS, is an important contribution to these largely open questions. They found that the ability to discriminate between stimuli that cue for neighboring sound locations significantly improves with practice. Yet, the susceptibility to improvement seems to differ between the two prominent cues for sound localization—interaural time and intensity differences.

The first systematic study of practice-induced improvement in a simple discrimination task was conducted about 150 years ago by Volkman in the somatosensory modality. He found that within several hours of practice, the minimal distance needed for subjects to distinguish between single and double stimulation on the skin was often halved. Subsequent studies found that in some areas of the skin the threshold may be reduced to as little as 1% of its initial level by extensive practice over several weeks [see description of Tawney (26) in ref. 4]. What were the changes underlying this dramatic improvement? Boring (ref. 5, pp. 480–483) suggested that learning replaces attention in fine tuning of the decisions made between vaguely different sensations. Although at the cognitive level this view has not been radically changed (6), the ability to map decision-making mechanisms to specific neuronal sites has gone through substantial progress.

The progress in psycho-anatomical mapping was enabled by advances in our understanding of the neural substrate of stimulus representations. An important tool used as a behavioral probe to assess the site of modification is characterization of learning transfer and specificity. The rationale is the following. Different levels along the hierarchies of stimulus processing encode stimuli differently. For example, in the visual pathways, neurons in low-level areas encode small regions in the visual fields whereas neurons in higher-level areas encode large regions (7). Thus, for example, if learning to discriminate between stimuli consistently presented at a given retinal position involves modifications at low-level areas where neighboring positions are represented separately, improvement should not transfer even to adjacent untrained positions. On the other hand, if learning occurs at high-level areas, it is expected to transfer to more remote novel positions. This rationale was applied in several studies of visual learning. When detection or discrimination of an element at a given retinal position was substantially improved with practice, and improvement was specific to the trained position, this positional specificity was interpreted as evidence for changes at an early visual cortical area (e.g., refs. 8–12, although see ref. 13). Indeed, one of the surprising findings of visual training was the prevalence of relatively fine positional specificity.

Wright and Fitzgerald (3) examined the specificity of improvement in the discrimination of interaural time and intensity differences. These cues are computed in separate nuclei (medial and lateral superior olive, respectively) already at the first binaural interaction stage in the auditory pathways. Behavioral evidence indicates that these separately computed cues are combined at some higher-level stage and form an integrated spatial perception (14). The initial convergence stage precedes the primary auditory cortex, although this initial convergence (at the level of the inferior colliculus) is probably only partial (15). Thus, if learning occurred at a high-level stage where cues had already been combined, it would have been position-specific and yet would transfer between the two localization cues. Instead, learning these two cues had different characteristics, in particular with respect to learning rate. Thus, learning seems to have occurred before an integrated percept is formed. Moreover, listeners did not fully transfer learning to novel tone frequencies (from 4 to 6 kHz), although frequency was not explicitly a relevant parameter. This specificity implies that learning occurred at a level in which these frequencies (less than an octave apart) are represented separately. Although our current understanding of primates' auditory hierarchies is limited, this finding, too, supports the conclusion that learning did not occur at very high levels in the auditory pathways. For example, neurons in the cat's secondary auditory field, AII, a station higher in the auditory processing hierarchy than the primary auditory area, are quite broadly tuned compared with the relatively narrow tuning of area AI (16). Taken together, these data suggest that learning occurred at intermediate stages along the auditory pathways, perhaps at the primary auditory area and/or below.

Although learning specificity is informative and useful with respect to pinpointing the sites of underlying modifications, it poses difficulties in designing optimal training regimes. For example, how should we train a person to become an expert in a perceptual task like sound localization? Would it be more effective to train for each cue separately? Which stimulus parameters should we train for? If learning does not transfer across frequencies then one should perhaps train with the full range of frequencies, and potentially with other parameters between which the listener should learn to discriminate. One should note that stimulus specificity is found particularly when a fixed narrow set of parameters is consistently trained. Because this procedure seems to be the one yielding the greatest improvement, it may still be the most effective type of training.

Is there a way to choose or select the site of neuronal modifications and consequently affect the degree of generalization? Learning typically begins at generalizing high-level sites, as evident from findings that initial learning transfers quite a bit to novel stimuli (refs. 17–19 and see also discussion in ref. 3). This initial stage seems to involve perceiving the “gist” of the task. This stage is crucial to the process of guiding subsequent improvement. If the trained participants fail to observe (18) or sense (20) the required discriminations, even substantial training will not lead to improvement.

Modifications underlying subsequent learning probably occur at lower-level areas. Evidence suggests that the selection of these areas depends on the training procedure (18). Procedure affects participants' attention and attention dictates learning specificity (21). When participants attend a broad range of parameters, their discrimination abilities are improved within this entire broad range. For example, in the visual domain, if observers attend a broad area because elements are presented at two remote retinal positions, they learn discriminations also at intermediate untrained positions (22). Thus, in the auditory modality, if listeners were trained on localization discrimination with two quite different tone frequencies (e.g., 4 and 8 kHz), perhaps learning would have taken place at higher auditory areas, which may enable attention to a broad range of frequencies. In that case, learning would transfer to untrained intermediate frequencies.

Another aspect of perceptual learning found by Wright and Fitzgerald (3), which is also characteristic of visual learning, is the substantial variability between subjects. Not only does initial performance vary, but also learning rate and learning specificity greatly differs between participants (e.g., ref. 23). This variability suggests that learning occurs at somewhat different areas or neuronal populations in different individuals (18). But beyond intersubject variability, the amount and rate of learning seems different for the two localization cues. In discriminating interaural intensity differences, listeners who have had intensive practice improved substantially more than listeners who received only brief training. On the other hand, in discrimination of interaural time differences, there was no clear difference in improvement between participants with prolonged compared with brief practice. In both groups, some listeners improved whereas others failed to show clear improvement. The different patterns of learning may reflect a difference in the susceptibility of sites coding for intensity and time difference, respectively, to training induced modifications. Deciphering whether the difference stems from different accessibility to attentional mechanisms or from different degrees of plasticity within lower-level mechanisms, may require more direct physiological assessments.

Another aspect of perceptual learning found by Wright and Fitzgerald is the substantial variability between subjects.

The finding of the different learning patterns, so that one cue (interaural time difference) seems to be learned substantially faster than the other (by most, although not all listeners), is related to another open question in the field of perceptual learning. Under some conditions learning proceeds very fast. Perceptual learning is occasionally almost as abrupt as explicit, insight phenomena (e.g., ref. 24), although this is not the typical case. Often substantial improvement is found within hundreds of trials (e.g., ref. 25) and further practice does not seem to induce subsequent changes. Yet, typically, attaining an asymptotic level performance requires thousands of trials (e.g., refs. 9–12 and 21). What makes the difference? Is it the difference in training regime so that an effective training procedure yields fast learning whereas an ineffective one requires long practice before a substantial change is found? The findings of Wright and Fitzgerald (3) imply that there must be more than training procedure, because the training procedures were similar (although not identical) for these two cues. The difference is not between simple and complex tasks either because the two interaural cues seem similar in complexity. The crucial factor determining the rate of perceptual learning remains a puzzle.

Perceptual learning paradigms are becoming a tool for both changing neuronal representations and exploring them. The first is intuitive. Training is aimed at modifying abilities. The second stems from our view that learning specificity reveals the nature of the modified neuronal representations. Understanding the nature of the relevant representations may in turn prove useful in guiding the design of training regimes to be more efficient and more effective.