Major, rapid performance improvements in perceptual training are often dismissed as 'task' or 'procedural' learning because they are fast and generalize within a task. We assessed the contributions of perceptual and procedural learning to improvement in an auditory tone frequency learning task in humans and found that perceptual learning accounted for between 76% and 98% of the rapid early performance improvement.
Perceptual learning refers to performance changes, brought about through practice or experience, that improve an organism's ability to respond to its environment1. A recent resurgence of interest in perceptual learning has been sparked by findings of a previously unexpected degree of plasticity in the adult brain2, correlations between trained performance and neural representations3, and successes in using perceptual learning as a tool to help ameliorate language disorders4,
5.
Demonstrating the occurrence of perceptual learning presents a number of difficulties, one of which is distinguishing it from procedural learning. Procedural learning refers to improvement in performance on a task that results from learning the response demands of the task. A rapid early phase of learning that generalizes to similar tasks has been taken by some investigators as indicative of procedural learning6. To minimize the likelihood that observed improvements in performance are due to procedural learning, many studies provide participants with pretraining to familiarize them with the procedure7,
8,
9. This practice reflects the cautious assumptions that perceptual learning is slow and is restricted to improving the precision with which specific physical characteristics of individual stimuli are represented, stored or compared. However, these assumptions have not been tested, and it may be that the rapid initial phase of performance improvement that is often observed reflects a genuine change in perception rather than an improved ability to perform the experimental task. For example, generalization of learning in binaural hearing, from interaural level difference to interaural time difference discrimination6, may reflect improvement in the ability to store and compare abstract representations of sound lateralization rather than representations of discrete differences in loudness or timing of sound arriving at each ear. Here, we considered learning that is specific to the perceptual judgment as 'perceptual' and contributions to improvement resulting from aspects of task performance not involved in making a perceptual judgment as 'procedural'.
With these definitions, we investigated the contributions of perceptual and procedural learning to early performance improvement on a two-interval, two-alternative forced choice (2I-2AFC) frequency discrimination target task ("Which tone is higher?"). We compared early performance improvement on this task in four groups of participants with different prior experience of training tasks sharing certain perceptual or procedural aspects with the target task. Group 1 trained on the target task. Group 2 trained on a different procedure—an AXB frequency discrimination task ("Was the middle tone the same pitch as the first or third tone?"). Groups 3 and 4 were trained on the same procedure (2I-2AFC), but group 3 trained on a different auditory task (intensity discrimination: "Which tone was louder?") and group 4 trained on a visual contrast discrimination task without any auditory training (for a complete description of tasks and stimuli, see Supplementary Methods online). After one block of training, all groups were tested on two blocks of the target 2I-2AFC sound frequency discrimination task.
If early rapid performance improvement largely reflects procedural learning, we would expect to see little difference between participants with similar experience of the procedure used in the target task but different experience of the perceptual judgment (groups 1, 3 and 4). In contrast, if early rapid performance improvement largely reflects perceptual learning, groups 3 and 4 would show more improvement over the test blocks than groups 1 and 2 because they had not experienced the perceptual task during training. If performance improvement reflects some kind of learning specific to auditory stimuli (familiarity with 1-kHz tones, for example), group 4 would be expected to show more learning over the test blocks than the other groups.
Frequency discrimination sensitivity (the reciprocal of discrimination thresholds; see Supplementary Methods online) in the two test blocks of the target task for the four experimental groups is shown in Figure 1a. We used sensitivity because those data were normally distributed in each block, allowing for parametric group comparisons. Groups 3 and 4, trained on intensity and visual discrimination, showed similar improvement between test blocks, whereas those trained on the two frequency discrimination tasks showed little or no performance improvement between test blocks (Fig. 1b). Sensitivity increases between test blocks can thus be divided into two sets according to whether training was on a frequency discrimination task (groups 1 and 2) or not (groups 3 and 4): a Tukey-B post hoc test identified these as homogenous subsets, so the data were combined. Participants in groups 3 and 4, trained on nonfrequency discrimination tasks, showed significant improvement across test blocks (F1,38 = 43.3, P < 0.001), whereas those in groups 1 and 2, trained on frequency discrimination tasks, did not (F1,38 = 1.42, P = 0.24). We concluded, therefore, that groups 1 and 2 received some amount of frequency-specific (perceptual) training during the training block, which groups 3 and 4 received during the first testing block. The dominant aspect of training was thus specific to frequency discrimination. Unpublished data from our group (see Supplementary Fig. 1 online) support this conclusion by showing similar results even earlier, in the first 200 trials.
Figure 1. Improvement in frequency discrimination in four groups (n = 20 in each group) with differing prior training.
(a) Mean frequency discrimination sensitivity levels. Note that the bar marked by a star is a threshold measured using an AXB paradigm rather than the 2I-2AFC paradigm used for all other sensitivity measures. (b) Mean frequency discrimination sensitivity increase between blocks. Error bars, s.e.m. Informed written consent was obtained from all listeners. The study was approved by the University of Nottingham (UK) Department of Psychology Ethics Committee. *P = 0.002; **P < 0.001 on one-sample t-tests.
We used a model to quantify the relative contributions of perceptual and procedural learning (see Supplementary Note online for a detailed description of the model). Following the suggestion that a learning curve can be thought of as a combination of procedural and perceptual learning curves10, we modeled frequency discrimination performance improvement as the linear sum of procedural and perceptual learning components. Because training groups had various experience of the procedural and perceptual aspects of the task, certain learning components were hypothesized to be the same for different groups. The equations describing this model were, however, underdetermined. Imposing the simple assumption that all parameters in the model must be 0 allowed us to calculate upper and lower limits of the procedural and perceptual learning components. We analyzed these limits taking into account the experimental error on the values of the derived components. This analysis revealed that the vast majority, between 76% and 98%, of early performance improvement was perceptual learning, with procedural learning contributing a mere 2−24%. Moreover, procedural learning accounted for a greater proportion of later (between the two test blocks) than earlier (between training and testing) performance improvement, in contrast with the common assumption that perceptual learning dominates only when procedural learning is nearing completion.
The definition of perceptual learning adopted here is quite general and is likely to incorporate more than one learning process. Two components of perceptual learning in frequency discrimination have been described previously: one that generalized across frequencies and a (smaller) one that was frequency specific9. Early frequency discrimination performance improvement (first 700 trials) that generalized across frequencies has been demonstrated11 and found to show a degree of frequency specificity to more protracted (11,000 trials) perceptual learning in the frequency range coded peripherally by temporal mechanisms7. Rapid early learning is not confined to simple stimuli (isolated pure tones) but occurs even for more complex tonal patterns12. Changes in human auditory evoked potentials have been seen in response to small frequency differences in complex tonal patterns after a mere 6 min of training13. Rapid early learning is not limited to the auditory system; improvements on a similar time scale have also been observed in visual texture segregation14 and hyperacuity15.
The early, rapid phase of learning is likely to be more important from an applied perspective. Training that improves the ability to make a certain perceptual judgment and that generalizes to nontrained circumstances could be of greater utility to training programs designed to tackle suspected auditory deficits4,
5 than the more protracted, more stimulus-specific, later phase of perceptual learning.
We have shown here that early frequency discrimination performance improvements, which in the past have been shown to generalize across frequencies, contain a significant component specific to the performance of frequency discrimination judgments. On these grounds, and on the basis of the common finding that the major proportion of total learning occurs early in training (e.g., S.A., D.J.C.H. and D.R.M., unpublished observations), we suggest that most perceptual learning happens early, within the first 500−600 trials of a frequency discrimination task (see also Supplementary Fig. 1). Studies that provide other than very brief pretraining may therefore miss interesting and dramatic early perceptual learning.
Acknowledgments We thank S. Cirstea for assistance with statistical analysis. This research was entirely supported by the Medical Research Council (UK).
0 allowed us to calculate upper and lower limits of the procedural and perceptual learning components. We analyzed these limits taking into account the experimental error on the values of the derived components. This analysis revealed that the vast majority, between 76% and 98%, of early performance improvement was perceptual learning, with procedural learning contributing a mere 2−24%. Moreover, procedural learning accounted for a greater proportion of later (between the two test blocks) than earlier (between training and testing) performance improvement, in contrast with the common assumption that perceptual learning dominates only when procedural learning is nearing completion.
