The performance of humans with perceptual tasks can be improved by repetition. In recent years, much effort has been invested in understanding the characteristics of perceptual learning (Dosher & Lu, 1999; Fahle & Poggio, 2002; Gilbert, 1994; Hochstein & Ahissar, 2002; Karni & Bertini, 1997; Sagi & Tanne, 1994). Various frameworks were developed to isolate different systems of perceptual learning, with emphasis given to different components in the hypothesized information processing system. A classification of learning processes according to the time scale of learning revealed the existence of a learning process with a fast time scale (min) being less selective for stimulus features, and a process with a slower time scale (days) being more selective for stimulus features (Karni & Sagi, 1993). The differences in selectivity to stimulus features were taken to suggest that the fast process is task driven, optimizing extra-stimulus aspects of the task, and that the slow process is stimulus driven, improving stimulus processing in sensory areas of the brain. Both processes were assumed to depend on the presence of both a stimulus and a task, basing on results showing that learning is specific to those stimulus aspects that are relevant to the task (Ahissar & Hochstein, 1993; Karni & Sagi, 1995; Seitz & Watanabe, 2003). Here we attempt to better dichotomize the stimulus/task aspects of learning by suggesting a new classification of perceptual learning, basing on the statistical information available to the observer.

According to signal detection theory (SDT), performance on detection and discrimination tasks depends on two complementary processes: a sensory process and a decisional process. The sensory process is activated by the stimuli encountered during task performance; the decisional process is thought to mediate a response by having access to contextual information related to the task as well as to the sensory information. Contextual information may include accumulated knowledge about the frequency of occurrence of the stimuli, and the required stimulus-response mapping, thus, can be viewed as task related according to the stimulus/task classification (Sagi & Tanne, 1994). Both the “task” and the “decision” processing components need to be better defined to allow for a detailed mapping between the stimulus/task classification and the stimulus/decision classification. Here it is sufficient to note that both the “task” and the “decision” are goal directed, being dependent on the current goal of the system. Earlier views considered sensory and goal-directed processes as a hierarchy of processing modules. However, recent results from electrophysiology (Ito, Westheimer, & Gilbert, 1998; Lee, 2002; Lee & Mumford, 2003; Li, Piech, & Gilbert, 2004), psychophysics (Bar, 2003; Freeman, Sagi, & Driver, 2001; Juan & Walsh, 2003), and theoretical neuroscience (Bar, 2003; Lee & Mumford, 2003; Mumford, 1992; Ullman, 1995) point to a strong coupling between bottom up and top down influences.

In the current study, we introduced a method to detect decision-driven (statistical) learning processes. Such learning is viewed as modification of decision strategies that were developed to take advantage of stimulus regularities at a given behavioral context. In our method, the same stimuli and the same task are used in both the training and the post-training sessions but the probability that the practiced stimuli will appear is changed. In the post-training task, the different stimuli are randomly interleaved within a test session so that the precise stimulus that is presented on a given trial cannot be predicted (mixed by trial method). When the mixed-by-trial (MBT) method is used, the ability of the observers to retain their learning results is tested under conditions of increasing stimulus uncertainty. We hypothesize that processing modules that are minimally affected by extra-retinal features (i.e., as the probability of stimulus occurrence) will retain their level of improved performance in conditions of stimulus uncertainty. However, processing modules that make use of contextual/statistical information available to the observer will fail under uncertainty. Thus, a failure to retain the learning results when stimulus statistics is changed implies learning by way of a change in the decision strategy. Here this method is applied to explore the learning of contrast discrimination.

Contrast discrimination is a basic visual task whereby the observers report which of two stimuli appears to have a higher contrast. Performance with contrast discrimination is measured as the contrast threshold for detecting a contrast increment at various contrast-base levels (pedestals). Contrast discrimination was shown to be stable across repetitions when practiced on a wide range of contrasts (Dorais & Sagi, 1997; Tsodyks, Adini, & Sagi, 2004; Zenger & Sagi, 2002). Improved performance has been reported for conditions where (i) the target is flanked by nearby maskers during practice (Adini, Sagi, & Tsodyks, 2002), and (ii) when a single (fixed) contrast level is practiced (Yu, Klein, & Levi, 2004). Both methods (i and ii) lead to similar improvements in the threshold for contrast discrimination at 0.5 base contrast; however, do they reflect the same learning mechanisms? To resolve this issue, we used the MBT method to compare the performance of the two procedures under conditions of contrast uncertainty. The main results reported here were presented in the 2003 Annual Vision Sciences Society Meetings in Sarasota, Florida.

We conducted contrast discrimination experiments under different experimental conditions, using a temporal two-alternative forced-choice (2AFC) procedure. We measured the minimal difference in contrast needed to discriminate two foveally centered Gabor signals (GSs) that differed only in their contrast. The results were described by a threshold versus contrast (TvC) function (Legge, 1981; Legge & Foley, 1980). Methodological details relevant to specific experiments are described in the sections related to these experiments.

A temporal 2AFC procedure was used. To measure contrast discrimination threshold, each trial consisted of two stimuli presented sequentially: One of the frames contained a Gabor stimulus of a certain base contrast (a pedestal of contrast C_b ); the other frame contained the sum of the pedestal and the target signal (target contrast C_t , with a total stimulus contrast of C_b + C_t ) (Figure 1a). Before each trial, a small white fixation circle was presented at the center of the screen. The observers, when ready, pressed a key to activate the trial sequence, which consisted of (1) a no stimulus interval (500 msec), (2) a stimulus presentation (90 msec), (3) a no stimulus interval (1,000 msec), and (4) a second presentation (90 msec). We measured the just noticeable difference in contrast between the two stimuli (threshold value of C_t ). The observers reported which of the stimuli appeared to have a higher contrast (that is, which of the stimuli contained the target). Pressing a key indicated the decision, and an auditory feedback signal was given for an incorrect response. A staircase method was used to determine the contrast threshold at a level of 79% correctness (Levitt, 1971): The contrast of the target was increased by 0.1 log units after every incorrect response and decreased by 0.1 log units after three consecutive correct responses. A block was terminated after 8 reversals of the staircase procedure (which were approximately 50 trials), and the geometrical mean of the last 6 reversal values was used as a threshold estimate. All staircases started with a high target contrast that enabled error-free detection or discrimination. The time sequence of the experiment was designed to prevent contrast adaptation (Hammett & Snowden, 1995; Magnussen & Greenlee, 1986); thus, we used a relatively long ISI. Also, we used a relatively small number of reversals per block, (8 reversals, as compared with, e.g., 12 reversals in Yu et al., 2004, and 15 reversals in Sowden, Rose, & Davies, 2002), and a more than 15-min break every 10 blocks to reduce effects due to fatigue.

To measure the TvC function of the observers, we used seven different base contrasts (0, 0.03, 0.05, 0.06, 0.12, 0.25, and 0.51) in each session, which started and ended with a measurement of the contrast-detection threshold (C_b = 0). The target contrast (C_t ) at the beginning of the staircase of each experimental condition was above the estimated threshold for that condition. That is, for each base contrast, 0 < C_b < 0.51, we started from a Weber's fraction C_t /C_b > 1. Because of technical limitations, the base contrast 0.51 was an exception: We started the staircase with a target contrast of 0.28, having a Weber's fraction of 0.28/0.51 = 0.55. With base contrasts 0-0.12, we started the staircase with a target contrast of 0.20 or 0.24 (depending on the observer). The starting target contrast for C_b = 0.25 was C_t = 0.28. A session lasted about 30-45 min. Observers participated in up to two successive sessions on a single day. In experiments where two sessions were practiced on the same day, a 15-to-40-min break separated the sessions. Within a session, base contrasts were ordered according to two methods:

Stimuli were displayed as gray-level modulation on a Mitsubishi Diamond Pro 2060u monitor, using a PC computer with an Intel Pentium IV processor. The video monitor had 1600 × 1200 pixels occupying an 18.6º × 13.7º area. The mean display luminance was 30 cd/m² in an otherwise dark environment. The stimuli were viewed from a distance of 125 cm.

The stimuli consisted of one target signal (at the fixation point), and one pedestal signal (at the target location). In some of the experiments, we also had 2, 4, 6, 8, or 10 flankers that differed from the target and pedestal, only in their location and contrast (Figure 1). The spatial luminance distribution of each of the target, pedestal, and flanker signals was described by a Gabor function, a cosine grating multiplied by a Gaussian envelope (Daugman, 1985), with a vertical orientation, and σ = λ = 0.15º. We define contrast as A/I₀, where A is the amplitude of the cosine function generating the Gabor function and I₀ is the screen mean luminance. The contrast of the flankers, C_f , was 0.3. The contrast of the pedestal signal (the base contrast), C_b , was changed according to the different experimental conditions. In experiments where flankers were used, the target and flankers were collinearly aligned with 2λ (0.3º) spacing. The target/nontarget frames in the 2AFC procedure were marked by four white crosses that were placed at ΔX = ±4.6º, ΔY= ± 4.6º relative to the fixation; the size of the crosses was 80 pixels at the initial, introductory session. In the other sessions, we reduced the size of the crosses to 36 pixels each.

The observers were high-school or undergraduate students with normal or corrected-to-normal vision, and were naive as to the purpose of the experiments. None had any previous experience with a psychophysical experiment. In this study, observers participated in several experiments as described in detail in the relevant sections.

Experiments 1 and 2 tested the absence of learning by repetitions in the contrast discrimination task, in the multiple-contrast experiment, for both the blocked and the MBT practice. Experiments 3-6 tested the effect of various practice methods on the TvC curves; the MBT method was used to measure the pre- and post-training thresholds. Post-training thresholds were recorded from only one session (the first session with the MBT method that followed the training period) to avoid bias due to re-learning or extinction of learning.

Six observers participated in this experiment; each practiced 4-5 daily sessions of the multiple-contrast CD task, using the blocked method. All observers participated in an introductory session, on the first day that they came to the lab. In this session, their contrast-detection threshold was measured 9 times, in 9 separate blocks of trials. This was done to eliminate the possibility that the resulting perceptual learning (if found) reflects general purpose learning, as learning the time sequence of the experiment, the rules for correct response, and learning of the necessary motor skills for the task (correct key press). In previous studies of this type, some of our new observers experienced some difficulties in detecting the time interval of the target/nontarget frames, in particular when the stimuli (target and/or pedestal) were at low contrast. To eliminate the possibility that our learning effects included the learning of the stimulus timing, we marked the stimulus frames with large crosses (80 pixels each) during the first 8 blocks of the initial phase. In the last block of this session, the size of the crosses was reduced to 36 pixels each, verifying a similar performance level. The measurements of the TvC curves followed this session.

We compared the TvC curve that was obtained on the first day of practice to the curve obtained on the fourth day. Figure 2 shows the TvC curves from the practice sessions for six observers, and the average TvC curves that were obtained on the first and fourth days of practice (across six observers). As can be seen from Figure 2, the TvC curves that were measured on the first and fourth days of practice do not differ significantly.

Five observers participated in this experiment; each practiced 4-5 daily sessions of the contrast discrimination task with multiple contrasts using the MBT method. Figure 3 shows the TvC curves for these practice sessions for five observers, and the average TvC curves that were obtained on the first and fourth days of practice (across five observers). All the observers had at least two sessions of this task with the blocked method, before practicing the task using the MBT method. As can be seen from Figure 3, no significant difference was found between the thresholds that were measured on the first and fourth days of practice. However, the slope of the TvC curve decreased from 0.63 on the first day to 0.44 on the fourth day (with the slope of the difference curve being 0.194, SE = 0.052, F = 13.87, df = 3, p < .05, linear regression analysis of log(ΔC_t) vs. log(C), slope range between base contrasts 0.05 and 0.5). This barely significant change reflects a small increase (0.13-0.09 log units; that is, 20-35% increase) in the threshold for contrast discrimination in the low range of base contrasts (between 0.05 and 0.12,) and a small decrease in the threshold for discrimination at pedestal contrasts 0.25 and 0.50 (0.05-0.07 log units; that is, ~15%). This change of slope may suggest that although there was no significant change in each of the individual thresholds, some sort of learning effect did take place while practicing the task using the MBT method. This learning can be explained by a change in decision strategy. It is possible that when observers are uncertain about stimulus contrast, they optimize decision only for a limited contrast range, with this range initially set to low contrasts and after some practice shifts to higher contrast. Alternatively, it is possible that the contrast transducer function had changed.

Surprisingly, on the average, the experimental procedure that was used to measure the TvC curves had no significant effect on the thresholds for discrimination.  Figure 4 compares the average TvC curves obtained with the blocked method (black) versus the MBT method (red). Each of the three observers practiced four sessions of the CD task with the blocked method, followed by another four sessions of the task with the MBT method. As can be seen in Figure 4, no significant difference was found between the thresholds for contrast discrimination that were measured in the two methods, in any of the base contrasts. The results of these experiments (1 and 2) confirm earlier findings showing that contrast discrimination performance cannot be improved by practicing the task with a wide range of contrasts (Dorais & Sagi, 1997; Tsodyks et al., 2004; Zenger & Sagi, 2002).

After confirming the absence of learning in the above procedures, we examined how several learning procedures affected the TvC curves. The contrast interleaved method (MBT) was used to measure the pre- and the post-training curves.

The results obtained in Experiment 2 showed almost identical TvC curves for the blocked and the mixed (MBT) contrast methods. However, in Experiment 2 the MBT measurements were preceded by the blocked measurements. Here we checked whether the order of practice methods (Blocked before MBT) affected the result. Three observers without previous experience in contrast discrimination experiments performed this task. Each started the experiment by practicing three sessions of the multiple contrasts CD task with the MBT method. Following this initial stage, they practiced three sessions with the blocked method, and then were re-tested using the MBT method. We checked for (1) learning effects during the initial stage of testing (CD task with the MBT method, as in Experiment 2), (2) learning effects with the blocked method (as in Experiment 1), and (3) a change between the pre- and post-training TvC curves obtained with the MBT method. To test the learning effect in each of the tested procedures, we compared the first day of practice with the third day of practice. No significant change was found with any of the tested base contrasts following the pretraining sessions, either with the MBT method (Figure 5a) or with the blocked method (Figure 5b), in agreement with the results of Experiments 1 and 2.

Interestingly, although we found no learning effect with either of the two methods, we did find a significantly improved threshold for discrimination with pedestal contrast 0.5, using the MBT method, following practice with the blocked method (Figure 6). This effect was mostly due to one observer (BN), who had a very high threshold with pedestal contrast 0.5 when tested with the MBT method. However, after he was exposed to this pedestal contrast in the blocked method, he improved dramatically (0.45 log units).

The results presented in Figure 6 provide some evidence, though not too strong, for the existence of a fast-learning effect during practice with the multi-contrast CD task, using the blocked method (see Discussion). The present results, however, do not enable an accurate estimation of the time involved, and we assume it to be less than one practice session (Figure 6b). To separate this fast-learning effect from the other possible learning processes, all the observers in Experiments 4-6 (except observer SR) had at least two sessions of the contrast discrimination task with the blocked method, before being tested with other methods.

Recently Yu et al. (2004) reported learning effects in CD using a smaller range of base contrasts. In a previosus study (Adini et al., 2002), we failed to find such an effect, and here we thought to re-examine this issue. The three observers who participated in Experiment 3 continued the practice with four out of the seven base contrasts for another eight days. In each session they practiced the discrimination of the set of four contrasts twice, using the blocked method. In the last block of each session, we re-measured the threshold for detection (the sequence stimulus contrasts during a session were 0, 0.12, 0.25, 0.51, 0, 0.12, 0.25, 0.51, and 0). During the first four days of the practice, the observers had one session of practice each day they came to the lab. These sessions did not produce any learning effect, thus the observers were given two sessions of practice per day during the following four days of practice, with a 15-30-min break in between. We checked whether there was a learning effect during practice with the partial range method, and re-measured the multi-contrasts CD task, using the contrast interleaved method (MBT), to check if the practice led to a perceptual learning effect. No learning effect was found throughout the practice with the partial range of four contrasts. Figure 7 shows the average thresholds that were measured on the first day of the practice with four contrasts (green triangles up), and the average of the thresholds that were taken on the last day of this practice (violet triangle pointing down). Moreover, no change was found between the pretraining (black solid lines) and the post-training (red dotted lines) TvC curve. Interestingly, on the average (N = 3), the observers showed a small decrease in the threshold of discrimination with the pedestals that were not in the partial set that was practiced.

This experiment was motivated by the findings of Yu et al. (2004), showing a learning effect in CD when using a single, fixed contrast pedestal. Six observers practiced contrast discrimination using a single, fixed base contrast (C_b = 0.5). They practiced the task for 3-5 days, with two sessions daily, separated by a break of 10-30 min. Each of the two daily sessions contained 9-10 blocks (~40-50 trials per block) and lasted about 30 min. The observers practiced four-to-eight sessions of multiple-contrast CD tasks prior to the practice with a single contrast. We checked for a learning effect when the single contrast was practiced, and for a difference between the pre- and post-training TvC curves (using both the blocked method and the MBT method). We found that all the observers had their contrast discrimination thresholds reduced with the 0.5 pedestal, when tested with the fixed base contrast (0.5, as practiced). Figure 8 shows the threshold versus block number that was measured during three days of practice, averaged across the six observers. On the average, the threshold for CD on the last day of practice with the fixed base contrast was smaller by 0.15 ± 0.02 log units than the threshold that was measured in the context of the pretraining multiple-contrast CD experiment. This result confirms the presence of learning in CD, as found by Yu et al. (2004). Next we checked whether this learning passes the MBT test. Figure 9 and 10 compare the TvC curves measured before and after the practice with the single base contrast, using the blocked method (Figure 9) and the contrast interleaved method (MBT; Figure 10). On each graph, we marked the average performance of discrimination, which was measured on the last day of practice with the single base contrast (blue star). The data show that when the observers were tested in the multiple-contrast CD task with the blocked method, the improved performance with the practiced contrast was highly specific to the trained contrast and did not transfer to the whole TvC curve (Figure 9).

Furthermore, this specific learning effect largely disappeared when the different contrasts were randomly mixed by trial during testing (Figure 10). Interestingly, this latter post-learning test (MBT) showed some deterioration in performance with low contrast pedestals (in two out of three observers). We speculate that this effect happened because the observers had difficulties readapting their decision strategy to the MBT condition during the single session they were given. The specificity for contrast found here is in agreement with the findings of Yu et al. (2004). The failure of this learning effect to show in the uncertainty condition indicates that the improvement in performance is due to the development of a contrast-specific decision strategy and not due to changes in the transducer function.

Observers practiced the multi-contrast contrast discrimination task, as in Experiments 1 and 2, in the presence of collinear chains of Gabor signals. The flankers differed from the target only in their location and contrast (Figure 1b). The spacing between the GSs in the chain was 2λ. The flankers contrast was 0.3. Prior to the practice with context, we measured the lateral masking curve of the observers (Polat & Sagi, 1993) to establish a baseline for the context effect. The practice with the context was carried out in two cycles: In each cycle, the observers performed the multi-contrast CD task, in the presence of collinear flankers. The number of flankers varied from subsession to subsession (N = 2, 4, 6, 8, and 10 flankers). After each practice cycle, we measured a single TvC curve, using the contrast interleaved method (MBT).

We compared the TvC curves that were obtained before and after the practice with the context to test for a learning effect. The practice sessions were carried out using either the ordered blocked method (Experiment 6a) or the MBT method (Experiment 6b).

Three observers practiced CD with chains (flankers) using the blocked method. Before that, all had practiced the CD task with the blocked and the mixed-by-trial methods (four-to-eight sessions each), and with the single base contrast method (C_b = 0.51; four sessions). After the observers practiced four sessions of the fixed CD condition, we re-measured their TvC curve, using the MBT method. This measurement served us as their prepractice TvC curve for later comparisons.

Results: Figure 11 compares the pre- and post-training TvC curves for this experiment. As can be seen, after practice with chains of GSs with the blocked method, the observers showed improved contrast discrimination thresholds for all base contrasts tested with the MBT method. Note that the improvement obtained was smaller than the one reported in our previous study (Adini et al., 2002), though covering a larger range of base contrasts. Some of these differences are probably due to the different measurement methods, blocked (Adini et al., 2002) and MBT (in the present study). Context-dependent learning may also depend on the contextual stimuli used and on their variations during training. Yu et al. (2004), using a fixed-length elongated Gabor signal instead of varied length Gabor chains, failed to find such a learning effect. The detailed conditions required for this learning effect to take place are not yet clear, and the main result to note here is that the learning effect was not affected by the relatively high uncertainty in the base contrast that was introduced during the pre- and post-training tests.

The second group of three observers started the experiment by training the multi-contrast CD task using the blocked (four sessions) and the MBT method (four sessions), before practicing with the chain. This group practiced the context learning using the contrast-interleaved method (MBT), which led to a significant improvement in the threshold of discrimination with only base contrasts of 0.5 and 0.05 (Figure 12). Practice with context induced a clear learning effect with the blocked training method, which extended to all the pedestal contrasts, and a more limited effect when training was carried out under conditions of contrast uncertainty (MBT). A similar advantage of learning with the blocked method, over the MBT method, was found in Experiment 3.

In this study, several experimental procedures were used with varied stimulus uncertainty to test the dependence of learning results on uncertainty. The experimental results are summarized in Table 1. It was hypothesized that stimulus-driven and decision-driven learning mechanisms differ in their sensitivity to uncertainty. Two types of learning were identified, differing in their generalization to conditions of stimulus uncertainty. One type of learning could be executed only under stable and predictable stimulus conditions, whereas the other type generalized to conditions of stimulus uncertainty. This distinction is meaningful with respect to real-life applications, because we expect to have the gain achieved through perceptual learning also when the stimulus is encountered outside of the precisely trained context. The ability to generalize learning to conditions that do not strongly depend on the stimulus probability of appearance implies that the role of decision strategy is minimized. Learned tasks that pass the uncertainty test are expected to be of the “low-level” type, and are minimally affected by the behavioral context.

The dichotomy established here was used to detect learning processes that are driven by decision mechanisms. To this end, we tested a multiple-contrast discrimination task, in a situation where the observers could not predict the base contrast in any upcoming trial, and thus could not use contrast-specific strategies. We applied several learning methods and checked the transfer (or generalization) of the learning to this uncertainty situation. We found that when the training was carried out with a single, specific contrast (Experiment 5), the learning could be applied only when the observers could predict the upcoming trial (i.e., the whole session, and/or the whole block of trials had the same base contrast), suggesting a decision-driven learning mechanism. However, the practice procedure with the flankers led to a learning effect that passed the MBT test, and, therefore, is in accord with the stimulus-driven learning mechanism, suggested in Adini et al. (2002).

In Experiment 5, we found that all the observers performed better on CD after practicing with a fixed base contrast (C_b = 0.5), in agreement with Yu et al. (2004). The mean improvement was 0.15 log units ± 0.02 SE. The size of this learning effect is similar to the mean increase in performance (0.14 log units) that was found for the contrast-detection task by Sowden et al. (2002). Our results show a saturation of learning after the first two or three blocks of practice, with further improvement shown on the second day (Figure 8), indicating a consolidation period (Karni & Sagi, 1993). This finding is different from that of Yu et al. (2004), who found continuous improvement within each session, which was partially lost at the following session. Overall, the day-to-day improvement in the study of Yu el al. (2004) was relatively small and much of the learning effect could be attributed to some transient, within-session adaptation process. In agreement with Yu et al. (2004), we did find the learning effect to be specific to the trained contrast (Figure 10). Furthermore, our results with the MBT method showed that the learning result could not be generalized even to the same base contrast when all base contrasts were randomly interleaved in each block tested (Figure 10). We concluded that after practicing with the single base contrast, the observers were able to use contrast-specific strategies to handle the practiced contrast, and their contrast-transducer function was not affected (as indicated by the TvC curve).

An ideal measurement tool for learning should not produce learning. The MBT method fulfils this requirement, showing no significant learning in contrast discrimination (Experiment 2 and 3). This could be attributed to reduced efficiency of selection processes involved in learning when uncertainty regarding stimulus parameters exists (compare the results of Experiment 6a and 6b). Surprisingly, the added uncertainty did not affect discrimination thresholds before learning, as CD thresholds did not differ significantly between the blocked and the MBT methods in pre-learned observers (as in Experiment 2, but not after learning as in Experiment 5). This latter result was found to depend on the testing order, showing increased thresholds with the MBT method when tested before the blocked method (Experiment 3). We took this result (Figure 6) as evidence for a fast-learning process that had taken place during practice using the blocked method. This effect also saturated very quickly (within the first block/session), because no significant learning effect was seen in the data obtained with the blocked method (data presented in Figure 2, Figure 5b, and Figure 6b). The learning effect could only be observed when the observer moved from the high-uncertainty conditions (MBT) to the no-uncertainty condition (blocked), and back to the MBT method. Note that this fast-learning effect, though obtained with the blocked method, was transferred to the MBT method, suggesting that it involved modification of low-level networks. As shown in Figure 6, on the average, this learning effect is not too strong and is significant only for the highest base contrast (0.5). The literature indicates that learning processes due to improved selection are relatively fast, in particular when a “difficult” task is preceded by an easy one (Ahissar & Hochstein, 1997; Liu & Weinshall, 2000; Rubin, Nakayama, & Shapley, 1997). Thus, the effect might reflect improved selection among multiple-contrast filters (Geisler & Albrecht, 1997) during practice with blocked contrasts. On this account, the observers monitor selected contrast channels. It is possible that unpracticed observers do not use their knowledge about base contrast to weigh the contrast channels to achieve better performance, but with repeated practice, they learn to select and better weight the appropriate channels to produce a response (Lu & Dosher, 2004). Such a selection can be implemented by association between the different processing units, within or between the different levels of processing, enabling the retainment of the decision scheme for future use. Note that a strong interaction between top-down (i.e., task demands) and bottom-up modulation is involved in perceptual learning (Ahissar & Hochstein, 1993; Meinhardt, 2002; Shiu & Pashler, 1992). Thus, decision-driven perceptual learning may induce changes at the sensory level, possibly by selecting the behaviorally relevant neuronal population as the target for the learning process. This selection process, often termed “gating of learning,” is an essential factor in sensory development during the critical period (Held & Hein, 1963), and some of the neuronal correlates were recently identified (Fregnac, Shulz, Thorpe, & Bienenstock, 1988). The results of Experiment 6 are in agreement with this concept. We found a difference in the magnitude of the learning effect, depending on whether stimuli were mixed (MBT) or blocked in a session, with the blocked method producing a stronger learning effect (Figure 12 vs. Figure 11). It is possible that the gating process fails when stimulus properties are not well specified.

In summary, we introduced the MBT method to study perceptual learning. This method is based on manipulating the uncertainty level of the specific attribute that is relevant to the task (i.e., contrast). With this method, the same task and stimulus are used in both training and post-training tests. Thus, it is neither the task transfer nor the stimulus transfer that is being tested, but rather the effect of uncertainty regarding the value of the tested stimulus attribute. Learning mechanisms that rely on attribute-dependent strategy but not on low-level transducers are expected to fail when learning is tested under conditions where the stimulus attribute is randomly varied between presentations.

We thank Yuval Avivi, Yoram Bonneh, and Nitzan Censor for helpful discussions, and Tal Adini for the icon. This research was supported by the Basic Research Foundation, administrated by the Israel Academy of Science and Humanity.