Abstract
Repeatedly exercising a perceptual ability usually leads to improvement, yet it is unclear whether the mechanisms supporting the same perceptual learning could be flexibly adjusted according to the training settings. Here, we trained adult observers in an orientation-discrimination task at either a single (focused) retinal location or multiple (distributed) retinal locations. We examined the observers’ discriminability (N = 52) and bias (N = 20) in orientation perception at the trained and untrained locations. The focused and distributed training enhanced orientation discriminability by the same amount and induced a bias in perceived orientation at the trained locations. Nevertheless, the distributed training promoted location generalization of both practice effects, whereas the focused training resulted in specificity. The two training tactics also differed in long-term retention of the training effects. Our results suggest that, depending on the training settings of the same task, the same discrimination learning could differentially engage location-specific and location-invariant representations of the learned stimulus feature.
Keywords
Perceptual abilities, such as the ability to discern a slight change in stimulus orientation, can be significantly enhanced by practice. The cortical loci and information processing stages in which such perceptual learning effects occur have been debated for decades. On the one hand, by using simple and restricted training settings, early studies on visual perceptual learning (VPL) discovered long-lasting practice effects that are restricted to the trained visual-field location (Schoups et al., 1995; Shiu & Pashler, 1992). The specific learning effects are correlated with, and ascribed to, specific refinement in low-level, location-specific visual representations (Jehee et al., 2012; Jia et al., 2020; Schoups et al., 2001) as well as strengthened connectivity from early visual cortex to decision areas (Jia et al., 2020). On the other hand, it has been shown that learning transfer across locations is promoted by manipulation of training settings such as attention deployment (Donovan et al., 2015; Hung & Carrasco, 2021), task difficulty (Ahissar & Hochstein, 1997), precision demands (Jeter et al., 2009), and irrelevant tasks (Xiao et al., 2008), suggesting the involvement of higher level processes that are less location sensitive.
To offer a general account of learning transfer and specificity observed under various conditions, an overfitting hypothesis has been proposed (Sagi, 2011): Training on a very restricted stimulus set usually leads to overfitting of neural processes to the very details of stimulus settings and thus results in overspecificity of the learning effects, whereas manipulation of training procedures that can eliminate overfitting would induce learning transfer. An integrated reweighting theory with computational model (Dosher et al., 2013; Dosher & Lu, 2017) has also been proposed to reconcile the diverse observations on VPL and the viewpoints on the underlying mechanisms. According to the reweighting model, learning takes place by optimizing readout of sensory information from both low-level, location-specific representations and high-level, location-invariant representations; to what degree each of the two representation levels contributes to learning could be affected by manipulation of training settings (Dosher et al., 2020; Talluri et al., 2015). The differential employment of the location-specific and location-invariant representations would in turn determine the degree of location specificity or transferability of the learned stimulus features. That is, specificity or transferability of VPL is not an all-or-none phenomenon but rather constitutes two extremes of a continuum; it is a false dichotomy to attribute VPL to either low-level or high-level cortical processes.
In view of the above studies, VPL is most likely mediated by a flexible mechanism that can differentially engage low- and high-level processes, depending on many factors such as the difficulty or precision of the task (Talluri et al., 2015) and the interactions between different training conditions (Dosher et al., 2020). However, it is unknown whether such a flexible learning mechanism could also apply to training with an identical stimulus and task.
Although specificity or transferability of VPL revealed by psychophysical studies is commonly used to deduce the learning mechanisms under different training conditions, learning transfer could be induced by different mechanisms. If an irrelevant task is introduced at a location (the transfer location) different from the trained VPL task and trained visual-field location, originally location-specific VPL becomes transferable to the transfer location associated with the irrelevant task (Wang et al., 2012; Xiao et al., 2008). This line of research using the so-called double-training paradigm has equated perceptual learning with cognitive or conceptual learning that is location independent (Hu et al., 2021; Wang et al., 2016; J. Y. Zhang et al., 2010). However, the location transfer of VPL can also be mediated by refined readout from location-invariant sensory representations (Dosher et al., 2013, 2020; Talluri et al., 2015) or by refined location-invariant representations per se. In addition to leveraging the location-invariant processes, in the case of double training, one could also achieve learning transfer to the location of the irrelevant task through a mechanism similar to location-specific VPL: Because of a complex interaction of the stimuli and tasks with the transfer location during double training, the functional connectivities might be strengthened between cognitive areas and the visual cortical region representing the transfer location where the irrelevant task is exercised (Wang et al., 2012; Watanabe & Sasaki, 2015). In a similar vein, because of the potential interaction of the stimulus and task with the transfer location, a pretest by itself at an untrained location suffices to enable learning transfer to the untrained location (Harris & Sagi, 2015; T. Zhang et al., 2010). Because of the aforementioned complications introduced by additional stimuli or tasks at the transfer location, the location specificity and transferability shown in previous VPL studies using different experimental designs do not necessarily indicate different mechanisms supporting learning. It is also unclear whether VPL can be generalized to an untrained location that is truly untouched before testing for learning transfer. Resolving this issue is critical for dissociating the contribution of location-invariant and location-specific processes to VPL and also for a better understanding of the flexible mechanisms of VPL.
Statement of Relevance
Learning generalization and retention are two fundamental issues in the field of learning research. In conventional studies on visual perceptual learning, an improved perceptual ability with training at a single visual-field location is characterized by its specificity and long-term retention. The current study found that distributed training across multiple visual-field locations enhanced perceptual ability to the same degree as focused training at a single location, but these two forms of training differed in location generalization and long-term retention. These distinctions suggest that, even for the same training task, the mechanisms supporting the same perceptual benefit can be flexibly adjusted according to the settings used for training. Our findings are helpful for resolving the debate on learning transfer and specificity whereby the mechanisms of perceptual learning are inferred.
To address the above issue, we trained human observers to practice a conventional orientation-discrimination task on an identical stimulus presented at a single visual-field location (focused training) or multiple visual-field locations (distributed training). We simplified the double-training paradigm to avoid the complications introduced by irrelevant stimuli and tasks or by the pretest, ensuring that the untrained locations would be untouched before testing for learning transfer. This design allows for a better dissociation of the above-mentioned factors and mechanisms that could contribute to VPL and its location transfer.
The current study employed the typical VPL of orientation discrimination, which has been shown to be correlated with changes in location-specific and location-invariant cortical representations: Orientation-discrimination training sharpens the orientation-tuning functions of neurons not only in the primary visual cortex (V1), which is location sensitive (Schoups et al., 2001), but also in the higher-tier area V4 (Adab & Vogels, 2011; Raiguel et al., 2006; Yang & Maunsell, 2004) and the posterior inferior temporal cortex (Adab et al., 2014), which become increasingly location insensitive. Learning-induced specific changes in neuronal orientation tuning would predict a bias in orientation perception (the tilt aftereffect, or TAE) near the trained orientation (Teich & Qian, 2003). Such a bias has indeed been reported after orientation-discrimination training (Chen & Fang, 2011). To examine whether learning and its generalization involve modifications to sensory processing and whether location-specific and location-invariant sensory representations differentially contribute to learning under the two training conditions, we compared the training effects on orientation discriminability and TAE as well as their retention at the trained and untrained locations.
Our hypothesis, which is based on the aforementioned empirical and theoretical accounts of VPL, is that distributed training—compared with focused training—promotes location generalization of both improved orientation discriminability and induced TAE. Our results support this hypothesis and suggest differential employment of location-specific and location-invariant representations for the same discrimination learning.
Method
Subjects
Thirty-two (23 female, nine male; age: M = 22.2 years, SD = 0.3) and 20 (15 female, five male; age: M = 20.8 years, SD = 0.6) naive human subjects with normal or corrected-to-normal eyesight were recruited from the undergraduate and graduate student population to participate in Experiments 1 and 2, respectively. The sample sizes were determined according to previous VPL studies (Dosher et al., 2020; Kattner et al., 2017; Xiao et al., 2008) and the balanced distribution of subjects across the visual-field locations tested (for details, see the Procedure section). All subjects were inexperienced with perceptual training and unaware of the purpose of the study. Each subject was required to sign an informed consent form. The current study conformed to the Declaration of Helsinki and was approved by the Ethics Committee of Beijing Normal University.
Visual stimuli and behavioral tasks
Visual stimuli were generated with Psychtoolbox-3 (Brainard, 1997), based in MATLAB (The MathWorks, Natick, MA), on a 27-in. LED monitor (Acer XB271HU; Acer Corporation, San Jose, CA; 2,560 × 1,440 pixels, 0.233 mm/pixel, 120-Hz frame rate). The luminance of the monitor was linearized by an 8-bit look-up table. The viewing distance was 100 cm. A head-and-chin rest was used to stabilize the subject’s head. All experiments were conducted in a dimly lit room.
We used Gabor patches to examine the subjects’ thresholds for orientation discrimination and their biases in perceived vertical before and after orientation-discrimination training. The Gabor stimuli were stationary, phase-randomized, Gaussian-windowed, sinusoidal gratings with 47% Michelson contrast, 0.3° standard deviation, 2 cycle/degree spatial frequency, and 49.3 cd/m2 mean luminance that was equal to the screen background (Figs. 1a, 2a, 2d, 3a, and 3d). We presented the Gabor stimuli at visual-field locations along 3° retinal eccentricity when the subjects gazed at a fixation point in the screen center. In each block of trials, only one fixed stimulus location was tested.
Orientation-discrimination learning under single- and multiple-location training (Experiment 1). (a) Experimental design: Left group of panels illustrates a trial in the orientation-discrimination task at 36° reference orientation. Right panel shows the four possible stimulus locations (dotted circles with annotations). (b and e) Learning curves were averaged across the subjects in Group 1 (n = 16) and Group 2 (n = 16), respectively. Individual subjects’ midtest thresholds at the trained and untrained locations are shown in (c) and (f). Midtest and posttest thresholds are compared in (d) and (g). Solid symbols represent the eight subjects whose eye movements were monitored throughout all experimental sessions. Error bars represent standard errors of the mean. FP = fixation point.
Effects of focused training at a single location on discriminability and bias in orientation perception (Experiment 2). (a and d) Experimental design shows the two locations (dotted circles) used for measuring the orientation threshold (70° reference orientation) and orientation bias (90° reference orientation), respectively. Thresholds (b) and biases (e) at the trained (LB or LC; squares) and untrained (LC or LB; circles) locations were respectively averaged across the subjects (n = 10). Individual subjects’ thresholds (c) and biases (f) were compared between Posttest 1 and Posttest 2 (gray symbols) and between Posttest 1 and Posttest 3 (black symbols) at both the trained (two left panels) and untrained (two right panels) locations. Error bars represent standard errors of the mean. FP = fixation point.
Effects of distributed training at two locations (LB plus LD, or LC plus LA) on discriminability and bias in orientation perception (Experiment 2). (a and d) Experimental design shows the four possible locations (a) used for measuring the orientation thresholds (70° reference orientation) and the two possible locations (d) for measuring orientation biases (90° reference orientation). Thresholds (b) and biases (e) at the trained location of LB or LC (squares) and the untrained location of LC or LB (circles) were respectively averaged across the subjects (n = 10). Individual subjects’ thresholds (c) and biases (f) were compared between Posttest 1 and Posttest 2 (gray) and between Posttest 1 and Posttest 3 (black) at both the trained (two left panels) and untrained (two right panels) locations. Error bars represent standard errors of the mean. FP = fixation point.
In both Experiments 1 and 2, the threshold for orientation discrimination was measured with a two-interval forced-choice task. Although similar learning curves can be obtained using a one-interval forced-choice method (Dosher et al., 2013, 2020; Schoups et al., 2001), the two-interval forced-choice task minimizes the effect of subjective bias on estimation of the orientation threshold by providing an explicit reference within either of the two stimulus intervals. In each trial, the fixation point was first displayed for 500 ms. Then, two Gabor patches were sequentially displayed at the same location, each for 90 ms with an interstimulus interval of 600 ms. The orientation of one Gabor was fixed (referred to as the reference), and the other Gabor (referred to as the probe) was identical to the reference except for a small difference in orientation (Fig. 1a, left panel). The subjects were required to report whether the second Gabor was tilted clockwise or counterclockwise relative to the first one. Error feedback was provided only during the training stages (Tables 1 and 2). The orientation difference between the reference and probe was controlled by a classic three-down-one-up staircase procedure, which resulted in a 79.4% convergence rate. The step size of the staircase was 0.05 log units. Each staircase (a block) consisted of eight reversals (~40–50 trials). The geometric mean of the last four reversals was taken as the discrimination threshold.
Experiment 1 Procedure
Note: Groups 1 and 2 first practiced the orientation-discrimination task, respectively, under the single- and multiple-location conditions (the first training stage), and then the conditions were switched for further training (the second training stage).
Experiment 2 Procedure
Note: Groups 1 and 2 were trained respectively at one or two visual-field locations.
Experiment 2 also measured the perceived vertical with a one-interval forced-choice method of constant stimuli before and after the orientation-discrimination training. The measurement at each location consisted of four blocks of 50 trials. Within a block of trials, each of five Gabor orientations (90°, 90° ± 3°, and 90° ± 6°) was repeated in 10 trials in a random order. The subjects were required to report whether the Gabor was tilted clockwise or counterclockwise relative to the vertical (90°; Figs. 2d and 3d). No feedback was given on the observer’s responses. We calculated the percentage of responses in which the Gabor was reported as clockwise tilted relative to the vertical. Using the response ratios on the five stimulus orientations, we constructed the psychometric curve and fitted it with a logistic function. The perceived vertical was the orientation at which the observer’s response was 50%. We defined the bias magnitude as the deviation of the physical vertical from the perceived vertical (i.e., 90° minus the perceived vertical).
Procedure
Before data collection in Experiments 1 and 2, all subjects were given approximately 50 practice trials to familiarize them with the behavioral tasks.
In Experiment 1, we trained the subjects to discriminate the Gabor orientation at a reference orientation of 36°. We measured orientation-discrimination thresholds at four locations spaced 2.05° apart along 3° visual-field eccentricity (referred to as LA, LB, LC, and LD; Fig. 1a, right panel). We evenly and randomly divided the 32 subjects into two groups (Table 1). One group was trained at one location (focused training); the other group was trained at three locations (distributed training). Specifically, each of the 16 subjects in Group 1 was randomly assigned to one of the four locations evenly (i.e., LA, LB, LC, or LD) in the first five daily training sessions. After a pretest (Session 1) and the first training stage (Sessions 2–5), we examined the location specificity of orientation-discrimination learning by comparing the midtest (Session 6) threshold between the trained location (single-location condition; Table 1) and three untrained locations (multiple-location condition; Table 1). To further verify the location specificity, in the second training stage (Sessions 7–9), we had the subjects continue with the same orientation task at the three untrained locations (Table 1). The total number of blocks of training trials per day was the same as in the first training stage. We assessed whether the training could further improve orientation discriminability by comparing the midtest (Session 6) and posttest (Session 10) thresholds. The 16 subjects in Group 2 were trained using a similar procedure except that multiple-location training preceded single-location training (for details, see Table 1).
Experiment 2 examined both the orientation-discrimination threshold and orientation-bias magnitude. We used the same orientation-training task as in Experiment 1 but at a reference orientation of 70° (20° clockwise from the vertical). With this setting, one would expect a relatively strong clockwise shift of the subjective vertical after the training (Chen & Fang, 2011; Pinchuk-Yacobi et al., 2016; Teich & Qian, 2003). We evenly and randomly divided the 20 subjects into Groups 1 and 2. They performed the orientation-discrimination task at either a single location (Group 1) or multiple locations (Group 2) for 5 days. Group 1 was trained at LB (n = 5) or LC (n = 5; Table 2 and Fig. 2a; single-location condition); Group 2 practiced at two locations, either LB plus LD (n = 5) or LC plus LA (n = 5; Table 2 and Fig. 3a; multiple-location condition). For both groups, we compared the posttraining orientation thresholds between the trained location of LB or LC and the untrained location of LC or LB (for details, see Table 2). Before and after the orientation-discrimination training, we also measured the bias in perceived vertical at both LB and LC to compare the training-induced TAE. All subjects were called back 2 weeks later (Posttest 2) and approximately 11 weeks later (Posttest 3) for a retest. In all testing sessions, the subjective vertical was measured before the orientation-discrimination threshold.
Note that, in both Experiments 1 and 2, the trained and untrained locations were counterbalanced across subjects to avoid a systematic pretraining difference in threshold between the trained and untrained locations, allowing us to examine learning transfer simply by comparing the posttraining thresholds across locations.
Eye movement control
For eight of the 32 subjects in Experiment 1, we monitored their eye movements using an EyeLink 1000 infrared eye tracker (SR Research, Kanata, ON, Canada; 250-Hz sampling rate). In a trial, if the observers shifted their gaze outside the fixation window (1.5° in radius) during stimulus presentation, the trial was aborted immediately. The percentage of aborted trials accounted for 5.83% ± 1.18% of the total trials. In 93.76% ± 1.47% and 61.67% ± 5.08% of the valid trials, respectively, the eye positions were located within 1° and 0.5° (radius). These fixation-controlled subjects showed orientation thresholds and location transfer that were comparable with those of the remaining subjects (compare the solid with the hollow symbols in Figs. 1c, 1d, 1f, and 1g), excluding the possibility of contamination of the results by eye movements.
Data analysis
All statistical analyses (analyses of variance, or ANOVA, and Bayesian analyses) were conducted using JASP software (JASP Team, 2021). The Bayesian statistics (Bayes factor, or BF10) computed with the default prior options were used to evaluate the evidence in favor of the null hypothesis.
Results
Experiment 1: comparison of orientation-discrimination learning and its generalization between focused and distributed training
Thirty-two naive subjects practiced the orientation-discrimination task at a single location (Group 1; n = 16) or multiple locations (Group 2; n = 16) for 5 days (Table 1 and Fig. 1a). We compared training-induced improvement in discriminability and its transfer to untrained locations between the two forms of training (focused vs. distributed).
The two different training procedures led to the same amount of discrimination learning at the trained location (Figs. 1b and 1e), as indicated by a mixed-design ANOVA, with group (1, 2) as a between-subjects factor and session (pretest, midtest) as a within-subjects factor—significant main effect of session: F(1, 30) = 139.70, p < .001, η p 2 = .82; no significant main effect of group: F(1, 30) = 0.15, p = .704, η p 2 = .01, BF10 = 0.29; no interaction between group and session: F(1, 30) = 2.94, p = .097, η p 2 = .09, BF10 = 1.02. These results also imply that, for both groups of subjects, their mean discrimination thresholds start at similar pretest levels and decrease to similar midtest levels at the trained location.
We next compared location transfer of learning between the two groups of subjects by conducting a mixed ANOVA on the midtest thresholds, with location (trained, untrained) as a within-subjects factor and group (1, 2) as a between-subjects factor. We found a significant interaction between group and location, F(1, 30) = 6.77, p = .014, η p 2 = .18, indicating a significant difference in location transfer between the two groups. Specifically, for Group 1, the untrained location showed a significantly larger threshold than the trained location (repeated measures ANOVA), F(1, 15) = 23.55, p < .001, η p 2 = .61 (Fig. 1b; for data from individual subjects, see Fig. 1c). In contrast, for Group 2, the difference between the trained and untrained locations was not significant, F(1, 15) = 4.46, p = .052, η p 2 = .23, BF10 = 1.46 (Fig. 1e; for data from individual subjects, see Fig. 1f). The specificity (Group 1) and transfer (Group 2) were further supported by a direct comparison of midtest thresholds across groups in the same conditions. In the multiple-location condition (untrained in Group 1 but trained in Group 2), Group 2 showed a lower threshold than Group 1 (one-way ANOVA), F(1, 30) = 8.80, p = .006, η p 2 = .23 (compare Fig. 1e with Fig. 1b, green squares in Session 6). In contrast, in the single-location condition (trained in Group 1 but untrained in Group 2), there was no difference in mean threshold between the two groups of subjects (one-way ANOVA), F(1, 30) = 0.00, p = .980, η p 2 = .00, BF10 = 0.34 (compare Fig. 1e with Fig. 1b, orange circles in Session 6). These analyses suggest that orientation-discrimination learning induced by the multiple-location training is completely transferable to untrained locations, but the same improvement induced by the single-location training is not.
To further verify the difference in learning transfer between the two forms of training, we trained the two groups of subjects for 3 more days in the orientation-discrimination task under their respective untrained conditions (multiple- and single-location conditions for Groups 1 and 2, respectively). The rationale of the second training stage is that no further improvement with training would indicate complete location transfer; otherwise, location specificity is present. We conducted a mixed ANOVA with session (midtest, posttest) and location (trained location, untrained location in the second training stage) as within-subjects factors and group (1, 2) as a between-subjects factor. We found a significant interaction between group and session, F(1, 30) = 7.52, p = .010, η p 2 = 0.20, indicating a differential learning effect between the two groups during the second training stage. The training significantly decreased the threshold in Group 1 (one-way ANOVA), F(1, 15) = 17.96, p < .001, η p 2 = .55, but not in Group 2, F(1, 15) = 0.37, p = .552, η p 2 = .02, BF10 = 0.37, further supporting the specificity in Group 1 and transfer in Group 2 observed in the first training stage.
Experiment 1 revealed that orientation-discrimination learning was location specific when the training was focused at one location; however, when two additional locations were added, learning was generalized to a new, untouched location where no stimulus and task had been involved before testing for learning transfer.
Experiment 2: probing the differential mechanisms underlying focused and distributed training
In this experiment, we compared the effects of focused and distributed orientation-discrimination training on subjective bias in orientation perception (i.e., training-induced TAE). Twenty naive subjects practiced the orientation-discrimination task at a single location (Group 1; n = 10; Table 2 and Fig. 2a; LB or LC) or two locations (Group 2; n = 10; Table 2 and Fig. 3a; either LB plus LD or LC plus LA) for 5 days. Considering that the training-induced TAE shown in previous studies markedly diminishes in about 1 week (Chen & Fang, 2011) and that improved performance using some training procedures does not deteriorate until several months after the training (Hung & Carrasco, 2021; Qu et al., 2010; Yashar & Carrasco, 2016), we retested the subjects 2 weeks later (Posttest 2) and approximately 11 weeks later (Posttest 3) to assess long-term retention of both training effects.
We first compared the improvement in orientation discrimination between focused (Group 1) and distributed (Group 2) training. A mixed-design ANOVA, with group (1, 2) as a between-subjects factor and session (Pretest, Posttest 1) as a within-subjects factor, showed results (Figs. 2b and 3b) consistent with those from Experiment 1: a significant main effect of session, F(1, 18) = 59.83, p < .001, η p 2 = .77; no significant main effect of group, F(1, 18) = 0.10, p = .755, η p 2 = .01, BF10 = 0.38; and no significant interaction between group and session, F(1, 18) = 0.00, p = .993, η p 2 = .00, BF10 = 0.39. These results indicate once again that the two different training approaches result in the same amount of learning at the trained location.
Next, we conducted detailed analyses of the two training effects (improved discriminability and training-induced TAE) induced by focused and distributed training, respectively.
To evaluate location specificity and long-term retention of the discriminability learning for Group 1 (focused training), we conducted a repeated measures ANOVA, with location (trained, untrained) and session (Posttest 1, Posttest 2, Posttest 3) as two influencing factors. We found a main effect of location, F(1, 9) = 23.51, p < .001, η p 2 = .72, but not session, F(2, 18) = 0.51, p = .610, η p 2 = .05, BF10 = 0.17, and no interaction between them, F(2, 18) = 0.41, p = .669, η p 2 = .04, BF10 = 0.27. Specifically, in Posttest 1, the trained location showed a significantly smaller threshold than the untrained location, F(1, 9) = 41.15, p < .001, η p 2 = .82 (Fig. 2b). These results suggest that single-location training leads to location-specific and fully retained orientation-discrimination learning.
To assess the training-induced TAE and its location specificity, we conducted a repeated measures ANOVA with two influencing factors of location (trained, untrained) and session (Pretest, Posttest 1). There was a significant interaction between these two factors, F(1, 9) = 17.20, p = .002, η p 2 = .66. Further analysis showed a significant bias increment at the trained location (one-way ANOVA), F(1, 9) = 37.11, p < .001, η p 2 = .81, but not the untrained location, F(1, 9) = 1.19, p = .303, η p 2 = .12, BF10 = 0.60 (Fig. 2e), indicating an entirely location-specific TAE after the single-location training.
To test whether the training-induced TAE would be a lasting effect, we conducted a repeated measures ANOVA with location (trained, untrained) and session (Posttest 1, Posttest 2, Posttest 3) as two factors. An interaction was observed between them, F(2, 18) = 6.17, p = .009, η p 2 = .41. Further analysis showed that the bias magnitude at the trained location was significantly reduced 2 weeks later (one-way ANOVA), F(1, 9) = 10.52, p = .010, η p 2 = .54 (Fig. 2e, Posttest 1 vs. Posttest 2; for data from individual subjects, see Fig. 2f), but the bias was still larger than that at the untrained location, F(1, 9) = 6.94, p = .027, η p 2 = .44 (Fig. 2e, Posttest 2).
The results shown in Figure 2 indicate that, following the single-location training, the orientation-discrimination learning is location specific but long-lasting and that the training-induced TAE is entirely location specific and partially persistent.
The data from Group 2 (distributed training; Fig. 3) were analyzed in a way similar to Group 1.
To examine location specificity and long-term retention of the improved discriminability after distributed training, we conducted a repeated measures ANOVA with two influencing factors of location (trained, untrained) and session (Posttest 1, Posttest 2, Posttest 3). There was a main effect of session, F(2, 18) = 4.02, p = .036, η p 2 = .31, but not location, F(1, 9) = 0.81, p = .390, η p 2 = .08, BF10 = 0.30, and there was no significant interaction between the two factors, F(2, 18) = 0.95, p = .404, η p 2 = .10, BF10 = 0.25 (Fig. 3b; for data from individual subjects, see Fig. 3c). These results indicate that, after distributed training, the enhanced discriminability is fully transferred to the untrained location and that the learning effects at both the trained and untrained locations diminish slightly.
To estimate the training-induced TAE and its location transfer, we conducted a repeated measures ANOVA with location (trained, untrained) and session (Pretest, Posttest 1) as two factors. There was a significant interaction between them, F(1, 9) = 6.31, p = .033, η p 2 = .41 (Fig. 3e). One-way ANOVA showed that multiple-location training increased the orientation bias at both the trained, F(1, 9) = 52.76, p < .001, η p 2 = .85, and untrained, F(1, 9) = 67.79, p < .001, η p 2 = .88, locations, with a smaller increment at the untrained location, F(1, 9) = 5.28, p = .047, η p 2 = .37, indicating partial transfer of the TAE induced by multiple-location training.
To evaluate the retainability of the training-induced TAE, we conducted a repeated measures ANOVA with location (trained, untrained) and session (Posttest 1, Posttest 2, Posttest 3) as two influencing factors. A significant interaction between them was observed, F(2, 18) = 4.06, p = .035, η p 2 = .31 (Fig. 3e). The bias at the untrained location remained unchanged 2 weeks later (Posttest 1 vs. Posttest 2, one-way ANOVA), F(1, 9) = 0.74, p = .413, η p 2 = .08, BF10 = 0.51 (Fig. 3e), whereas the bias at the trained location significantly decreased (Posttest 1 vs. Posttest 2), F(1, 9) = 6.31, p = .033, η p 2 = .41, and reached a level comparable with the untrained location (Posttest 2, trained vs. untrained), F(1, 9) = 2.40, p = .156, η p 2 = .21, BF10 = 0.85 (for data from individual subjects, see Fig. 3f). These results suggest that distributed training leads to two dissociable bias components: a location-specific, short-lived component and a location-independent, long-lasting component.
Experiment 2 (Figs. 2 and 3) revealed two major differences between the two forms of training despite a similar improvement in discrimination ability. First, in comparison with focused training, distributed training promotes location generalization of both discrimination learning and training-induced TAE. Second, for focused training, the specificities of both discrimination learning and training-induced TAE are strong and long-lasting, but for distributed training, the specificity of discrimination learning is negligible and the specificity of training-induced TAE is short-lived.
Discussion
The current study showed that, in comparison with focused training at a single location, distributing the identical training task across multiple locations led to the same improvement in discriminability but more generalization of this learning effect and the training-induced TAE. The two forms of training also differed in long-term retention of the two training effects. These results suggest that location-specific and location-invariant representations could be differentially employed for the same learning.
Accumulating evidence suggests that VPL involves a broad network of brain regions (for reviews, see Dosher & Lu, 2017; Li, 2016; Maniglia & Seitz, 2018; Watanabe & Sasaki, 2015). Three distinct mechanisms can contribute to learning-induced improvement in orientation discriminability.
First, learning-induced refinement in orientation representation in multiple visual areas, as seen in V1 (Jehee et al., 2012; Schoups et al., 2001), V4 (Jehee et al., 2012; Yang & Maunsell, 2004), and the posterior inferior temporal cortex (Adab et al., 2014), can well account for our observations: The location specificity of both training effects (the induced TAE and improved discriminability) following the focused training suggests refined orientation representations in location-specific areas such as V1 and V2; in contrast, distributed training may engage more orientation-sensitive but location-insensitive areas such as V4 and the posterior inferior temporal cortex, resulting in more location transfer of both training effects.
Second, optimization of readout, or optimization of the connectivities between visual representations and cognitive processes, has been considered to be a possible mechanism of VPL (Dosher & Lu, 2017; Li, 2016; Maniglia & Seitz, 2018; Watanabe & Sasaki, 2015). This mechanism can partly explain our observations: Optimized connectivities between the cognitive areas and the location-specific or location-invariant areas can respectively explain the specificity and generalization of the improved discriminability; however, it remains unclear whether the readout optimization by itself would predict the TAE that we observed.
Third, VPL and its transfer have been ascribed to changes in cognitive rather than sensory processes (Hu et al., 2021; Xiao et al., 2008), but cognitive learning alone cannot explain our observations of learning-induced TAE.
Based on the above discussion, our findings are in favor of the first mechanistic account.
Group 1 subjects in Experiment 1 showed plateaued discriminability after training at a single location, but the mean threshold at this location was further reduced by subsequent distributed training at the other three locations (Fig. 1b). The additional and transferable improvement further supports that distributed training may engage a mechanism involving location-invariant representations, different from the location-specific representations employed during the first focused-training stage. For Group 2 subjects, after the first distributed-training stage, location-invariant representations might have been optimized to a plateau and become more orientation informative than location-specific representations; therefore, no further behavioral improvement was observed in the second focused-training stage.
Focused and distributed training also differed in learning retainability. The enhanced discriminability remained unchanged approximately 11 weeks after focused training (Fig. 2b), consistent with previous studies (Vogels & Orban, 1985; E. Zhang & Li, 2020). In contrast, the distributed training effects slightly decreased 2 weeks later (Fig. 3b). In fact, similar deteriorations of VPL have also been reported when the training involves multiple orientations (Hung & Carrasco, 2021), multiple locations (Qu et al., 2010), and complex tasks (Yashar & Carrasco, 2016). Our present finding, together with the previous observations on VPL deterioration, implies that when multiple or complex tasks and stimuli are used for training, learning generalization may take place at the expense of long-term retention.
According to previous studies (Chen & Fang, 2011; Pinchuk-Yacobi et al., 2016), training-induced TAE is not simply a form of passive sensory adaptation due to repeated stimulus exposure; instead, it could be an indication of learning-induced changes in orientation-sensitive visual areas. The current study showed a dissociable TAE component that was location specific and short-lived (Figs. 2e and 3e). This volatile component may reflect dynamic changes in location-specific representations because it exhibits temporal dynamics and location specificity that are very similar to the activation patterns seen in V1 over the course of orientation-discrimination learning: A location-specific enhancement in V1 activity is present during early training, but the elevated activities drop back to pretraining levels after 2 weeks of training despite maintained behavioral performance (Yotsumoto et al., 2008).
Studies on the interference between different learning tasks may also provide an insight into the flexible mechanisms for learning generalization. It has been shown that training with intermixed tasks disrupts learning (Dosher et al., 2020; Seitz et al., 2005) and that an intertask interval (~1 hr) helps alleviate the interference (Seitz et al., 2005). Such a time window for learning consolidation could also be helpful for integrating information about different training conditions interleaved in different blocks of trials, which might in turn affect learning mechanisms and thus learning generalization. Moreover, the interference between different tasks or conditions depends on their similarity and shared representations; location-invariant representations involve overlapping populations of neurons to encode similar orientations at different visual-field locations, incurring interference of orientation-discrimination learning at different locations (Dosher et al., 2020). The similarity or proximity between different training conditions (and between their neural representations) may also affect the learning mechanisms and learning generalization. Sequential practice on different perceptual tasks sharing a common high-level task structure produces fast learning of new tasks having the same structure (Kattner et al., 2017). In particular, it has been shown that, when visual and tactile orientation-discrimination training share the same stimulus orientation, orientation representation at a conceptual level could be engaged, enabling learning transfer across sensory modalities (Hu et al., 2021). For distributed training in the present study, location-insensitive cortical areas representing the identical stimuli displayed at different locations are likely to subserve learning transfer across locations.
In conclusion, depending on the exact task and stimulus settings used for training, the brain is able to flexibly use differential mechanisms to support learning, leading to differential behavioral characteristics. In particular, as demonstrated in the current study, such a flexible mechanism even applies to learning with the identical stimulus and task, but the present behavioral results alone cannot provide a definite answer to the intricate cortical mechanisms. Understanding the general rules that the brain uses for selection of neural processes will provide insights into the nature of learning and also have implications for clinical rehabilitation and skill training.
Footnotes
Acknowledgements
We thank Xibin Xu, Zhou Zhao, and Qing Zhu for technical assistance.
Transparency
Action Editor: Alice Cronin-Golomb
Editor: Patricia J. Bauer
Author Contribution(s)
Yangyang Du and Gongliang Zhang contributed equally to this work
