Reactivation-induced memory integration prevents proactive interference in perceptual learning

Abstract

We acquire perceptual skills through experience to adapt ourselves to the changing environment. Accomplishing an effective skill acquisition is a main purpose of perceptual learning research. Given the often observed learning effect specificity, multiple perceptual learnings with shared parameters could serve to improve the generalization of the learning effect. However, the interference between the overlapping memory traces of different learnings may impede this effort. Here, we trained human participants on an orientation discrimination task. We observed a proactive interference effect that the first training blocked the second training at its untrained location. This was a more pronounced effect than the well-known location specificity in perceptual learning. We introduced a short reactivation of the first training before the second training and successfully eliminated the proactive interference when the second training was inside the reconsolidation time window of the reactivated first training. Interestingly, we found that practicing an irrelevant task at the location of the second training immediately after the reactivation of the first training could also restore the effect of the second training but in a smaller magnitude, even if the second training was conducted outside of the reconsolidation window. We proposed a two-level mechanism of reactivation-induced memory integration to account for these results. The reactivation-based procedure could integrate either the previously trained and untrained locations or the two trainings at these locations, depending on the activated representations during the reconsolidation process. The findings provide us with new insight into the roles of long-term memory mechanisms in perceptual learning.

Introduction

Early theories of visual perceptual learning considered the improvement of visual functions through experience as a result of perceptual plasticity (Hochstein & Ahissar, 2002; Sagi & Tanne, 1994). A hallmark of perceptual learning that supported this proposal was the initially observed location, feature, or eye specificity of the learning effects (Ahissar & Hochstein, 1997; Ball & Sekuler, 1987; Crist et al., 1997; Fahle, 1997; Karni & Sagi, 1991; Shiu & Pashler, 1992). However, the accumulating evidence revealed by succeeding investigations pointed out that the underlying mechanism of perceptual learning is not purely perceptual. These studies have found that the learning specificity could also be explained by higher level factors such as restricted attentional processing (Donovan & Carrasco, 2018; Donovan, Szpiro, & Carrasco, 2015; Xiao et al., 2008; Xiong, Zhang, & Yu, 2016) and modified reading out of sensory information from decision areas (Dosher & Lu, 1998; Jia et al., 2018; Jia et al., 2020; Law & Gold, 2008). Not beyond our expectation, the roles of mnemonic processing in perceptual learning are drawing increasing attention as well. Recent investigations have suggested that perceptual learning effect is associated with the changed representation of visual stimuli in short-term memory (Jia et al., 2021) and refined consolidation process of long-term memory (Bang et al., 2018; Tamaki et al., 2020; Yang, He, & Fang, 2022). 

One of the ultimate goals of perceptual learning research is to promote its application in the real world (Deveau, Ozer, & Seitz, 2014; Johnston et al., 2020; Sha, Toh, Remington, & Jiang, 2020). That would require the evidence of performance improvement in the contexts that are not identical to that of the training. Intuitively, this could be achieved by facilitating the generalization of the learning effect (Harris, Gliksberg, & Sagi, 2012; Kattner, Cochrane, Cox, Gorman, & Green, 2017). However, the widely observed learning specificity in perceptual learning was an obvious barrier against this idea. To address this issue, several procedures based on double training (Wang, Zhang, Klein, Levi, & Yu,2014; Xiao et al., 2008), attentional cueing (Donovan & Carrasco, 2018; Donovan et al., 2015), or concept-related categories (Tan, Wang, Sasaki, & Watanabe, 2019; Wang et al., 2016) have been proposed to improve the transfer of learning effect to untrained stimuli or retinal locations. These approaches, although successful in demonstrating the generalized learning effect, required other manipulations in addition to the trained task. 

To extend the acquired learning effect to the untrained conditions, we could also conduct the training with modified parameters. The hope is that by sharing a subset of the parameters with the original training, the new training may require reduced effort to accomplish. However, interference is a ubiquitous phenomenon when multiple learnings with shared parameters take place (Anderson, 2003; Herszage & Censor, 2018; Postman & Underwood, 1973) and perceptual learning was not an exception (Bang et al., 2018; Been, Jans, & De Weerd, 2011; Huang, Zhang, & Li, 2022; Seitzt et al., 2005; Shibata et al., 2017; Yotsumoto, Chang, Watanabe, & Sasaki, 2009). Particularly, several studies have shown that the interference can be observed between two trainings that were separated by at least one day, indicating that long-term memory mechanism was involved in and may provide the solution for the interference between multiple perceptual trainings (Bang et al., 2018; Huang et al., 2022). 

It has long been suggested that the consolidation of learning-related memory traces is not a static process. Reactivating a consolidated memory can transfer it to an unstable state in which the memory can be modified and reconsolidated (Dudai, 2006; Elsey, Van Ast, & Kindt, 2018; Lee, Nader, & Schiller, 2017; Nader & Hardt, 2009). The cycle of reactivation and reconsolidation could occur more than once and constitutes a crucial part of the learning flexibility in daily experience. Interestingly, recent studies have shown that reactivation could facilitate the perceptual learning of texture discrimination task (Amar-Halpert, Laor-Maayany, Nemni, Rosenblatt, & Censor, 2017; Chen & de Beeck, 2021; Klorfeld-Auslender et al., 2022). The strengthened memory through the reactivation and reconsolidation cycles was proposed to account for the facilitation effect (Amar-Halpert et al., 2017). Furthermore, the procedures based on reactivation were shown to resolve the interference of multiple learnings in motor learning (Herszage & Censor, 2017) and reward learning (Huang & Li, 2022). These results were accompanied by the suggestion that memory integration triggered by the reactivated memory traces of the original learning and the novel memory of the new learning during the reconsolidation contributed to the prevention of the interference. This notion was consistent with the proposal that reactivation-induced integration of overlapping memory traces could facilitate the generalization of learning effect (Morton, Sherrill, & Preston, 2017; Ritvo, Turk-Browne, & Norman, 2019; Schlichting & Frankland, 2017). However, given the unique phenomenon of the location specificity in perceptual learning, it was unclear whether and how the reactivation-based procedure could facilitate the efficiency of multiple perceptual trainings by preventing interference between them. More specifically, how the location representation is involved in the integration process remained a critical issue to address. 

The present study consisted of five psychophysical experiments to examine the role of memory reactivation and its induced memory integration in solving the potential interference between two perceptual learnings (Figure 1A). Experiments 1 established a baseline learning effect for the visual orientation discrimination task after five days of training (i.e., the first training). Experiment 2 demonstrated a proactive effect when a second training session at the untrained location was added 2 days after the first training. Experiment 3 showed that the interference could be eliminated by introducing a brief reactivation of the first training immediately before the second training. Experiment 4, in which a 6-hour interval was introduced between the reactivation and the second training, showed that conducting the second training inside the reconsolidation window of the first training was critical for preventing the interference. Finally, in experiment 5, we introduced a tilt discrimination task after the reactivation but before the 6-hour interval. This manipulation could also restore the effect of the second training, but in a smaller magnitude, leading us to propose a two-level mechanism for the reactivation-induced memory integration in perceptual learning. We totally recruited 80 participants with 16 participants for each experiment. This sample size is larger than the typical multiday perceptual learning studies to ensure sufficient statistical power when conducting between-subjects comparison. In a recent study (Hung & Carrasco, 2021) that examined the location specificity and transfer of the learning effect of the orientation discrimination task, nine participants were recruited for each training group and robust results were observed. Therefore, we believe that the sample size of the present study was appropriate to detect the potential training effects. 

Figure 1.

 

Experimental design. (A) Procedures of the experiments. (B) Orientation discrimination task. (C) Tilt discrimination task.

Methods

Participants

We totally recruited 80 participants with 16 participants from Peking University for each experiment (experiment 1: 12 females and 4 males, mean age = 19.50 years; experiment 2: 11 females and 5 males, mean age = 19.50 years; experiment 3: 7 females and 9 males, mean age = 19.69 years; experiment 4: 11 females and 5 males, mean age = 19.13 years; experiment 5: 12 females and 4 males, mean age = 19.75 years). Participants had normal or corrected-to-normal vision and no known neurological or visual disorders. They provided written informed consent before the experiment. The local ethics committee approved the study. 

Stimuli

Stimuli were generated in MATLAB (MathWorks, Natick, MA, USA) using Psychtoolbox 3 package (Brainard, 1997; Pelli, 1997) and presented on a dark-gray background (∼36 cd/m²) on a 20-inch CRT monitor (resolution of 1024 × 768 pixels and a 60-Hz refresh rate). Gamma correction was applied to the monitor. Gabor stimuli (spatial frequency = 1.5 cycle/degree, contrast = 0.8, Gaussian filter sigma = 0.6, diameter ≈ 3.5°) were positioned in either upper-left or lower-right visual field with an eccentricity of 6.5°. The experiment was conducted in a dimly lit room. A chinrest was used to stabilize the head and maintain an 85-cm viewing distance. 

Procedure

The present study consisted of five psychophysical experiments (Figure 1A). All experiments were completed within 9 days that consisted of a pretest session (day 1), first training sessions (days 2–6), a midtest session (day 7), and a post-test session (day 9). The experiments differed from each other in day 8. 

Orientation discrimination task

As shown in Figure 1B, each trial of the orientation discrimination task began with a 1,000-ms fixation interval, followed by the presentation of the reference Gabor stimulus for 200 ms at periphery (upper left or lower right). After another fixation interval of 600 ms, the test Gabor stimulus was shown for 200 ms at the same location as the reference. Participants were instructed to judge whether the test stimulus was tilted clockwise or counterclockwise relative to the reference stimulus with a keyboard within 1,500 ms from the offset of the test stimulus. The base orientation of the reference-test pair was 55° or 145°, counterbalanced across participants. One stimulus in the pair was randomly assigned to the base orientation and the other stimulus’ orientation was the base orientation plus a deviation. The orientation difference (i.e., the deviation) between reference and test stimuli was controlled by a three-down one-up staircase method (step size = 0.5°). This method converged to 79.4% correct responses. For the first block of each condition, the initial orientation difference was 5°. The stopping criterion of a block was 15 reversals or 100 trials, whichever was met first. The threshold in each block was determined by the mean orientation difference of the last eight reversals. The initial orientation difference of the subsequent blocks was set to be the threshold of the last block of the same condition. 

Test sessions

In each test session, participants completed three runs of the orientation discrimination task. Each run consisted four stimulus conditions with one block for each condition. The four conditions were Loc1–Ori1, Loc1–Ori2, Loc2–Ori1, and Loc2–Ori2 (Loc1 was the location for the first training, Loc2 was the location for the second training, Ori1 was the stimulus orientation for both trainings). The order of the four conditions in each run was randomized. There was no feedback during the test sessions. There was a practice of 120 trials before the pretest session, with 30 trials for each condition. Auditory feedback was provided after incorrect choices in the practice. 

Training sessions

Participants were trained on the orientation discrimination task with Gabors presented at the same orientation and location throughout the training. The stimuli conditions for the first and second trainings were Loc1–Ori1 and Loc2–Ori1, respectively. The orientation (55° or 145°) and location (upper left or lower right) of the trained Gabor stimulus were counterbalanced across participants. For the first training at Loc1, each participant was trained for 5 days with one session in each day. Each session consisted of 16 blocks. For the second training of day 8 at Loc2, each participant was trained for 8 blocks. There was auditory feedback after incorrect choices. The threshold of each training session was determined by the mean of the thresholds from the session. 

Here, the ratio of (second learning trials)/(first training trials) was 1/10, slightly smaller than our previous study (approximately 1/6) (Huang & Li, 2022), which had demonstrated that the procedure of reactivation plus new learning was effective in that the second learning requires much less training trials to reach a comparable performance level as the first learning. The main consideration for decreasing this ratio was that the first training of the present study was much longer as compared with the previous study and the orientation discrimination training was known to partially transfer to untrained location (i.e., Loc2 in the second training). We, therefore, expected to observe the restored effect at Loc2 with comparable but smaller ratio of trials for the second training session. 

Reactivation session

The reactivation session was identical to the first training session except that the participants were only trained for two blocks. We noted that previous investigations that are related to the present study used only a few trials for reactivation (Amar-Halpert et al., 2017; Herszage, Sharon, & Censor, 2021). The present study aimed to decrease the interference by integrating the memory traces of two trainings. It has been suggested that memory integration requires strong coactivation of the two memories (Ritvo et al., 2019). A straight way to induce strong activation of the first training is to provide more trials during the reactivation. Here, we chose a comparable number of trials as in our previous study (Huang & Li, 2022), as well as in a recent perceptual learning study (Bang et al., 2018). However, we do not exclude the possibility that reactivating the first training with less trials could also lead to similar effect. Future investigations are required to address this issue. 

Tilt discrimination task

As shown in Figure 1C, each trial began with a 1,000-ms fixation interval, followed by the presentation of a Gabor stimulus at Loc2 for 200 ms. The participants were required to judge whether the test stimulus was tilted clockwise or counter-clockwise relative to the vertical with a keyboard within 1500 ms from the offset of the test stimulus. There were five deviations from the vertical: −2.25°, −1.50°, −0.75°, 0°, 0.75°, 1.50°, and 2.25°. There was a practice of 28 trials at the beginning of the tilt discrimination task session. The formal experiment of the session consisted of 4 blocks with 140 trials in each block. 

Eye tracking

Participants were instructed to look at the fixation dot and their eye movements were monitored using Eyelink 1000 Plus (1,000 Hz, monocular). We ensured the maintenance of the fixation in the following four aspects. 1) Before the experiment, we emphasized the importance of fixating at the fixation point to the participants. 2) We monitored participants’ eye movement throughout the experiment. 3) Participants who could not fixate in the practice session were excluded from participating the experiment. 4) During the experiment, the experimenter monitored the fixation position of the participants. If there was a large deviation from the central fixation point, the participant would be reminded in the between-blocks rest. 

Data analyses

The present study focuses on the two trained stimulus conditions (Loc1–Ori1 and Loc2–Ori1) and the influence of reactivation manipulations on the training effects of the two conditions. Therefore, the main results are shown only for these two conditions of interest (referred as Loc1 and Loc2 because they shared the same orientation) to avoid the complexity of conducting too many comparisons in statistics. 

Repeated measures two-way analysis of variance on threshold, with session (pretest and midtest) and location (Loc1 and Loc2) as factors, was performed to examine the effect of the first training. Repeated measures two-way analysis of variance on threshold, with session (midtest and post-test) and location (Loc1 and Loc2) as factors, was performed to examine the effect of the second training. 

To decrease the impact of initial threshold on the training effect in between-experiment comparison, we calculated the mean percent improvement (MPI) to evaluate the performance improvements between test sessions. MPI_pre-mid was defined as ([pretest threshold – midtest threshold]/pretest threshold) × 100 to measure the improvement from pretest to midtest session. MPI_mid-post was defined in a similar fashion for the improvement from midtest to post-test session. Mixed analyses of variance with experiment as a between-subjects factor and location (Loc1 and Loc2) as a within-subjects factor were performed to compare the training effect across experiments. 

All data have been made publicly available at the Open Science Framework and can be accessed at https://osf.io/n6vuz/?view_only=39756bfce90249fa8e933a88776e0e94. 

Results

Experiment 1

Experiment 1 served as a baseline for the training effect of the first training at Loc1 (Figure 2). We were interested in comparing the training effects at Loc2 between different combinations of experiments, given our experimental manipulations. Therefore, we did not particularly assign a baseline training effect for Loc2. Training improved participants’ discrimination performance, as revealed by the decreased threshold from the first session, mean 2.36° ± 0.629°, to the fifth session, mean, 1.96° ± 0.477°, of the training phase, paired t test, t(15) = 4.36, p = 0.001, Cohen's d = 1.09, BF₁₀ = 63.29. The training effect was further evidenced by the significant main effect of session, F(1,15) = 53.14, p < 0.001, \({\rm{\eta }}_p^2\) = 0.78, BF₁₀ > 100, when we compared the estimated discrimination thresholds at the trained location (Loc1) and an untrained location reserved for the second training (Loc2) before (pretest) and after (midtest) the training sessions. The training effect was larger at Loc1 as revealed by the significant interaction between session and location, F(1,15) = 6.17, p = 0.025, \({\rm{\eta }}_p^2\) = 0.29, BF₁₀ = 3.68. The thresholds in the two locations were statistically indistinguishable in the pretest session, t(15) = 0.48, p = 0.64, Cohen's d = 0.12, BF₁₀ = 0.28. In midtest session, the threshold at Loc1 was significantly lower than that of Loc2, t(15) = 2.97, p = 0.01, Cohen's d = 0.74, BF₁₀ = 5.65. 

Figure 2.

 

Results of experiment 1. (A) Thresholds in the five sessions of the first training. (B) Thresholds in the test sessions for the locations of the first training (Loc1) and second training (Loc2). Error bars represent standard errors.

Experiment 2

Experiment 2 introduced a one-half session training of the same orientation at Loc2 and the results are shown in Figure 3. The second training was done two days after the first training to ensure that the first training was well-consolidated. 

Figure 3.

 

Results of experiment 2. (A) Thresholds in the five sessions of the first training. (B) Thresholds in the test sessions for the locations of the first training (Loc1) and second training (Loc2). Error bars represent standard errors.

The first training improved participants’ discrimination performance, as revealed by the decreased threshold from the first session, mean, 2.50° ± 0.640°. to the fifth session, mean, 1.78° ± 0.352°, of the training phase, paired t test: t(15) = 5.55, p = 0.001, Cohen's d = 1.39, BF₁₀ > 100. The training effect was further evidenced by the significant main effect of session, F(1,15) = 58.06, p < 0.001, \({\rm{\eta }}_p^2\) = 0.80, BF₁₀ > 100, when we compared the estimated discrimination thresholds at Loc1 and Loc2 before and after the training sessions. The training effect was larger at Loc1 as revealed by the marginally significant interaction between session and location, F(1,15) = 3.40, p = 0.085, \({\rm{\eta }}_p^2\) = 0.19, BF₁₀ = 1.61. The thresholds in the two locations were statistically indistinguishable in the pretest session, t(15) = 0.05, p = 0.964, Cohen's d = 0.01, BF₁₀ = 0.26. In the midtest session, the threshold at Loc1 was significantly lower than that of Loc2, t(15) = 2.56, p = 0.022, Cohen's d = 0.64, BF₁₀ = 2.90. 

The results showed no significant difference from experiment 1 in the effects of the first training, MPI_pre-mid, F(1,30) = 0.44, p = 0.514, \({\rm{\eta }}_p^2\;\)= 0.01, BF₁₀ = 0.36. However, the second training at Loc2 was not effective because neither the threshold difference between the midtest and post-test sessions, F(1,15) = 2.41, p = 0.141, \({\rm{\eta }}_p^2\;\)= 0.14, BF₁₀ = 0.43, nor the interaction between session and location, F(1,15) = 1.68, p = 0.215 \({\rm{\eta }}_p^2\;\)= 0.10, BF₁₀ = 0.93, was significant. Also, the threshold changes from the midtest to post-test sessions were not significantly different between the two experiments, F(1,30) = 0.01, p = 0.928, \({\rm{\eta }}_p^2\;\)< 0.01, BF₁₀ = 0.28. The main effect of location was marginally significant, F(1,30) = 3.42, p = 0.074, \({\rm{\eta }}_p^2\;\)= 0.1, BF₁₀ = 2.33. The interaction between location and experiment was not significant, F(1,30) = 0.11, p = 0.739, \({\rm{\eta }}_p^2\;\)< 0.01, BF₁₀ = 0.43. These results revealed a proactive interference effect from the first training to the second training, possibly owing to the 10-fold difference in the amount of the training between them. 

Experiment 3

Huang and Li (2023) showed that conducting a new color-reward learning during the reconsolidation window of the reactivated old color-reward learning facilitated their integration and prevented them from interfering each other. This result agreed with the proposal that high degree of coactivation of two memory traces could lead to their integration and prevent the interference between them (Wammes, Norman, & Turk-Browne, 2022). Therefore, we conducted experiment 3 in which a brief reactivation of the first training was added before the second training. This manipulation would potentially induce the coactivation of the two trainings because the second training would be temporally overlapped with the reconsolidation window of the first training. 

The first training improved participants’ discrimination performance, as revealed by the decreased threshold from the first session, mean, 2.59° ± 0.744°, to the fifth session, mean, 1.95° ± 0.47°, of the training phase, paired t test: t(15) = 4.26, p = 0.001, Cohen's d = 1.07, BF₁₀ = 53.35. The training effect was further evidenced by the significant main effect of session, F(1,15) = 47.93, p < 0.001, \({\rm{\eta }}_p^2\) = 0.76, BF₁₀ > 100, when we compared the estimated discrimination thresholds at Loc1 and Loc2 before and after the training sessions. The training effect was larger at Loc1 as revealed by the significant interaction between session and location, F(1,15) = 5.90, p = 0.028, \({\rm{\eta }}_p^2\) = 0.28, BF₁₀ = 3.10. The thresholds in the two locations were statistically indistinguishable in the pretest session, t(15) = 0.03, p = 0.977, Cohen's d = 0.01, BF₁₀ = 0.26. In the midtest session, the threshold at Loc1 was significantly lower than that of Loc2, t(15) = 3.75, p = 0.002, Cohen's d = 0.94, BF₁₀ = 21.79. 

As shown in Figure 4, adding a brief reactivation of the first training before the second training resulted in a significant effect for the second training, F(1,15) = 25.26, p < 0.001, \({\rm{\eta }}_p^2\) = 0.63, BF₁₀ > 100, and this effect was larger at Loc2, significant interaction between session and location: F(1,15) = 14.17, p = 0.002, \({\rm{\eta }}_p^2\) = 0.49, BF₁₀ = 15.84. The threshold at Loc1 was statistically lower than that of Loc2 in the midtest session, t(15) = 3.75, p = 0.002, Cohen's d = 0.94, BF₁₀ = 21.79. However, in the post-test session, the thresholds at the two locations were statistically indistinguishable, t(15) = 0.82, p = 0.426, Cohen's d = 0.21, BF₁₀ = 0.34. Further between-experiments comparisons in MPI_mid-post revealed that the effects of the second training were significantly larger in experiment 3 as compared with experiment 1, F(1,30) = 14.31, p = 0.001, \({\rm{\eta }}_p^2\) = 0.32, BF₁₀ = 10.39; interaction: F(1,30) = 0.01, p = 0.944, \({\rm{\eta }}_p^2\) < 0.01, BF₁₀ = 0.33, and experiment 2, F(1,30) = 18.84, p < 0.001, \({\rm{\eta }}_p^2\) = 0.39, BF₁₀ = 6.35; interaction: F(1,30) = 0.1, p = 0.75, \({\rm{\eta }}_p^2\) < 0.01, BF₁₀ = 0.39, suggesting that combining the reactivation with the second training successfully release the inhibition of learning at the previously untrained location (i.e., Loc2). Additionally, the performance at Loc1 was further improved after the second training, one-sample t-test against 0 for the MPI_mid-post: t(15) = 2.76, p = 0.015, Cohen's d = 0.69, BF₁₀ = 4.01. 

Figure 4.

 

Results of experiment 3. (A) Thresholds in the five sessions of the first training. (B) Thresholds in the test sessions for the locations of the first training (Loc1) and second training (Loc2). Error bars represent standard errors.

Experiment 4

To validate that the coactivation of the two trainings was the key factor in restoring the effect of the second training, we conducted experiment 4 in which a 6-hour interval was placed between the reactivation and second training sessions (Figure 5). We chose 6 hours as the time window for the process of reconsolidation based on previous literatures (Bang et al., 2018; Huang et al., 2022; Huang & Li, 2022; Nader, Schafe, & Le Doux, 2000). Because the second training was conducted after the first training's reconsolidation, no coactivation of the two trainings would occur and their integration was not possible. 

Figure 5.

 

Results of experiment 4. (A) Thresholds in the five sessions of the first training. (B) Thresholds in the test sessions for the locations of the first training (Loc1) and second training (Loc2). Error bars represent standard errors.

The first training improved participants’ discrimination performance, as revealed by the decreased threshold from the first session, mean, 2.58° ± 0.617°, to the fifth session, mean, 1.90° ± 0.707°, of the training phase, paired t test: t(15) = 4.23, p = 0.001, Cohen's d = 1.06, BF₁₀ = 50.65. The training effect was further evidenced by the significant main effect of session, F(1,15) = 152.32, p < 0.001, \({\rm{\eta }}_p^2\) = 0.91, BF₁₀ > 100, when we compared the estimated discrimination thresholds at Loc1 and Loc2 before and after the training sessions. The training effect was larger at Loc1 as revealed by the significant interaction between session and location, F(1,15) = 10.77, p = 0.005, \({\rm{\eta }}_p^2\) = 0.42, BF₁₀ = 18.92. The thresholds in the two locations were statistically indistinguishable in the pretest session, t(15) = 0.64, p = 0.53, Cohen's d = 0.16, BF₁₀ = 0.31. In the midtest session, the threshold at Loc1 was significantly lower than that of Loc2, t(15) = 4.56, p < 0.001, Cohen's d = 1.14, BF₁₀ = 89.43. 

The results revealed no significant effect of the second training, F(1,15) = 1.15, p = 0.301, \({\rm{\eta }}_p^2\) = 0.07, BF₁₀ = 0.5. The effects of the second training as defined by MPI_mid-post were not significantly different from those of experiment 2, F(1,30) = 1.47, p = 0.235, \({\rm{\eta }}_p^2\) = 0.05, BF₁₀ = 0.44; interaction: F(1,30) = 0.78, p = 0.383, \({\rm{\eta }}_p^2\) = 0.03, BF₁₀ = 0.52, but were significantly smaller than those of experiment 3, F(1,30) = 16.34, p < 0.001, \({\rm{\eta }}_p^2\) = 0.35, BF₁₀ = 46.22; interaction: F(1,30) = 2.66, p = 0.113, \({\rm{\eta }}_p^2\) = 0.08, BF₁₀ = 1.01. Re-emergence of the proactive interference after introducing this 6-hour interval confirmed the important role of reactivation-induced memory integration in experiment 3. 

Experiment 5

We then asked what was the role of location representation in the observed proactive interference in experiment 2. Simply paying attention to Loc2 seemed not sufficient to solve the interference because the second training of experiment 2 directed the participants’ attention to Loc2 and yet no learning occurred. It was the learning at Loc2 that was inhibited by the first training, potentially due to the learned low priority of attention during the first training (Huang & Li, 2023; Wang & Theeuwes, 2018). Experiment 3 demonstrated an effective approach to foster the integration of the two trainings and thus release the inhibition of new learning caused by the intensive first training. Because the two trainings differed only at their stimulus locations, we would expect that the coactivation of the representations of the two locations was a part of the integration process. To test this hypothesis, we conducted experiment 5 in which a tilt discrimination task at Loc2 was performed immediately after the reactivation and before the 6-hour interval. With this setting, any difference in performance change after the second training between experiments 4 and 5 would be attributed to the addition of the tilt discrimination task. 

The first training improved participants’ discrimination performance, as revealed by the decreased threshold from the first session, mean, 2.38° ± 0.645°, to the fifth session, mean, 1.91° ± 0.689°, of the training phase, paired t-test: t(15) = 2.75, p = 0.015, Cohen's d = 0.69, BF₁₀ = 3.94. The training effect was further evidenced by the significant main effect of session, F(1,15) = 70.24, p < 0.001, \({\rm{\eta }}_p^2\) = 0.82, BF₁₀ > 100, when we compared the estimated discrimination thresholds at Loc1 and Loc2 before and after the training sessions. The training effect was larger at Loc1 as revealed by the significant interaction between session and location, F(1,15) = 7.06, p = 0.018, \({\rm{\eta }}_p^2\) = 0.32, BF₁₀ = 14.59. The thresholds in the two locations were statistically indistinguishable in the pretest session, t(15) = 1.13, p = 0.277, Cohen's d = 0.28, BF₁₀ = 0.44. In midtest session, the threshold at Loc1 was significantly lower than that of Loc2, t(15) = 2.48, p = 0.025, Cohen's d = 0.62, BF₁₀ = 2.57. 

We expected that adding the tilt discrimination task would integrate the two locations and release the inhibition of new learning at Loc2. Indeed, we found that the effects of the second training were restored, F(1,15) = 17.87, p = 0.001, \({\rm{\eta }}_p^2\) = 0.54, BF₁₀ = 1.33 (Figure 6), because the training effects defined by MPI_mid-post were significantly larger than those of experiment 4, F(1,30) = 4.81, p = 0.036, \({\rm{\eta }}_p^2\) = 0.14, BF₁₀ = 1.3; interaction: F(1,30) = 2.43, p = 0.13, \({\rm{\eta }}_p^2\) = 0.08, BF₁₀ = 1.07, suggesting that the location-based learned attentional suppression may play an important role in the observed proactive interference. However, the restored effects were significantly smaller than those of experiment 3, F(1,30) = 6.95, p = 0.013, \({\rm{\eta }}_p^2\) = 0.19, BF₁₀ = 1.51; interaction: F(1,30) = 0.03, p = 0.872, \({\rm{\eta }}_p^2\) < 0.01, BF₁₀ = 0.29, and there was no further improvement at Loc1, one-sample t-test against 0 for the MPI_mid-post: t(15) = 0.10, p = 0.919, Cohen's d = 0.03, BF₁₀ = 0.26, after the second training, suggesting that only location representations were integrated to unlock the inhibition of learning at Loc2 in experiment 5. 

Figure 6.

 

Results of experiment 5. (A) Thresholds in the five sessions of the first training. (B) Thresholds in the test sessions for the locations of the first training (Loc1) and second training (Loc2). Error bars represent standard errors.

Discussion

A summary of the results across experiments is shown in Figure 7. The reactivation manipulations in experiments 3 and 5 revealed two related but different results (Figure 7B). In experiment 3, the reactivation of the first training occurred immediately before the second training, resulting in significant training effect at Loc2 and additional improvement at Loc1. In experiment 5, the reactivation of the first training was immediately followed by a tilt discrimination task at Loc2 and the second training was conducted 6 hours later. The second training could also be protected from the proactive interference in this procedure, but an additional performance improvement at Loc1 was not observed. Importantly, the performance improvement from the midtest to post-test session was smaller in experiment 5 as compared with that of experiment 3. We suggest a two-level mechanism of reactivation-induced memory integration to account for these results. 

Figure 7.

 

Summary of the results. MPIs after the (A) first training (MPI_pre-mid) and (B) second training (MPI_mid-post) are shown for each experiment. Mixed analyses of variance with experiment as a between-subjects factor and location (Loc1 and Loc2) as a within-subjects factor were conducted to examine the effects of the second training between experiments. Asterisks indicate the significant main effects of experiment. There was no significant interaction between experiment and location in the comparisons. Error bars represent standard errors. *p < 0.05, ***p < 0.001.

The first level of the mechanism is to release the inhibition of new learning at the previously untrained location (Loc2) by integrating the two locations during the reconsolidation process. This can be achieved by the coactivation of the two locations when the location representation of Loc2 was activated during the reconsolidation window of the reactivated first training at Loc1. In the present study, the representation of Loc2 was activated by the second training and the tilt discrimination task in experiment 3 and experiment 5, respectively. Accordingly, we have observed significant improvement at Loc2 after the second training in both experiments. However, only releasing the inhibition of new learning at Loc2 in experiment 5 was not sufficient to further elevate the performance at Loc1. We would attribute the additional improvement at Loc1 after the second training in experiment 3 to the second level of the mechanism. At the second level, the memory traces of the trainings at the two locations were integrated owing to their coactivation that occurred when the second training was conducted during the reconsolidation of reactivated first training. This process could form an integrated memory of the two trainings so that training at Loc2 would benefit the performance at both locations, as was evident in experiment 3. Together, these results suggest that which level of the mechanism dominates the reactivation-induced integration process depends on the representations that are coactivated during the reconsolidation. 

A recent theory named the nonmonotonic plasticity hypothesis (NMPH) suggests that coactivation of two overlapping memories could lead to memory integration or differentiation depending on the level of the coactivation (Ritvo et al., 2019; Wammes et al., 2022). The NMPH states that the two memories would be integrated if they are strongly coactivated. Otherwise, if one memory is strongly activated and the unique components of the other memory are moderately active, the coactivation would be moderate one and result in differentiation of the two memories. In experiment 3, to reactivate the first training, we simply had the participants performing the training task that differed from the second training only at the stimulus location. Hence, the activation of the first training during its reconsolidation was most likely at a high level given the concurrently performed second training. Similarly, the representations of the two locations were also likely to be strongly coactivated in experiment 5. In both experiments, the NMPH and other models of memory integration (Morton et al., 2017; Schlichting & Frankland, 2017) would predict the occurrence of integration that agreed with the observed results. 

The first training showed partial transfer of learning effect to the untrained location in all experiments, ps < 0.01 (Figure 7A). Partial transfer of learning effect after extensive training was observed in a number of previous studies with different training tasks, such as motion direction discrimination (Zhang & Li, 2010), Vernier task (Wang et al., 2014), and foveal orientation discrimination (Xiong et al., 2016). Particularly, in a recent study in our laboratory (Jia et al., 2021), a similar orientation discrimination task was trained for 6 days and the partial transfer of learning effect to the untrained location was also observed. Transfer of learning effect in perceptual learning depends on many factors and it is still an on-going topic that deserve further exploration. Importantly, the partial transfer of learning effect to the untrained location in all experiments was in sharp contrast with the blocking of the further learning at the untrained location in experiment 2. Previous studies have shown that the transfer of learning to untrained locations can be elevated by double training method that was suggested to release the attentional suppression to the untrained location by performing an unrelated task at that location (Wang et al., 2014; Xiao et al., 2008; Zhang et al., 2010). In these studies, the unrelated task was generally performed for several sessions in order to release the learned attentional suppression. This setting was consistent with the results in experiment 2, in which a short training session at the untrained location was not sufficient to release the inhibition of new learning. However, by adding a brief activation of the first training before the second training, we enabled the new learning at the untrained location and induced further improvement for the first training, demonstrating that reactivation-induced memory integration is a practical way to foster the efficiency of multiple perceptual learnings. Further, our results also validated the idea that reactivation-induced integration could facilitate the generalization of learning (Morton et al., 2017; Ritvo et al., 2019; Schlichting & Frankland, 2017). 

Our results suggest an encoding phase hypothesis that the integration occurred during the second training and does not support a retrieval hypothesis that pattern separation was induced by the reactivation to prevent interference at the retrieval stage. If pattern separation was realized during the reconsolidation process, it would likely to generate two separate associations for the two locations, and thus prevent their interference during the retrieval (Huang et al., 2022). However, the retrieval hypothesis cannot explain the additional performance improvement at Loc1 after the reactivation and second training in experiment 3. If the two trainings were dissociated during the reconsolidation, the second training would not benefit the performance at the location of the first training. Considering that contextual cueing could facilitate pattern separation and prevent interference of perceptual learning (Huang et al., 2022), we suggest that both the pattern separation (memory differentiation) and memory integration could benefit the efficiency of multiple perceptual learnings. The difference was the ways that memory differentiation and memory integration were induced, because contextual cueing could promote the former and reactivation plus new learning promotes the latter. 

The present study found that the reactivation prevented interference between the first and the second trainings by integrating the two trainings, which was consistent with a study in the area of motor learning (Herszage & Censor, 2017). However, there were other motor learning studies that revealed impaired performance of the preestablished motor memory (i.e., the first training) (Gabitov et al., 2017) or the newly acquired motor memory (i.e., the second training) (Gabitov et al., 2019) owing to the reactivation before the second training. An obvious difference between our study and Gabitov et al.’s was the domain of the learning. The other important difference was the different amounts of the second training. In our study, the amount of the second training was only 1 in 10 of the first training, while the first and second trainings were equalized in the studies of (Gabitov et al., 2017; Gabitov et al., 2019). The ratio between the two trainings might contribute to whether the reactivation causes integration or interference. However, this question remains open, to be further investigated. 

We also noted that the present study adopted MPI as the measurement of the training effects, instead of the raw performance difference between sessions. Percentage improvement is a widely used index for measuring training effect in perceptual learning literature (Shibata et al., 2017; Tan et al., 2019; Xiao et al., 2008; Xiong et al., 2016). The advantage of calculating the MPI is that it could potentially decrease the impact of the initial threshold on the training effect between individual subjects. However, adopting MPI may also undermine the potential confound that is introduced by large variation at baseline condition. We addressed this potential issue by analyzing the data with approximately matched pretest performance across experiments. The results replicated the findings reported here. However, we believe that a better solution to overcome this issue in future studies is to increase the sample size of the experiments or adopting within-subject experimental design. 

In summary, the present results suggest a two-level mechanism for reactivation-induced memory integration and its roles in preventing proactive interference in perceptual learning. These findings could be adopted in future applications of various forms of learning that suffer from the interference from previous learning experience. 

Acknowledgments

Supported by grants from Science and Technology Innovation 2030-Brain Science and Brain-Inspired Intelligence Project (2021ZD0200204) and the National Natural Science Foundation of China (32271104). 

All data have been made publicly available at the Open Science Framework and can be accessed at https://osf.io/n6vuz/?view_only=39756bfce90249fa8e933a88776e0e94. 

Commercial relationships: none. 

Corresponding author: Sheng Li. 

Email: [email protected]. 

Address: School of Psychological and Cognitive Sciences, Peking University, 5 Yiheyuan Road, Haidian, Beijing 10087, China.