(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Feature-specific reactivations of past information shift current neural encoding thereby mediating serial bias behaviors [1]

['Huihui Zhang', 'School Of Psychological', 'Cognitive Sciences', 'Beijing Key Laboratory Of Behavior', 'Mental Health', 'Peking University', 'Beijing', 'Pku-Idg Mcgovern Institute For Brain Research', 'Huan Luo']

Date: 2023-03

The regularities of the world render an intricate interplay between past and present. Even across independent trials, current-trial perception can be automatically shifted by preceding trials, namely the “serial bias.” Meanwhile, the neural implementation of the spontaneous shift of present by past that operates on multiple features remains unknown. In two auditory categorization experiments with human electrophysiological recordings, we demonstrate that serial bias arises from the co-occurrence of past-trial neural reactivation and the neural encoding of current-trial features. The meeting of past and present shifts the neural representation of current-trial features and modulates serial bias behavior. Critically, past-trial features (i.e., pitch, category choice, motor response) keep their respective identities in memory and are only reactivated by the corresponding features in the current trial, giving rise to dissociated feature-specific serial biases. The feature-specific automatic reactivation might constitute a fundamental mechanism for adaptive past-to-present generalizations over multiple features.

Funding: This work was supported by the National Science and Technology Innovation STI2030-Major Project 2021ZD0204100 (2021ZD0204103 to H.L.), National Natural Science Foundation of China (31930052 to H.L.), and China Postdoctoral Science Foundation (2020M680166 to H.Z.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Here, we aim to understand the dynamic neural mechanisms of how multiple features in the previous trial confront and modify the present, from which the feature-specific serial bias arises. In two experiments, participants performed an auditory categorization task with their brain activities recorded using electroencephalography (EEG). First, behavioral results exhibit concurrent component-specific serial biases, i.e., repulsive for tone pitch and motor response and attractive for category choice. Importantly, the neural representations of features in the previous trial—pitch, category choice, and motor response—are reactivated and emerge simultaneously as the corresponding features in the present trial. Most crucially, the current neural representations exhibit behaviorally congruent shifts, i.e., being attracted or repulsed from the past in a feature-specific way, which further correlates to serial bias behavior.

For serial bias to occur, previous trial information must leave memory traces that shape perception and decision-making in the current trial, a process that presumably involves memory reactivation and operation. It is known that long-term memory can be reactivated to enter and potentially impact working memory (WM) when necessary [ 19 , 20 ]. Interestingly, recent WM studies suggest that although information could be retained in an “activity-silent” state [ 21 – 27 ], memory operations still rely on the activation of neural representations [ 28 , 29 ]. Accordingly, previous studies have demonstrated that past-trial information is reactivated when processing new inputs, a process that potentially contributes to serial bias [ 30 – 32 ]. Moreover, a recent study observed that previous information remained in activity-silent traces reappeared before the stimulus, and the reactivation strength correlates with the strength of serial bias [ 33 ]. This is in line with computational modeling studies that advocate a purely “activity-silent” synaptic plasticity mechanism for serial bias, which can be enhanced by past information reactivation [ 33 – 35 ]. As for the direction of serial bias, i.e., attractive or repulsive, the neural basis remains debated. Using consecutive trials containing the same or orthogonal orientations, one fMRI study shows neural evidence for an attractive bias in early visual cortex [ 36 ]. In contrast, recent MEG and fMRI studies [ 37 , 38 ] indicate a repulsive bias during current sensory encoding in early visual cortex, but no evidence for an attractive bias. Based on computational modeling, they propose a higher-level attractive integration [ 38 ].

Crucially, the information contained in the past trial is not simple but rather rich and manifold, encompassing multitudes of features at various levels, e.g., physical properties, abstract categories, and response actions. Different features are found to be associated with different serial biases, in attractive or repulsive directions, even within the same task [ 7 , 8 , 11 – 13 ]. A two-stage model employing efficient coding and Bayesian inference reconciles the repulsive and attractive serial bias in perceptual decisions by, respectively, linking them to sensory adaptation and post-perceptual Bayesian inference ([ 5 , 14 ]; but see [ 15 – 18 ]). However, the neural mechanism underlying the multifeature serial bias remains obscure.

The regularities and recurrences of our world render the past always relevant to the present [ 1 ]. Indeed, perceptual experience and decision-making at any moment are constantly intertwined with previous information and likewise impact the future [ 2 ], a characteristic allowing for perceptual stability and adaptive optimization. The past-to-present influences could occur automatically and span a relatively long time scale. For example, the perceived feature in the current trial (e.g., location, orientation, and category) tends to be systematically shifted by that in the previous trial, even though trials are independent of each other and several seconds apart, namely the “serial bias” effect [ 3 – 10 ].

There are also some incongruencies between Experiments 1 and 2. First, motor response emerges simultaneously with pitch and choice, suggesting that instead of being an output stage, the motor system actively participates in decision-making at the early stage [ 11 , 42 ]. In addition, past-trial reactivation and the corresponding behavioral bias were not observed in Experiment 2. One possible explanation is that early planning of choice-motor mapping could disrupt latent memory of previous motor response traces.

Finally, Experiment 2 also replicated the behaviorally relevant neural shifting results, for both pitch and category choice ( Fig 5E ). Specifically, neural representations of the current-trial pitch (blue) and category choice (green) displayed repulsive and attractive shifts by previous features, respectively (cluster-based permutation test, one-sided, corrected; pitch: 224 to 314 ms and 334 to 384 ms after the tone onset, p < 0.01; category choice: 634 to 674 ms after the tone onset, p = 0.056). Moreover, the neural shifts were correlated with serial bias behavior for both pitch ( Fig 5F ) and category choice ( Fig 5G ), i.e., showing significant neural shifts for the High-bias but not Low-bias group (cluster-based permutation test, one-sided, corrected; pitch: 234 to 304 ms and 344 to 394 ms after tone onset, p < 0.05; category: 594 to 674 ms after tone onset, p = 0.017). Notably, consistent with the lack of behavioral serial bias for motor response ( Fig 4C ), we found neither significant past-trial reactivation ( Fig 5C , orange) nor neural shift for motor response ( Fig 5E , orange).

(A) Grand average decoding performance for current-trial features as a function of time following the sustained white noise, for pitch (blue), category choice (green), and motor response (orange). Note that the response–cue frame occurred before white noise onset (−0.5 s, dashed vertical line). Time 0 denotes the white noise onset. Vertical dashed lines from left to right denote the onset and offset of response–cue frame, and the pure tone onset, respectively. Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, p < 0.05, one-sided, corrected) for each feature. Shadows represent SEM. ( B ) Cross-correlation coefficient time course as a function of temporal lag, for each current-trial feature-pair (green: pitch vs. category choice; brown: pitch vs. motor response). Shadow represents 95% confidence interval (bootstrapping, N = 5,000). Horizontal colored lines denote significant temporal clusters (permutation test, one-sided, corrected, p < 0.05) for each feature pair. Horizontal dashed lines represent the 95% permutation threshold (corrected for multiple comparison). Vertical dashed line indicates the peak correlation coefficient indexing the relative activation lag. ( C ) The same as A but for past-trial features. ( D ) Cross-correlation coefficients between past-trial reactivation and current-trial neural response, as a function of temporal lag, for pitch (blue) and category choice (green). The peak at 0 ms temporal lag (vertical dashed line) supports their co-occurrence. Other denotations are the same as B. ( E ) Grand average neural representational shift (Shift_dist) as a function of time relative to the white noise onset, for pitch (blue), category choice (green), and motor response (orange). Positive and negative values correspond to attractive and repulsive direction, respectively. Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, one-sided, corrected, p < 0.05 for pitch, p = 0.056 for category) for each feature. Shadows represent SEM. ( F, G ) Participants were divided into two groups based on the serial bias behavior in pitch and category choice, respectively. Grand average Shift_dist of High-bias (colored lines) and Low-bias (black line) groups for pitch ( F ) and category choice ( G ). Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, one-sided, corrected, p < 0.05). Data supporting this figure found here: https://osf.io/4cwv7/?view_only=3a4885189ebf46aaacf05ef109821d03 .

Importantly, as shown in Fig 5 , Experiment 2 did not show an early reactivation of past-trial information after the response–cue frame or the white noise but, instead, largely replicated Experiment 1. First, neural representation of the current-trial features emerged after the tone stimulus ( Fig 5A , cluster-based permutation test, p < 0.001, one-sided, corrected; −26 to 834 ms for pitch; 14 to 884 ms for category choice; 44 to 1,044 ms for motor response), with similar latencies ( Fig 5B ; pitch versus category choice: correlation coefficients, 0 ms lag, bootstrapping, 95% CI = [−10, 20] ms, N = 5,000; pitch versus category choice: permutation test, one-sided, corrected, −360 to 400 ms; pitch versus motor: correlation coefficients, −40 ms lag, bootstrapping, N = 5,000; pitch versus motor: 95% CI = [−150, 0] ms, permutation test, one-sided, corrected, p < 0.05, − 490 to 360 ms). Second, past-trial pitch (blue) and category choice (orange) features could also be decoded from neural response ( Fig 5C ; cluster-based permutation test, one-sided, corrected; pitch: 154 to 554 ms and 574 to 674 ms after tone onset, p < 0.05; category: 84 to 584 ms and 714 to 824 ms after tone onset, p < 0.040). Critically, similar to the findings of Experiment 1, past-trial reactivations co-occurred with the current-trial neural response of features, for both pitch and category choice ( Fig 5D ; cross-correlation coefficients: 0 ms lag for both pitch and category choice, bootstrapping, 95% CIs = [0, 0] ms, N = 5,000; permutation test, one-sided, corrected, pitch: −390 to 380 ms, category choice: −340 ms to 400 ms).

First, participants exhibited similar serial bias behavior as Experiment 1. All the history-dependence models outperformed the history-free model ( Fig 4B ; Model 2 versus Model 1: ΔAIC = −2,148; Model 3 versus Model 1: ΔAIC = −1,151; Model 4 versus Model 1: ΔAIC = −4,406). Furthermore, the pitch and category choice showed negative (regression coefficient: mean = −0.76; 95% CI = [−0.96, −0.55]; t(58,196) = −7.24, p < 0.001, one-sample t test) and positive serial bias (mean = 1.74; 95% CI = [1.51, 1.97]; t(58,196) = 14.90, p < 0.001, one-sample t test), respectively ( Fig 4C , left). Meanwhile, Experiment 2 did not show serial bias for motor response ( Fig 4C right, “Switch”: mean = 0.50, SD = 0.014; “Stay”: mean = 0.50, SD = 0.014; “Switch” versus “Stay”: paired-sample t test, t(29) = −0.66, p = 0.52), consistent with previous results [ 11 ].

( A ) Experiment 2 employed the same auditory categorization paradigm as Experiment 1, with only one difference, i.e., the response–cue screen appeared before the white noise (dotted box). Participants categorized the following 0.25-s pure tone (dark blue) embedded in a 1.75-s sustained white noise (grey line) into “high” or “low” category, by using the corresponding hand defined in the response–cue frame (dotted box) appeared at the beginning of each trial. ( B ) Model comparison results. ΔAIC of Model 2 (blue; current trial + previous pitch), Model 3 (green; current trial + previous category choice), and Model 4 (grey; current trial + previous pitch + previous category choice), compared to Model 1 (current trial only). ( C ) Left: Regression coefficients for previous pitch (blue) and previous category choice (green) extracted from the winning model (Model 4, * in model comparison). Error bars represent 95% confidence interval. Right: No motor response serial bias, i.e., no difference between “Switch” trials (dark orange) and “Stay” trials (light orange). Each circle represents individual participant. (***: p < 0.001). Data supporting this figure found here: https://osf.io/4cwv7/?view_only=3a4885189ebf46aaacf05ef109821d03 .

In Experiment 1, pitch and category choice information of the previous trial was reactivated by the corresponding event, i.e., pure tone ( Fig 2B ). Meanwhile, since the pure tone occurred early in each trial ( Fig 1A ), it might be the first task-relevant event (i.e., pure tone) rather than the feature-specific event that reactivated the past-trial information. To test the possibility and also confirm the findings of Experiment 1, we designed Experiment 2 (N = 30) during which participants performed the same auditory categorization task as Experiment 1, except now the response–cue frame appeared at the beginning of each trial ( Fig 4A ). If the pitch and category choice reactivations are indeed due to the early task-relevant event, we would expect their reactivations right after the response–cue frame. Moreover, since the white noise occurred at a fixed temporal lag after the response–cue frame in Experiment 2 ( Fig 4A ), we could also test whether the reactivation simply derives from temporal prediction.

The lack of behavioral correlations could be a result of the low statistical power of pitch neural shift analysis given its 5 levels. In order to address this issue, we performed a group-division analysis to boost the signal-to-noise ratio. Specifically, for each feature, all the participants (N = 30) were divided into two groups of the same size—High-bias and Low-bias—based on their serial bias in behavior (see Methods for details). As shown in Fig 3C, 3D and 3E (right), the High-bias group displayed a significant neural shift for pitch ( Fig 3C , blue; cluster-based permutation test, one-sided, corrected, 204 to 244 ms and 654 to 724 ms after the tone onset, p < 0.05), category choice ( Fig 3D , green; 484 to 664 ms after the tone onset, p < 0.001), and motor response ( Fig 3E , orange; 64 to 224 ms after the response cue onset, p < 0.001), while not for the Low-bias group ( Fig 3C, 3D and 3E , black line).

Finally, by examining two neural indexes—past-trial reactivation and neural shift—we assessed the behavioral relevance of the neural signature of serial bias. First, as shown in Fig 2E , the past-trial reactivation strength, defined as the mean decoding performance within the significant temporal clusters ( Fig 2B ), is correlated with the corresponding serial bias behavior across participants, for pitch, category choice, and motor response (Pearson’s correlation, N = 30, r = −0.42, p = 0.022 for pitch; N = 29, r = 0.40, p = 0.033 for choice; N = 29, r = 0.39, p = 0.036 for choice). Note that the negative correlation for pitch is due to its repulsive serial bias effect. Meanwhile, as shown in Fig 3C, 3D and 3E (left), the inter-subject correlation between neural shift and behavior is only significant for category choice (Pearson’s correlation, N = 30, r = 0.40, p = 0.028) and motor response (Pearson’s correlation, N = 29, r = −0.52, p = 0.0039), but not for pitch (Pearson’s correlation, N = 29, r = −0.29, p = 0.88).

Fig 3B plots the neural shift for the three features, with positive and negative values denoting attraction towards and repulsion from the past feature, separately. Pitch (blue line) showed a repulsive bias, rising shortly after pure tone (cluster-based permutation test, two-sided, corrected; 294 to 344 ms after the tone onset, p = 0.0056), consistent with the repulsive direction in serial bias behavior ( Fig 1B and 1D ). In contrast, the category choice (green line) displayed an attractive shift towards previous information, occurring relatively late (cluster-based permutation test, two-sided, corrected; 504 to 574 ms after the tone onset, p = 0.0058), congruent with the attractive bias in behavior ( Fig 1C and 1D ). Finally, the neural code of motor response (orange line) was shifted away from that of the prior trial (cluster-based permutation test, two-sided, corrected; 74 to 194 ms after the response cue, p = 0.005), again in line with behavior ( Fig 1E ). Notably, the neural shift results still held when trials preceded by unambiguous pitches (i.e., f1, f5) were removed ( S4E Fig ), consistent with the behavioral results ( S4A , S4B and S4C Fig ). Thus, the findings demonstrate direct neural evidence for serial bias, revealing a behaviorally congruent shift in neural representation for pitch, category choice, and motor response.

(A) Illustration of neural representational shift analysis. Left: Influence of previous f2 on current f1 as an example. Neural templates were built for f1 (dark blue circle) and f2 (light blue circle) base on the k-fold cross validation process. For each trial with current f1 and previous f2, its neural distance to template f1 and template f2 were computed, resulting in D1, D2, and their difference M_diff (D2-D1). As baselines, trials that have current f1 but previous f1 were selected and were analyzed using the same analysis, resulting in M_diff baseline . The difference between M_diff baseline and M_diff (Shift_dist) characterizes the neural representational shift in f1 by prior f2. Right: Neural representation of current-trial f1 is attracted toward (upper, positive Shift_dist values), repulsed from (lower, negative Shift_dist values), or not affected (middle, around zero Shift_dist values) by prior f2. ( B ) Grand average neural representational shift (Shift_dist) as a function of time following white noise onset, for pitch (blue), category choice (green), and motor response (orange). Positive and negative values correspond to attractive and repulsive direction, respectively. Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, two-sided, corrected, p < 0.05) for each feature. Shadows represent SEM. ( C-E ) Behavioral correlates of neural shift. Left: Scatterplot of behavioral bias vs. neural shift for pitch ( C ), category choice ( D ), and motor response ( E ). Neural representation shift is the mean shift_dist across significant clusters in B. Serial bias behaviors for pitch, category choice, and motor response are indexed by regression coefficients of past pitch, past category choice, and difference between “switch” and “stay,” respectively. Each dot represents individual participant, with black dots denoting excluded participants as outliers. Right: Participants were divided into two groups based on the serial bias behavior in pitch, category choice, or motor response, respectively. Grand average neural representational shift (Shift_dist) of High-bias (colored lines) and Low-bias (black line) groups for pitch ( C ), category choice ( D ), and motor response ( E ). Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, one-sided, corrected, p < 0.05). Data supporting this figure found here: https://osf.io/4cwv7/?view_only=3a4885189ebf46aaacf05ef109821d03 .

Fig 3A (left panel) exemplifies the general idea to test pitch serial bias, e.g., quantifying the influence of past-trial f2 on current-trial f1. First, we built the neural templates for f1 (dark blue) and f2 (light blue) based on the k-fold cross validation process (see details in Methods ). Next, we chose trials (current f1, previous f2) and computed the neural distance between each trial in the subsample and the f1 and f2 templates, separately, yielding D1 and D2, from which M_diff (D2-D1) was obtained. For comparison, we chose trials (current f1, previous f1) as baselines and did the same distance computation, yielding the M_diff baseline (D2-D1). Finally, the difference between M_diff and M_diff baseline was calculated (Shift_dist) to quantify the influence of the previous f2 on the neural representation of current f1. If previous f2 attracts f1 neural representation, we would expect smaller M_diff than M_diff baseline , yielding positive Shift_dist values, and vice versa. There are three possibilities—attractive, no effect, and repulsive ( Fig 3A , right panel)—corresponding to positive, around zero, and negative Shift_dist values, respectively. A similar idea has been applied to category and motor response (see Methods for details; S3 Fig ). The neural shift analysis was performed at each time point, for each feature, and in each participant.

Even though past-trial reactivations co-emerge with current-trial neural responses, it remains unclear at what stage and in which direction (attractive or repulsive) past-trial information biases current neural processing. We thereby examined direct neural evidence for serial bias. Specifically, we developed a novel analysis by accessing whether the neural representation of the current-trial feature would be attracted toward or repulsed from the preceding feature.

Most importantly, as shown in Fig 2D , all the three features showed temporally aligned profiles between current-trial neural responses and past-trial reactivations (permutation test, one-sided, corrected, p < 0.05; pitch: blue, −460 to 420 ms; category choice: green, −410 to 440 ms; motor response: orange, −380 to 370 ms), peaking at 0 ms time lag (bootstrapping, N = 5,000, 95% CIs = [0 0] ms for pitch and category choice, and [−60, 110] ms for motor response), advocating simultaneous activation of present and past information for each feature in the present trial.

We next examined the temporal relationship between the current-trial neural responses and past-trial reactivations, by calculating their cross-correlation coefficients over time. First, features in the current trial showed varied latencies in their neural response ( Fig 2C ). Specifically, the pitch versus category choice (green line) correlation coefficient was significant from −400 to 520 ms time lag (permutation test, one-sided, corrected, p < 0.05), peaking at 10 ms lag (bootstrapping, 95% CI = [0, 70] ms, N = 5,000), while the pitch versus motor response (brown line) showed peak around −250 ms lag (bootstrapping, N = 5,000, 95% CI = [−270, −230] ms; permutation test, one-sided, corrected, p < 0.05, − 600 to 300 ms). Thus, pitch lagged category choice by 10 ms and led motor response by 250 ms. The latter result is well expected given the 250 ms interval between pure tone and response cue frame.

Overall, features of the preceding trial occurring seconds before and retained in memory are reactivated by specific events in the current trial, i.e., pitch and category choice by the tone stimulus and motor response by the response cue. It is noteworthy that past-trial features, given their maintenance in memory, could potentially be reactivated by any triggering event in the current trial, yet the findings support a feature-specific reactivation temporal profile.

Most importantly, we found that past-trial features were reactivated by the corresponding event in the current trial ( Fig 2B ). Specifically, previous-pitch decoding (blue line, Fig 2B ) was at chance level prior to the pure tone and rose right after tone onset (cluster-based permutation test, one-sided, corrected; significant clusters: 184 to 524 ms, 584 to 774 ms, and 954 to 1,024 ms, p < 0.05). Similarly, previous category choice decoding (green line, Fig 2B ) emerged after the tone onset (cluster-based permutation test, one-sided, corrected, significant clusters: 84 to 154 ms, 184 to 634 ms, and 654 to 784 ms, p < 0.05). In contrast, rather than being triggered by the pure tone, previous motor response (orange line, Fig 2B ) occurred after the response cue frame (cluster-based permutation test, one-sided, corrected; clusters: 164 to 244 ms and 264 to 604 ms after the response cue onset, p < 0.05). Moreover, future trial information could not be decoded ( S2 Fig ), and even after excluding trials preceded by f1 and f5 that belong to the unambiguous categories, past-trial pitches and category choices still showed significant reactivations in the current trial ( S4D Fig ).

Interestingly, as shown in Fig 2B , past-trial features, i.e., pitch (blue), category choice (green), and motor response (orange) could also be decoded form the neural response of the current trial. Crucially, to exclude potential current-trial confounding when decoding past-trial features, we performed the past-trial decoding analysis for each of the same current features and then combined the results (see Methods for details).

Overall, multiple features of the current trial could be successfully decoded from the neural response, i.e., the pitch and category information about the pure tone emerges right after the tone onset, and the neural code of motor response arises following the response cue frame.

(A) Grand average decoding performance for current-trial features as a function of time following the sustained white noise, for pitch (blue), category choice (green), and motor response (orange). The pure tone (blue rectangle) was embedded in a 2.25-s sustained white noise (grey horizontal line). Vertical dashed lines from left to right denote the tone onset and response cue frame, respectively. Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, p < 0.001, one-sided, corrected) for each feature. Shadows represent SEM. ( B ). The same as A but for past-trial features. Note that the past-trial decoding is performed by fixing the current-trial feature. ( C ). Cross-correlation coefficients as a function of temporal lag for each current-trial feature pair (green: pitch vs. category; brown: pitch vs. motor response). Shadow represents 95% confidence interval (bootstrapping, N = 5,000). Horizontal colored lines denote significant temporal clusters (cluster-based permutation test, one-sided, corrected, p < 0.05) for each feature pair. Horizontal dashed lines represent the 95% permutation threshold (corrected for multiple comparison). Vertical dashed line indicates the significant peak indexing the relative activation lag. ( D ) Cross-correlation coefficients between past-trial reactivation and current-trial neural response, as a function of temporal lag, for each feature (blue: pitch; green: category choice; orange: motor response). The peak at 0 ms temporal lag (vertical dashed line) supports the co-occurrence of past-trial reactivation and current-trial neural response. ( E ) Correlation between past-trial reactivation strength and serial bias behavior across participants (left: pitch; middle: category choice; right: motor response). Reactivation strength is calculated as the mean decoding performance across significant clusters in B. Serial bias behaviors for pitch, category choice, and motor response are indexed by regression coefficients of past pitch, past choice, and difference between “switch” and “stay,” respectively. Each dot represents individual participant, with black dots denoting excluded participants as outliers. Data supporting this figure found here: https://osf.io/4cwv7/?view_only=3a4885189ebf46aaacf05ef109821d03 .

As shown in Fig 2A , the pitch (blue) and reported category (green) of the current trial could be decoded shortly after the pure tone (cluster-based permutation test, p < 0.001, one-sided, corrected; pitch: 84 to 1,144 ms after tone onset; category: 84 to 934 ms after tone onset). Furthermore, the motor response decoding ( Fig 2A , orange) occurred after the response cue frame (cluster-based permutation test, p < 0.001, one-sided, corrected; motor response: 44 to 994 ms after response cue frame). This is well expected since participants could only determine the responding hand after the response cue.

We first performed a time-resolved multivariate decoding analysis [ 27 , 39 – 41 ] to examine the neural representations of multiple features of the current trial—pitch, category, and motor response. A linear regression analysis was used to fit the neural dissimilarity to feature dissimilarity, and the time-resolved regression coefficients denote the decoding performance, for each feature, at each time point and in each participant (see details in Methods ).

Four generalized linear mixed-effects models (GLMMs) were built to account for the behavioral performances. Model 1 assumes that the category choice only depends on the current pitch and a constant bias (base model; no serial bias). Models 2 and 3 consider additional contributions from previous pitch or previous category choice, respectively. Model 4 takes account of contributions from both pitch and category choice in the preceding trial. Models 2 to 4 were evaluated against Model 1 by comparing the Akaike information criterion (AIC) values. As shown in Fig 1D (left), all the three models were better than Model 1 (Model 2: ΔAIC = −2,079; Model 3: ΔAIC = −657; Model 4: ΔAIC = −4,369), indicating the influence of past-trial information on current perception. Moreover, Model 4 outperformed Model 2 and Model 3, supporting that pitch and category choice history together affect the current category decision. Importantly, as shown in Fig 1D (right panel), Model 4 showed negative coefficients for past-trial pitch (mean = −0.79; 95% CI = [−0.89, −0.69]; t(58,002) = −15.44, p < 0.001, one-sample t test), but positive coefficients for past-trial category choice (mean = 1.70; 95% CI = [1.41, 1.99]; t(58,002) = 11.59, p < 0.001, one-sample t test), confirming the repulsive and attractive serial bias for pitch and category choice, respectively. Moreover, the same analysis after excluding trials preceded by f1 and f5, the two pitches associated with unambiguous categories, showed the same results ( S4A , S4B and S4C Fig ), supporting that the repulsive and attractive serial biases for pitch and category choice do not just arise from the preceding unambiguous pitches.

As shown in Fig 1B and 1C , both pitch and category choice exhibited the serial dependence effect, yet in different directions. Specifically, the pitch of the current tone tends to be perceived away from that of the preceding pitch, i.e., repulsive serial bias ( Fig 1B ). For example, the same f3 on the current trial was more likely categorized into the “high” category when preceded by lower pitch (light blue) compared to when preceded by higher pitch (dark blue). In contrast, the categorization performance displayed an attractive serial bias, such that the same f3 on the current trial was more likely to be reported as the “high” category when the previous choice was also “high” ( Fig 1C ). Moreover, this attractive bias decreased as the distance between previous and current pitches increased (M 0 = 0.46, 95% CI 0 = [0.38, 0.51]; M 1 = 0.27; 95% CI 1 = [0.21, 0.32]; M 2 = 0.11, 95% CI 2 = [0.06, 0.14]; M 3 = −0.00, 95% CI 3 = [−0.038, 0.026]; M 4 = −0.02, 95% CI 4 = [−0.073, 0.026]; ps < 0.001 for distance ≤ 2), suggesting it is not a repetitive bias caused by lapsing.

( A ) Auditory categorization task paradigm. Upper: Pitch of the pure tone stimulus was pseudorandomly selected from 5 fixed frequencies (f1, f2, f3, f4, and f5) between 180 Hz (low category; light green) and 360 Hz (high category; dark green). Lower: In each trial, participants categorized a given 0.25-s pure tone (dark blue; selected from f1, f2, f3, f4, and f5) embedded in a 2.5-s sustained white noise stream (grey line) into “high” (closer to 360 Hz) or “low” (closer to 180 Hz) category. After the pure tone, a response–cue screen appeared, based on which participants used the corresponding hand to made responses (right inset). ( B ) Pitch serial bias (aggregate results across participants). “High” category choice percent as a function of the current pitch, for different pitch in previous trial (Light-to-ark color lines denote low-to-high pitches). Each circle represents aggregated data for each condition. Solid lines represent the logistic regression fits. ( C ) Category serial bias (aggregate results across participants). “High” category choice percent as a function of the current pitch, for different category choice in the previous trial (light green: low category choice; dark green: high category choice). The subpanel shows how the category choice serial bias is modulated by the previous-current pitch distance. Error bars represent 95% confidence interval. ( D ) Model comparison results. Left: ΔAIC of Model 2 (blue; current trial + previous pitch), Model 3 (green; current trial + previous category choice), and Model 4 (grey; current trial + previous pitch + previous category choice), compared to Model 1 (current trial only). Right: Regression coefficients for previous pitch (blue) and previous category choice (green) extracted from the winning model (Model 4, * in model comparison). Error bars represent 95% confidence interval. ( E ) Motor response serial bias (left or right hand). “Switch” trials (dark orange): percentage of trials that are different from previous trial in motor response. “Stay” trials (light orange): percentage of trials that are the same as previous trial in motor response. Each circle represents individual participants. (**: p < 0.01, ***: p < 0.001). Data supporting this figure found here: https://osf.io/4cwv7/?view_only=3a4885189ebf46aaacf05ef109821d03 .

In Experiment 1, 30 participants performed an auditory pitch categorization task with their 64-channel EEG activities recorded, by pressing buttons to indicate the category (“high pitch” or “low pitch”) of a pure tone embedded in a sustained white noise ( Fig 1A ). The pitch of the tone stimulus was pseudorandomly selected from 5 fixed frequencies between 180 Hz and 360 Hz (f1, f2, f3, f4, and f5), which were individualized to normalize task difficulty across participants (see details in Methods ). Moreover, to dissociate category choice and motor response, a response–cue frame appeared after the tone stimulus, based on which participants used the corresponding hand to make choices. Note that participants learned the definition of high- and low-pitch categories in a pretest, i.e., listening to 180 Hz and 360 Hz pure tones ( S1 Fig ).

Discussion

In two auditory categorization experiments, we reveal a new feature-specific reactivation neural mechanism for serial bias that occurs spontaneously across trials. Specifically, serial bias arises from the co-emergence of past reactivation and the neural coding of present inputs, which renders their interactions and the shifting of current information. Importantly, we provide converging behavioral and new neural evidence that the serial bias occurs in a feature-dissociated manner, i.e., sensorimotor (i.e., pitch and motor response) and abstract (i.e., category choice) features are repulsed from and attracted to the corresponding ones in the previous trial, respectively, within the same task. Moreover, the past-trial reactivation is not due to either early task-relevant events or temporal prediction but triggered by the occurrence of the feature-specific event. Taken together, features in the current trial automatically trigger the corresponding memory traces of past trials, and their co-occurrence leads to serial bias. This reactivation might reflect a fundamental operational mechanism for past-to-present adaptive optimization.

The past-to-present influences occur in a wide range of paradigms and contexts [2,43,44], of which the serial bias represents an extreme phenomenon. It refers to the involuntary, systematic shifting of current perception by previous trials and appears on many features [3,4,6–10,15,45]. The typical serial dependence effect is in an attractive direction [3,10,15,45], while the repulsive bias and individual difference have also been observed [7,8,46,47]. Here, by carefully disentangling features in the experimental design, we revealed, within the same task, the dissociated serial bias for pitch, category, and motor response. We found that pitch and motor response were repulsed from past information, while the category displayed an attractive direction.

What accounts for the dissociated bias direction for pitch and category choice? We postulate that the repulsive bias for pitch reflects a sensory adaptation process. First, the repulsive bias occurs not just for unambiguous pitches in the previous trial (f1, f5), wherein participants might have a reverse expectation (e.g., after “low” categorical choice for f1, they expect the next to be a “high” category), but also for pitches associated with ambiguous categories (i.e., f2, f3, f4). Second, the neural shift of pitch occurs at a relatively early latency, indicating its sensory nature, consistent with previous findings [37,38]. About the attractive serial bias for category choice, there are currently two major views. One posits that both attractive and repulsive serial biases arise at the perceptual stage [15–18], while the two-stage framework attributes the repulsive and attractive serial bias to perceptual and post-perceptual stages, respectively [5,8,14,38,48], by incorporating efficient coding and Bayesian inference. Our results could not distinguish the two views, since the EEG recordings lack enough resolution to localize the involved brain regions. Future studies are needed to examine the spatiotemporal neural correlates of the attractive serial bias.

For the past to influence the present, either voluntarily or spontaneously, prior information should leave traces in memory. It has long been viewed that the WM process relies on persistent firing [49–52] or frequency-specific neural oscillations [53,54]. Interestingly, information could also be retained in an “activity-silent” way [21–25]. However, the “activity-silent” mode only passively maintains information [22] (including ours; see [55–57]), and memory manipulation still relies on the reactivations of WM to active states [28,29]. Consistent with the view, we demonstrate the co-emergence of past reactivation and present information, followed by the shifted neural representation of current features. Moreover, the reactivation profiles encompass multiple feature-specific reactivations that arise simultaneously as the corresponding features in the current trial. Thus, the memory traces are reactivated by corresponding features in the current trial, and the feature-specific co-occurrence in neural space contributes to the serial bias.

A recent interesting study revealed brief reactivation of past spatial information during the intertrial interval (ITI) [33], while here we observed the concurrent emergence of past reactivation and present information after the stimulus, as also revealed in other EEG findings [30–32]. Both findings support past reactivation from an activity-silent to an activity-based state for serial bias to occur, but at different time points, i.e., prestimulus versus poststimulus. The difference might arise from temporal anticipation in the task design, with a fixed ITI in the previous study but a randomized ITI in our experiment. Indeed, it is found that alpha-band power carries past-trial information only when the stimulus is temporally predictable [33]. Importantly, our Experiment 2 excludes the temporal expectancy interpretation, since even though the white noise appears at a fixed time after the response–cue frame, pitch reactivations are still triggered by the auditory tone rather than the white noise.

Due to the different tasks implemented here, we have opted to decode broad-band neural signals rather than specific neural oscillations. Alpha-band power has been mainly found in spatial WM tasks (e.g., [33,53,54,58]), while here participants performed an auditory categorization task entailing no spatial memory or attention. Additionally, we aimed to find the neural signature of multiple features, so focusing on broad-band signals rather than specific neural rhythms would be a more conservative choice. In fact, broad-band decoding has been widely used in a series of WM studies (e.g., [22,27,56,57]) and serial dependence works [30–32].

Partially consistent with previous findings [33], our alpha-band analysis during the prestimulus period revealed reactivations of past choice and motor response, but not pitch information (S5 Fig). Given the fixed time interval between noise and tone, our results indicate that the temporal anticipation of response-related features (choice and motor response) reactivates the corresponding information within the alpha-band. Moreover, persistent oscillatory activities have been found to convey past choice and motor information [11,12]. It is therefore also possible that prestimulus reactivation carried by alpha-band power reflects an enhancement of weakly but actively coded memory traces via an anticipation-induced decrease in cross-trial variability [59]. Nevertheless, reactivation on its own, either before or after stimulus, could not fully characterize the dynamic operation of serial bias. This is because reactivation only signals the presence of past-trial information, but how the past influences the present, as shown by serial bias behavior, still cannot be well accounted for. Furthermore, computational models postulate an integration of activity-based and activity-silent mechanisms for serial bias, i.e., activity-silent synaptic plasticity leads to serial bias that can be enhanced by past reactivation in WM tasks [33–35]. Meanwhile, they could not account for multifeature serial biases in non-WM tasks. Here, we provide direct neural evidence that current information is indeed shifted towards or away from the past (i.e., attractive, repulsive) following their co-occurrence.

Importantly, we extend previous single-feature emphasis to a more general framework, wherein multiple features and their respective serial biases are disentangled in both behavior and neural representations. It is recognized that even the simplest perceptual decision task encloses packs of features, e.g., physical properties, abstract categories, and response actions, collectively constituting an “event-file” [60]. Here within the same auditory categorization task, the past “event-file” is implicitly imprinted in memory and automatically passed to the next “event-trial” in a feature-encapsulated way. Importantly, not any events can trigger past information, also in line with previous findings [56,57]. For instance, white noise, despite being an auditory sound within the same sensory modality, failed to reactivate prior pitch information, and previous motor response, although apparently retained in memory, could not be triggered by either white noise or tone stimulus. Moreover, the fact that white noise reactivates memorized pitch [27,56] during maintenance but not here indicates different mechanisms for WM and serial bias. While WM task involves explicitly storing information in WM, serial bias automatically occurs without requiring participant to voluntarily retain past-trial information. The findings that white noise could not reactivate past-trial information but transcranial magnetic stimulation (TMS) could [33,61] are in line with the interpretation that white noise impulses do not reactivate synaptic traces [59]. Overall, past trial automatically leaves a memory trace within which multiple features keep their identities and specificities, exerting their respective impacts on the future.

Serial bias denotes a dramatic case about past-to-present influence since information in the previous trial, which occurs several seconds ago and in principle should be discarded, still biases the present trial. The phenomenon is therefore independent of several factors involved in many other paradigms, such as voluntary attention, task-relevant modulation, task-irrelevant capture, etc. [22,62,63]. Moreover, as trials are typically several seconds apart with random ITI in-between, serial bias could not arise from within-trial temporal effects either [43,44]. Furthermore, while potentially sharing similar WM storage underpinnings, serial bias essentially differs from WM studies that would instruct participants to at least retain certain information. Our findings thus implicate a presumably ubiquitous mechanism for temporal dependence in many cognitive processes, such as perception, attention, memory, and decision-making.

Taken together, every single present is intertwined with the past, yet the past-to-present influence is not feature-agnostic but feature-specific, allowing for information encapsulation through trial-by-trial updates. The current feature reactivates the corresponding traces left in memory, and their co-occurrence induces the neural interactions and generate serial bias, thereby facilitating automatic adaptive generalizations from past to present.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002056

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.