Dynamics of history-dependent perceptual judgment

Identical physical inputs do not always evoke identical percepts. To investigate the role of stimulus history in tactile perception, we designed a task in which rats had to judge each vibrissal vibration, in a long series, as strong or weak depending on its mean speed. After a low-speed stimulus (trial n − 1), rats were more likely to report the next stimulus (trial n) as strong, and after a high-speed stimulus, they were more likely to report the next stimulus as weak, a repulsive effect that did not depend on choice or reward on trial n − 1. This effect could be tracked over several preceding trials (i.e., n − 2 and earlier) and was characterized by an exponential decay function, reflecting a trial-by-trial incorporation of sensory history. Surprisingly, the influence of trial n − 1 strengthened as the time interval between n − 1 and n grew. Human subjects receiving fingertip vibrations showed these same key findings. We are able to account for the repulsive stimulus history effect, and its detailed time scale, through a single-parameter model, wherein each new stimulus gradually updates the subject’s decision criterion. This model points to mechanisms underlying how the past affects the ongoing subjective experience.

2 Supplementary Figure S1. Evolution of performance during the first 100 trials and illustration of the number of trials needed to infer a significant separation between PSEs in different stimulus ranges. a. Probability of categorizing stimuli as "strong" as a function of speed, with a stable range and boundary. Psychometric functions were fitted averaging data across all rats (corresponding to Figure 2a) from designated single trials at different stages of the session. b. Rank-sum tests between PSE values in the "high range" and in the "low range" conditions, based on trials from start of session to Trial # indicated on the abscissa. For each condition and each individual trial, a single PSE value was computed, averaging across sessions. Thus, for "high range" sessions the PSE was computed considering only trial 1, only trial 2, and so on until the last trial available; the same procedure was carried out for the "low range" sessions. The PSEs were then grouped cumulatively up to a given Trial # and tested for their difference between conditions. For example, for Trial # = 10, the 10 different PSEs computed from the first 10 trials of the "high range" were compared to the 10 different PSEs computed from the first 10 trials of the "low range". The p-values obtained by comparing the two groups of PSEs for each Trial # are reported on the ordinate, up to Trial # = 70. The dashed line denotes p-value = 0.05. Source data are provided as a Source Data file.
3 Supplementary Figure S2. Gradual psychometric function shift, excluding trials with incorrect n-1 choice, suggests that n-1 speed drives the repulsive bias. a. Probability of categorizing stimulus n as "strong" as a function of trial n speed, with curves grouped by speed of trial n-1, only using correct choices in trial n-1. Darker curves correspond to higher n-1 speed. Blue squares denote PSE. b. Bias of the trial n psychometric curve, depending on speed from trial n-1 (far left plot) to trial n-4 (far right plot), after removing incorrect trials in the corresponding trial n-1 to trial n-4. Shading of the squares denotes speed in trial n-1.
Squares correspond to 6 individual rats. Source data are provided as a Source Data file. 4 Supplementary Figure S3. Psychometric function parameters given n-1 choice and trial outcome. a. Distributions of parameters obtained by bootstrapping the psychometric functions grouped by n-1 choice (same data as shown in Fig. 3b). Trials were sampled with replacement (n = 1000), and both curves (n-1 judged weak, n-1 judged strong) were estimated for each artificial sample. A null difference could not be rejected for any parameter (90% confidence level). b. Distributions of parameters obtained by bootstrapping the psychometric functions grouped by trial n -1 outcome (shown in Fig. 3c). Trials were sampled with replacement (n = 1000), and both curves (n-1 incorrect, n-1 correct) were estimated for each artificial sample.
Confidence intervals around the resampled parameters did not exclude a null difference in midpoint or slope (confidence level < 90%), however it evidenced a difference in the asymptote parameters. Both lapse parameters,  (lower lapse rate) and  (upper lapse rate) slightly decreased after an incorrect trial (confidence = 92%, 97%, respectively), corresponding to an increase in lapse. Source data are provided as a Source Data file. 5 Supplementary Figure S4. Psychometric function shift and PSE regression slope, excluding trials with correct n-1 choice. a. Probability of categorizing stimulus n as "strong" as a function of trial n speed, with curves grouped by speed of trial n-1, only using incorrect choices in trial n-1. Darker curves correspond to higher n-1 speed. Blue squares denote PSE.
b. Bias of the trial n psychometric curve, depending on speed in trial n-1, after removing correct trials in trial n-1. Shading of the squares denotes speed in trial n-1. Squares corresponds to data, grouping all 6 rats. Source data are provided as a Source Data file. Qualitative predictions for a history-dependent change in lapse rate modulated by ITI duration.
Left: Two pairs of history-dependent psychometric curves following a short ITI (light blue) and long ITI (dark blue). The two pairs of curves have the same history-dependent midpoint 8 parameters, but asymmetrically different lapse rates: after a long ITI, the curves undergo a vertical shift in the opposite direction of the previous stimulus. Right: Lines represent the absolute difference between each pair of psychometric curves shown in the left plot. A historydependent change in lapse rate modulated by ITI duration would lead to a constant separation between the two lines, meaning that the effect of the ITI is linearly combined with the effect of the previous stimulus b. Qualitative predictions for a history-dependent change in midpoint, but not lapse rates, modulated by ITI duration. Left: Two pairs of history-dependent psychometric curves following a short ITI (light blue) and long ITI (dark blue). The two pairs of curves have different history-dependent midpoint parameters and same lapse rates: after a long ITI, the curves undergo a larger horizontal shift in the opposite direction of the previous stimulus. Right: Lines represent the absolute difference between each pair of psychometric curves shown in the left plot. A history-dependent change in midpoint modulated by ITI duration would lead to a larger separation between the two lines towards the center of the stimulus range, and a smaller separation towards the extremes. In other words, the effect of the ITI interacts with the effect of the previous stimulus, amplifying the horizontal shift. c. Rats' data support the model of b, a history-dependent change in midpoint modulated by ITI duration. Similarly, as in Supplementary Figure S5b, the slope of the regression line for each combination of stimulus n and stimulus n-1 (9x9) was computed, this time separately after short and long ITI. A larger separation between the two lines towards the center of the stimulus range, and a smaller separation towards the extremes suggests that ITI durations modulate the history-dependent change in midpoint, corresponding with a largely horizontal shift of the psychometric curves.
Two paired distributions of regression slopes (short ITI vs. long ITI) were obtained by bootstrapping rats' data. For each sample, we computed the difference between the two regression slopes in order to obtain a confidence interval. For each stimulus n, an asterisk indicates that the 90% confidence interval was excluding a null difference. Source data are provided as a Source Data file. 9 Supplementary Figure S7. Stimulus-dependent bias builds up between trials even when removing incorrect trials and considering only early-session trials. ITI is independent of speed of stimulus n-1 a. Psychometric curves on trial n plotted according speed of stimulus n-1 and ITI, only using correct choices in trial n-1 and early session trials (n<150). Dashed line is the average psychometric curve for all trials. b. Slope of the regression line fit between bias (PSE) and previous trial speed for all rats, separated for shortest 2 ITI quartiles (light blue) and longest 2 ITI quartiles (dark blue), as in Figure 5c. Shading represents 95% confidence intervals. All trials with incorrect choices in trial n-1 were removed and only early-session trials were considered (see above). Interaction term between the effects of previous speed and ITI durations on choice n, p=0.03 (two-tailed t-test of the regression coefficient). c. Median ITI from trial n-1 to n, grouped by speed of stimulus n-1. Source data are provided as a Source Data file. Point of subjective equality (PSE) in trial n with data grouped by choice in trial n-1. Data with n-1 speed = 0, excluding incorrect trials in trial n-1. We found a choice-repetition bias that was slightly larger following long ITIs (dark blue) as compared to short ITIs (light blue).
Middle: PSE depending on previous trial speed, following short (light blue) and long (dark  terms between the effect of stimulus n-2 and n-1 with the adjacent ITIs. GLM with the equation: p(judged strong)_n ~ 1 + Stimulus_n + Stimulus_n-1 + Stimulus_n-2 + Stimulus_n-2:ITI_n-1 + Stimulus_n-2:ITI_n-2. See main text for the explanation of the present analysis. Pvalues are computed from two-tailed t-tests of the regression coefficients. Source data are provided as a Source Data file.