Arousal-related adjustments of perceptual biases optimize perception in dynamic environments

Abstract

Prior expectations can be used to improve perceptual judgments about ambiguous stimuli. However, little is known about if and how these improvements are maintained in dynamic environments in which the quality of appropriate priors changes from one stimulus to the next. Using a sound-localization task, we show that changes in stimulus predictability lead to arousal-mediated adjustments in the magnitude of prior-driven biases that optimize perceptual judgments about each stimulus. These adjustments depend on task-dependent changes in the relevance and reliability of prior expectations, which subjects update using both normative and idiosyncratic principles. The resulting variations in biases across task conditions and individuals are reflected in modulations of pupil diameter, such that larger stimulus-evoked pupil responses correspond to smaller biases. These results suggest a critical role for the arousal system in adjusting the strength of perceptual biases with respect to inferred environmental dynamics to optimize perceptual judgements.

Figure 1. Dynamic sound-localization task

(a) Subjects listened via headphones to noise bursts with virtual source locations that varied along the frontal, azimuthal plane. The locations were sampled (points) from a Gaussian distribution (gray) with a mean that changed abruptly on unsignaled change-points (probability=0.15 for each sound) and a STD of 10° in low-noise blocks, 20° in high-noise blocks. The subjects listened passively to the sound sequence, except for occasional probe trials. All sounds except the probe sound were presented simultaneously with their corresponding locations on a semicircular arc shown on the isoluminant visual display, allowing subjects to develop priors on sound-source location based on both the auditory and visual signals and maintain a stable mapping between the two. (b) An example trial sequence showing the mean (solid line) and sampled (points) locations over 50 trials. Vertical dashed lines indicate randomly selected probe trials. (c) Probe-trial sequence. Using a mouse to control a cursor on the visual display, the subject reported: 1) the predicted location of the upcoming probe sound, followed by 250-ms fixation, presentation of the probe sound, then continued fixation for 2.5 s to allow for pupil measurements; 2) the estimated location of the probe sound; and 3) a high or low confidence report that the true location was within a small window centered on their estimate. The sound sequence then continued until the next probe. (d–f) Schematic illustrating the changing reliability and relevance of priors for the probe sounds in a and b, as indicated. Given a fixed-width likelihood function, more reliable and relevant priors have a stronger and more beneficial influence on the percept, here represented as the posterior, which is most uncertain (widest) in e and least uncertain in f.

Figure 2. Overall prediction and estimation performance

(a–c) Reported versus true (simulated) sound-source angle for an example subject for: (a) estimations from the control task; (b) predictions from the dynamic task (light gray points indicate change-point trials, on which the probe location was, by design, unpredictable); and (c) estimations from the dynamic task, including all trials. (d–f) Population summaries, plotted as in (a–c), with per-subject median values shown in black and the median of medians shown in red (n=29 subjects). For the dynamic tasks, median values were calculated in sliding 20° windows. Non-change-point trials were excluded from the predictions in (e). Note that the subjects’ perceptual reports (d and f) were biased slightly towards straight ahead at the far periphery. This bias, which likely reflected learned expectations that sounds were only played in the frontal plane, is accounted for in later analyses (β₅ and β₆ in Eq. 5). (g–i) STD of the perceptual errors from the dynamic task plotted versus the STD of: (g) the perceptual errors from the control task; (h) the prediction errors from the dynamic task; or (i) the expected STD of the perceptual errors, computed from the optimal, reliability-weighted combination of the control perceptual errors and the dynamic prediction errors. Points in g–i represent data from individual subjects. Prediction and perceptual errors were computed with respect to the simulated location of the probe sound.

Figure 3. Effects of task dynamics on performance

(a) STD of the subjects’ prediction errors (filled circles) as a function of the number of sounds after a change-point (SAC) in the generative mean azimuthal location, plotted separately for the two noise conditions (colors, as indicated; generative STDs are shown as dashed lines). For comparison, prediction-error STDs are shown for an approximately optimal predictive-inference model (open diamonds). Data from change-point trials (SAC=1) are not shown because locations were, by design, unpredictable on those trials. (b) Contrast values from a linear model describing individual subject (circles) and the approximately optimal model (each diamond represents analyses based on the same sound sequence experienced by the subject connected by a line) prediction-error STD in terms of (see inset in e): 1) the difference between change-point and non-change-point trials (CP), 2,3) the linear trend from SAC 2–6 for low-(Exp_low) or high-(Exp_high) noise trials, and 4) the difference between the two noise conditions (Noise). (c,d) Same conventions as in a,b but for perceptual errors on the dynamic task. Diamonds represent the theoretically predicted STD of perceptual errors computed from the optimal, precision-weighted combination of the subject- and condition-specific STDs of prior errors (circles in a, determined separately for each subject) and the subject-specific estimation-error STDs from the control task (the median value is shown as a horizontal dashed line; see Fig. 2g). (e,f) Same conventions as in a,b but for the frequency of high-confidence reports relative to overall frequency of high-confidence reports per subject. Diamonds represent the frequency of high-confidence reports corresponding to the theoretical perceptual errors in c, computed from the fraction of the theoretical posterior distribution within the confidence window. In a,c,e, circles and error bars are mean±sem of values measured from all 29 subjects. In b,d,f, points are data from individual subjects. Asterisks indicate sign-rank test for H₀: median value from the subject data=0, p<0.05. In each case, paired rank-sum test for H₀: median difference between subject data and theoretical prediction, p>0.087. In all panels, only data from sequences following noticeable change-points (changes in mean of at least twice the generative STD for SAC=1) were included.

Figure 4. Effects of task dynamics on perceptual bias

(a–c) Example data from a single subject illustrating the quantification of perceptual bias as the slope of the best-fit line to a scatter of the perceptual error versus the prediction error. Slopes close to zero reflect a low perceptual bias (i.e., the percept is unrelated to the prediction), as on change-point trials (b). Slopes closer to unity reflect a higher perceptual bias (i.e., the percept more closely matches the prediction), as on non-change-point trials (c). (d) Perceptual bias as a function of the number sounds after a change-point (SAC) in the generative mean azimuthal location, plotted separately for the two noise conditions (colors, as indicated). Circles and error bars are mean±sem of values measured from all 29 subjects. Diamonds indicate the theoretically predicted perceptual bias from an optimal, reliability-weighted combination of the subject- and condition-specific predictions (Fig. 3a) and the subject-specific estimates from the control task (Fig. 2g). (e) Contrast values from a linear model describing individual subject (circles) and model (each diamond represents analyses based on the same sound sequence experienced by the subject connected by a line) perceptual bias in terms of (see inset in Fig. 3e): 1) the difference between change-point and non-change-point trials (CP), 2,3) the linear trend from SAC 2–6 for low-(Exp_low) or high-(Exp_high) noise trials, and 4) the difference between the two noise conditions (Noise). Asterisks indicate sign-rank test for H₀: median value from the subject data=0, p<0.05. Paired rank-sum tests for H₀: median difference between subject data and theoretical prediction, p<0.01 for CP, p=0.16 for Exp_low, p=0.78 for Exp_high, and p<0.01 for Noise. In d and e, only data from sequences following noticeable change-points (changes in mean of at least twice the generative STD for SAC=1) were included.

Figure 5. Individual differences in perceptual bias

(a, b) Relationship between overall (mean) perceptual bias and either overall localization ability (STD of perceptual errors on the control task, a) or overall prediction ability (STD of prediction errors from non-change-point trials on the dynamic task, b), after accounting for the other factor (hence “residual”) via linear regression. (c–f) The dependence of perceptual bias on various task conditions, plotted as functions of the dependence of prediction-error STD on the same conditions: c, d) the linear trend from SAC 2–6 in the low-noise (c) and high-noise (d) condition (Exp); e) change-point versus non-change-point trials (CP); and f) high-versus low-noise condition (Noise). In each panel, points represent data from individual subjects. Lines are linear regressions. Only data from sequences following noticeable change-points (changes in mean of at least twice the generative STD for SAC=1) were included.

Figure 6. Dynamic modulation of perceptual bias by normative and non-normative factors

(a) Comparison of a parameter-free normative model (ribbons indicate mean±SEM simulated perceptual bias for the same task sequences experienced by the subjects) and the subjects’ behavior (points and errorbars are mean±SEM from 29 subjects), shown as a function of sounds after a change-point (SAC) for the two noise conditions (colors, as indicated). (b) Comparison of the linear model shown in panel e to behavior. Conventions as in panel a. (c,d) Dependence of the normative factors used in both models on task conditions: (c) prior relevance, which measures the probability of the current sound coming from the same distribution as the previous sound; and (d) prior reliability, which measures the anticipated precision of the predictive distribution relative to the likelihood distribution prior to stimulus presentation. (e) Best-fitting parameter estimates from the linear model fit to behavioral data from each subject (points) and to simulations of the parameter-free normative model (thick and thin bars indicate 95% confidence intervals over simulated subjective values and over simulated mean values across subjects, respectively). PE=prediction error. Asterisks indicate coefficients with mean values that differed from zero (t-test, p< 0.05).

Figure 7. Pupil diameter reflects dynamic modulations of perceptual bias within individual subjects

(a) Mean±sem evoked pupil response from 29 subjects, defined as the pupil diameter relative to baseline during the measurement period. Red line indicates the time of the peak mean response (1.38 sec after stimulus presentation). (b–d) Baseline pupil diameter for trials sorted into bins according to relevance (b), reliability (c), and confidence (d). Relevance and reliability were binned in quintiles per subject, then each bin was combined across subjects. Confidence was divided into all trials with a low (0) or high (1) confidence report. Points and errorbars are mean±SEM from all values in each bin. (e–g) Same as b–d, but using the pupil diameter measured at the time of the peak response after accounting for the linear baseline dependencies. (h,i) Regression coefficients from a linear model accounting for modulation of baseline pupil diameter (h) or the evoked response (i) at each time-point using as predictors: 1) prior relevance, 2) prior reliability, 3) the upcoming confidence report, and 4) the residual perceptual bias from the linear model in Fig. 6d. Points and error bars in h and lines and ribbons in i represent mean±sem of values computed per subject and thus represent within-subject modulations. Points and lines/ribbons corresponding to relevance, reliability, and confidence use the same colors as in (b–g). Bold symbols in h and horizontal lines in i indicate periods for which H₀: value=0, p<0.05, after accounting for multiple comparisons.

Figure 8. Pupil diameter reflects individual differences in perceptual biases

(a,b) Mean baseline diameter for each subject (points) as a function of the perceptual bias (a; fits to the PE term in Fig. 6e) and relevance-dependent bias (b; fits to the PE*relevance term in Fig. 6e). (c,d) Mean evoked pupil response for each subject as a function of the perceptual bias (a) and relevance-dependent bias (b). Pupil responses were measured at the time of peak response (1.38 sec after stimulus presentation) and orthogonalized to subject baseline pupil measurements. (e,f) Regression coefficients describing the relationship between shared or unique variance (colors, as indicated) in PE and PE*relevance coefficients from the behavioral model and average baseline (e) or stimulus evoked (f) pupil diameter. Points and error bars in d and lines and ribbons in e represent the correlation coefficient and 95% confidence intervals of the estimate and thus represent across-subject modulations. Horizontal lines in e indicate periods for which H₀: value=0, p<0.05 after accounting for multiple comparisons.