Metacognitive evaluation of postdecisional perceptual representations

Abstract

Perceptual confidence is thought to arise from metacognitive processes that evaluate the underlying perceptual decision evidence. We investigated whether metacognitive access to perceptual evidence is constrained by the hierarchical organization of visual cortex, where high-level representations tend to be more readily available for explicit scrutiny. We found that the ability of human observers to evaluate their confidence did depend on whether they performed a high-level or low-level task on the same stimuli, but was also affected by manipulations that occurred long after the perceptual decision. Confidence in low-level perceptual decisions degraded with more time between the decision and the response cue, especially when backward masking was present. Confidence in high-level tasks was immune to backward masking and benefitted from additional time. These results can be explained by a model assuming confidence heavily relies on postdecisional internal representations of visual stimuli that degrade over time, where high-level representations are more persistent.

Figure 1.

Methods. (A) Stimuli for Experiments 1 through 4 and general procedure. On each trial, the observer was presented with a stimulus and asked to make one of two judgments. Judgments were either “which eye is darker? Left or right?” (a low-level task) or “were the eyes looking to the left of right of you?” (a high-level task). In Experiment 1, they were prompted to make their response after a white noise mask. A confidence forced choice judgment was made over two consecutive trials. (B) Stimuli for Experiment 5. White lines are added for illustration only and were not part of the stimuli. The stimuli were dots presented with dynamic motion to mimic human walking. In the low-level task (top), the chest and pelvis dots were rotated clockwise or counterclockwise and back to center and the participant was asked to judge the direction, whilst in the high-level task they judged the relative walking direction (to the left or right of directly toward them). (C) Individual (thin lines) and average (thick lines) Psychometric functions describing the proportion of rightward responses across presented stimulus strengths. Gaze deviation had no effect on low-level responses and contrast difference had no effect on high-level responses. (D) Proportion of rightward responses for trials chosen (filled) as more likely to be correct and trials declined (open) at the forced choice confidence response for the low-level (top) and high-level (bottom) tasks. Stimulus strength was normalized based on the psychometric function of each observer before aggregating across observers. The difference between chosen and declined seems to be small here because most trial pairs contained a similar stimulus difficulty. Confidence efficiency can be better appreciated by comparing choices according to the pairs, as in (E). (E) Proportion of trial pairs in which the first decision was chosen as more likely to be correct, by normalized stimulus strength across the two intervals for the low-level (top) and high-level (bottom) tasks. The more clearly cyan (0) and magenta (1) are divided along the diagonal, the more confidence is likely to discriminate correct from incorrect perceptual decisions. (F) Reaction times by stimulus strength in each task from Experiment 2. The top shows the average and 95% within subjects confidence (error bars) median reaction times by stimulus strength ordered as in (C). The consistency across tasks and observers can be better appreciated by the bottom, where stimulus strength is normalized by each individuals psychometric function, and reaction times are normalized by taking the difference from the mean (across both tasks) in units of standard deviation (z-scored) of the logarithm of reaction times. Individuals are shown in small markers, and the average in large markers.

Figure 2.

Results of Experiments 1 and 2. (A) Psychometric functions averaged over participants in Experiments 1 and 2. (B) Comparison of slopes of individual participants, solid markers show the group median, and vertical lines show ±1 standard deviation. (C) Median reaction times in Experiment 2 for individual participants, where solid markers show the group median, and horizonal lines show ±1 standard deviation. (D) Confidence efficiency (top row), confidence noise (middle), and confidence boost (bottom) in Experiment 1 (left column) and 2 (right). Markers show the mean, thin vertical lines show 95% Confidence Intervals, thick show ±1 SD, and horizontal histograms represent the sample density from bootstrap resampling. (E) Within subjects effect size for the difference in confidence efficiency (top), noise (middle), and boost (bottom), comparing the high-level and low-level tasks in Experiment 1 (left symbol) and 2 (right) based on hierarchical Bayesian modelling. Markers show the posterior mean and vertical line shows the 95% highest density interval. (F) Between subjects effect size for the difference in confidence efficiency (top), noise (middle), and boost (bottom), comparing the low- (left symbol) and high-level (right) tasks in Experiments 1 and 2. Markers show the posterior mean and vertical lines show the 95% highest density interval.

Figure 3.

Effects of late visual masking. (A) Confidence efficiency (left columns), confidence noise (middle), and confidence boost (right) in the low- and high-level tasks (left and right plots). Horizontal histograms show the sample distribution estimated by bootstrapping, the means are shown in the markers, and 95% CIs with thin vertical lines and ±1 SD with thick. The results of Experiments 1 and 2 are shown adjacent for comparison. (B) Within-subject effect size of the mask on confidence efficiency. The colored markers show the posterior mean and vertical lines show the 95% highest density interval. The open black markers show the between subjects comparisons from Experiments 1 and 2 for the same conditions. (C) Between subjects effect comparing the high-level and low-level tasks in the mask and no-mask conditions of Experiment 3, vertical lines show the 95% highest density interval. (D) Between-subjects effect sizes comparing confidence efficiency in Experiment 1 (laboratory experiment) and the mask condition of Experiment 3 (online experiment) for the low- and high-level tasks (different participants in Experiment 3) vertical lines show the 95% highest density interval. (E) Average psychometric functions across all Experiments and conditions. (F) Slope of the psychometric functions for each participant in Experiment 3, the large markers show the group medians, and vertical lines show ±1 standard deviation. Colors correspond with (A).

Figure 4.

Interaction between decision time and response cue time. (A) Confidence efficiency (left), confidence noise (middle), and confidence boost (right) in Experiment 4, in the low-level (magenta) and high-level (green) tasks. Thin vertical lines show 95% CI, thick show ±1 SD. (B) Within-subjects effects (short vs long duration to response cue) in the low- and high-level tasks in Experiments 4, vertical lines show the 95% highest density interval, open black markers show the predicted effect size comparing Experiments 1 and 2. (C) Effect size of the difference in confidence efficiency between Experiments 1 and 2 (colored markers), against the prediction based on the additive effects of the mask and duration to response cue from Experiments 3 and 4 (open black markers).

Figure 5.

Perceptual evidence accumulation for decisions and confidence. (A) Simulations of perceptual evidence accumulation and the effect of long and short response cues. The sequential sampling process (Equation 6) is shown to the left, where evidence is accumulated up to a collapsing bound (thick black curves). The median decision time is shown with the green line in these plots. This demonstration is based on the parameters fit to the median reaction time observers from Experiment 4. The red and black boxes correspond with the median postdecision time (≤200 ms after the response, when the next decision is cued) in the long and short duration conditions respectively. The top plot demonstrates continued accumulation with the same signal-to-noise ratio as predecision commitment, the bottom plot shows continued accumulation of noise only (no additional signal evidence after decision commitment). The distributions of final accumulated evidence (for each stimulus strength) are shown to the right, assuming accumulation continues until the next trial in the short and long duration to response cue conditions. These distributions of final accumulated evidence for confidence generate different predictions about confidence efficiency, shown in the bar plot to the right. Continued accumulation of the same signal + noise as prior to decision commitment (filled bars) predicts an increase in confidence efficiency with more time, whereas accumulating noise only predicts a decrease in confidence efficiency (open bars). (B) Demonstration of the relationship between postdecision time and decision time in the short (top) and long (bottom) duration to response cue conditions. The blue bars represent decision time, pink, nondecision time (NDT), and purple, the intertrial interval (ITI). In the long duration condition there is additional waiting time (WT, yellow) where the participant waits for the response cue. Assuming perceptual decision processes proceed in the same manner in both conditions (as there was no substantial difference in perceptual sensitivity), participants who made decisions faster (labelled ‘fast’) would have relatively longer postdecision time (NDT + WT) in the long duration condition compared with those who took longer (labelled “medium”). (C) Confidence efficiency in Experiment 2, splitting participants into three equal groups based on their reaction times. Thin error bars show 95% CI, thick show ±1 SD. (D) Within subjects effects for the difference in high-level and low-level confidence efficiency in (C). Error bars show 95% highest density interval. (E) Model predicted difference in confidence efficiency between the high-level and low-level tasks assuming confidence efficiency is based on continued accumulation with the same signal-to-noise ratio as the prior to decision commitment in the high-level task, and continued accumulation of noise only in the low-level task. The pattern of decreasing difference is similar to that of the data in Experiment 2. (F) Confidence efficiency in the low-level (top) and high-level (bottom) tasks for participants split into three RT groups based on their reaction times in the short duration condition in Experiments 4 (left) and 5 (right). Thin error bars show 95% CI, thick show ±1 SD. (G) Within-subjects effects for the difference between the short and long duration to response cue conditions in Experiments 4 and 5. Error bars show 95% highest density interval.