Metacognitive bias resulting from trade-off between local and global motion signals

Abstract

Visual confidence generally depends on performance in targeted perceptual tasks. However, it remains unclear how factors unrelated to performance affect confidence. Given the hierarchical nature of visual processing, both local and global stimulus features can influence confidence, but their strengths of influence remain unknown. To address this question, we independently manipulated the local contrast signals and the global coherence signals in a multiple-aperture motion pattern. The drifting-Gabor elements were individually manipulated to give rise to a coherent global motion percept. In both dichotomous direction-discrimination task (Experiment 1) and analog direction-judgment task (Experiment 2), we found stimulus-dependent biases in metacognition despite matched perceptual performance. Specifically, participants systematically gave higher confidence ratings to an incoherent pattern with clear elements (i.e., strong local but weak global signals) than a coherent pattern with noisy elements (i.e., weak local but strong global signals). We did not find any systematic effects of local/global stimulus features on metacognitive sensitivity. Model comparisons show that variation in local/global signals in the stimulus should be considered a factor influencing confidence, even after controlling for the effects of performance. Our results suggest that the metacognitive system, when generating confidence for a perceptual task, puts more weights on local than global signals.

Figure 1.

The LSNR-GSNR stimulus space. We independently manipulated the local contrast and global motion coherence to create stimuli that differed in the LSNR-GSNR space (central panel). Low LSNR corresponds to noisy elements (A and C), whereas high LSNR corresponds to clear grating elements (B and D). Low GSNR corresponds to a globally coherent motion pattern (A and B), whereas high GSNR corresponds to an incoherent motion pattern (C and D). Note that the global motion percept was rotational in Experiment 1, which is different from the rightward translational motion percept illustrated by the signal arrows (green) in this figure. Demo movies for example stimuli (including those for Experiment 2) can be found here: https://osf.io/mjtdh/.

Figure 2.

Schematic illustration of the dynamic range of local/global signals in the two-dimensional stimulus space. Colored lines represent the iso-performance lines for different levels of perceptual performance. Here, 75% (green line) is chosen to illustrate the dynamic range. The dotted lines mark the lower and upper bounds of the LSNR and the GSNR. These bounds define the dynamic range in this 2D stimulus space, within which local and global signals compensate for each other to maintain constant performance. The calibration sessions aimed at identifying the dynamic range for 75% accuracy in this 2D space by estimating the lower and upper bounds for LSNR and GSNR. Note that the iso-performance lines could be nonlinear in the actual data (depicted as linear here only for illustrations).

Figure 3.

Stimuli identified by the adaptive staircase procedure for each participant in Experiment 1, varying in the two-dimensional space of LSNR and GSNR. The middle-left panel shows how GSNR (averaged over trials) varied as LSNR varied, across participants. On average, when the local elements were noisy (i.e., low in LSNR; blue and light blue dots), the global motion pattern became coherent (i.e., high in GSNR) to maintain perceptual performance. Vice versa when local elements were clear (i.e., high in LSNR; red and light red dots). To summarize this relationship, we computed log(GSNR/LSNR) as a one-dimensional variable to describe the variation in this stimulus space. The bottom-left and the bottom-middle panels show how this one-dimensional variable changes with LSNR and GSNR. respectively. The diagonal panels (from top-left to bottom-right) show the distributions of LSNR, GSNR, and log(GSNR/LSNR), respectively.

Figure 4.

Similar perceptual performance (dʹ) across the four conditions. Each light thin line represents the data of one participant. Each bar represents the mean of each LSNR level. Error bars represent the 95% confidence intervals.

Figure 5.

Confidence ratings increased as Local SNR increased. Each light thin line represents the data of one participant. Each bar represents the mean Z value of each LSNR level. Error bars represent the 95% confidence intervals.

Figure 6.

Differences in confidence ratings between correct and incorrect responses across levels of local SNR. Each light thin line represents the data of one participant. Each bar represents the mean across participants. Error bar denotes 95% confidence intervals.

Figure 7.

The relationship between stimulus signal structure (as log(GSNR/LSNR)), perceptual performance (as dʹ), and confidence (as z(p(high-conf))).

Figure 8.

Procedure of Experiment 2. Participants first saw the multiple-aperture motion pattern with a translational global motion direction. Then they indicated the perceived global motion direction by turning a simulated dial (cyan pointer) using the computer mouse. Before confirming this direction-judgment response, participants adjusted the width of the bet span (thick white bar; symmetrical around the cyan pointer) so that the bet span would cover the global motion direction of the stimulus. The white number at the center (“75”) represents the wagering points and was updated in real time as participants adjusted the width of the bet span (wider bet span, lower wagering points). After the participants had submitted their response, trial-by-trial feedback would then be given, depending on whether the response was correct (top-right panel, with white bar covering the true stimulus direction indicated in green; point increment for this trial in green; total points so far after this increment in black) or incorrect (bottom-right panel, with white bar not covering the true stimulus direction indicated in red; point deduction for this trial in red; total points so far after deduction in black). For demo movies of the same example stimuli, please visit: https://osf.io/mjtdh/.

Figure 9.

Perceptual performance for each stimulus condition measured as precision in the direction-judgment task. (a): Illustration of the uniform-normal mixture fit (blue curve) to the distribution of the direction-judgment errors (20 bins of 36° each) for two participants in the three local-SNR level conditions. (b) Perceptual performance (measured inversely using perceptual precision) was worst when LSNR was the lowest (level 1). Each light thin line represents the data of one participant. The dark thick line represents the mean. Error bars represent the 95% confidence intervals.

Figure 10.

Lower mean confidence response for 0.15 LSNR, but response reached a plateau for 0.30 and 0.60 LSNRs. Each light thin line represents the data for one participant. The dark thick line represents the mean. Error bars represent the 95% confidence intervals.

Figure 11.

Lower SNR yields less confidence regardless of the precision of the response. Mean bet span for 0.15 SNR is significantly higher than 0.30 and 0.60 SNR; the bet span is interpreted as inversely related to confidence.

Figure 12.

The relationship between local/global signal structure in the stimulus (as log(GSNR/LSNR) ), perceptual performance (as precision estimated by sigma), and confidence (as the width of bet span).