Confidence Leak in Perceptual Decision Making

Abstract

People live in a continuous environment in which the visual scene changes on a slow timescale. It has been shown that to exploit such environmental stability, the brain creates a continuity field in which objects seen seconds ago influence the perception of current objects. What is unknown is whether a similar mechanism exists at the level of metacognitive representations. In three experiments, we demonstrated a robust intertask confidence leak-that is, confidence in one's response on a given task or trial influencing confidence on the following task or trial. This confidence leak could not be explained by response priming or attentional fluctuations. Better ability to modulate confidence leak predicted higher capacity for metacognition as well as greater gray matter volume in the prefrontal cortex. A model based on normative principles from Bayesian inference explained the results by postulating that observers subjectively estimate the perceptual signal strength in a stable environment. These results point to the existence of a novel metacognitive mechanism mediated by regions in the prefrontal cortex.

Keywords: attention; decision making; open data; open materials; perception.

Figure 1

Tasks for Experiments 1 and 2. The stimulus consisted of 40 X’s and O’s that were colored in red and blue. The observers’ tasks were to decide whether there were more X’s or more O’s (letter identity task), as well as whether there were more red or blue letters (color task). For each task, observers indicated their level of confidence. In Experiment 1 confidence was indicated using a 1–4 scale. In Experiment 2, for one of the tasks observers used a visual analog scale (VAS) by sliding a marker, while for the other question they decided whether to provide an answer in order to win a larger reward (thereby indicating high level of confidence), or to “opt out” and not give a response thus earning a smaller, guaranteed, reward (thereby indicating low level of confidence). To further minimize response priming, in Experiment 2 the VAS response was provided with a mouse, while the opt out response was provided with a keyboard.

Figure 2

Results of Experiment 1. A) Individual fits in a regression in which only the confidence on the letter identity task was used to predict the confidence on the color task. B) Mean confidence on the color task increases as a function of the confidence on the letter identity task. C) Mean accuracy on the color task does not vary as a function of the accuracy on the letter identity task. Open circles signify model fits.

Figure 3

Results of Experiment 2. A) Individual observers’ confidence on one task, measured with the visual analogue scale (VAS), as a function of the confidence on the other task, measured with opt-out procedure. Positive relationship was found for 17 of the 18 observers. B) Probability density functions for the VAS confidence on one task for trials in which the observers decided to opt out (i.e., low confidence) vs. trials in which observers provided a response (i.e., high confidence) on the other task. For display purposes the curves were smoothed with a 10-point moving average. The thick lines indicate data, while the thin lines indicate model fits.

Figure 4

A graphical depiction of our model. A) Discrimination between stimuli S₁ and S₂. Each stimulus follows a Gaussian likelihood function (thick lines) on an axis that denotes the total evidence available on a given trial. The posterior distributions (thin lines) are drawn for the case when the two stimuli have equal prior probability. Bayes-optimal thresholds are placed horizontally (black lines) based on predetermined cutoff values (in this case, 0.5 for the decision, and 0.75 for the confidence). These cutoffs correspond to vertical criteria defined in the likelihood space (green lines), which intersect the horizontal thresholds on the posterior distributions. Observers’ goal is to set stable cutoffs in the posterior probability space, as depicted in the inset. B) Solid lines represent the distributions from panel A, while dashed lines represent the expected distributions in a higher signal-to-noise environment. To remain consistent with the threshold placed at a cutoff of 0.75 on the posterior distributions, the confidence criteria (defined on the likelihood space) move “inward” (see the new criterion vs. baseline criterion in the figure). However, if the expectation for a high signal-to-noise environment is false, then the observer is using the dashed criteria (based on the observer’s expectations) to judge stimuli characterized by the solid distributions. This means that unbeknown to the observer, she is using a lower effective cutoff on the true posterior distribution (in the Figure above, the effective cutoff is ~0.6), naturally resulting in a higher proportion of high confidence responses.

Figure 5

Experiment 3. A) Observers were asked to decide on the orientation (clockwise vs. counter-clockwise) of briefly flashed Gabor patches. In different blocks either 2 (high attention condition) or 4 (low attention condition) patches were flashed, and observers indicated the orientation of a post-cued Gabor patch. B) Capacity d’ was higher in the high-attention, 2-stimulus condition. C) Individual data showing that the average confidence on the current trial is positively correlated with the confidence on the previous trial for all 20 observers. D) Confidence autocorrelation was significantly positive (both p’s < .001) for both the high and low attention conditions but was not significantly different between the two conditions (p = .8).

Figure 6

The inter-task confidence leak correlates negatively with metacognition, measured as the area under the Type 2 ROC curve (Type 2 AUC), suggesting that the process of confidence leak affects negatively observers’ metacognitive performance.

Figure 7

Higher confidence leak scores are predicted by lower gray matter volume in right prefrontal cortex (PFC). The image shows a T map thresholded at p = 0.005 uncorrected for display purposes, though analyses were performed using small-volume correction for multiple comparisons. The regions in right DLPFC and aPFC are shown in blue and green, respectively.