The metaperceptual function: Exploring dissociations between confidence and task performance with type 2 psychometric curves

Abstract

<p>Confidence can dissociate from perceptual accuracy, suggesting distinct computational and neural processes underlie these psychological functions. Recent investigations have therefore sought to experimentally isolate metacognitive processes by creating conditions where perceptual sensitivity is matched but confidence differs (“matched-performance / different-confidence”; MPDC). Despite these endeavors’ success, much remains unknown about MPDC effects and how to best harness them in experimental settings. Here we developed a principled approach to comprehensively characterizing MPDC effects through analyzing metaperceptual (i.e., type 2 psychometric) functions relating objective performance to subjective confidence across widely varying performance levels and experimental manipulations. We found that MPDC effect magnitude depends on stimulus properties, observers’ sensitivity level, and critically on trial type order (blocked or interleaved across stimulus property variations). Our findings provide the first comprehensive exploration of MPDC effects, offer a prescriptive guide to metaperceptual analysis, and suggest optimal experimental paradigms for experimentally isolating metacognition and awareness in future studies.</p>

Tuned inhibition in perceptual decision-making circuits can explain seemingly suboptimal confidence behavior

“…We developed the Differential Tuned Inhibition model to test this hypothesis. Our results show that such a network can explain counterintuitive behaviors reported in the literature [4,10,[12][13][14][16][17][18]44,45]. We further demonstrate that the model's special property of weighting less inhibited units more heavily in computing confidence is the key to capturing empirical findings, since control simulations demonstrate that the model fails to reproduce these findings when instead more inhibited units drive confidence.…”

“…Thus, we reasoned that these data could be well modeled by supposing that the High PE condition has both stronger stimulus drive (more coherently moving dots) and stronger noise (more randomly moving dots); the increase in both stimulus strength and noise allows for signal-to-noise ratio (and hence d’) to be similar, even as the same increases in signal and noise lead to higher absolute values of sensory evidence (and hence higher confidence). Indeed, previous work has demonstrated that increases in signal strength and noise can yield performance-matched increases in confidence under the assumption of fixed decision criteria [ 38 ], an assumption that is supported empirically in cases where the experimental conditions modeled with lower and higher variance in sensory evidence are randomly interleaved across trials [ 17 ]. We implemented this idea in the modeling by positing that the parameter controlling standard deviation of instantaneous sensory evidence, σ, would take on a larger value in the High PE condition.…”

“…How the brain calculates subjective decision confidence is still a topic of active debate [1,2,[39][40][41][42][43]. Although dominant models suggest that confidence reflects an optimal readout of the probability that a decision is correct [1][2][3][4][5][6][7][8][9], it appears challenging for such models to account for counterintuitive behaviors in which confidence and accuracy dissociate [10,11,13,14,16,17,44]. An alternative hypothesis suggesting that confidence reflects a heuristic reliance on decision-congruent evidence [10,14,16,18,21] captures many of these behaviors, and is supported by human intracranial electrophysiology [23].…”

“…In light of this discussion, and of the empirically observed tuned inhibition in cortical areas, a biologically plausible model of sensory evidence accumulation ought to implement more than one level of inhibition and consider how such variations in inhibition tuning may affect a neuron's role in the circuitry. Further, such stratification of tuned inhibition could provide a neural mechanism to explain findings that confidence judgments rely on the magnitude of decision-congruent evidence [10,14,17,18,[21][22][23]. Specifically, the output of less inhibited 'detection' neurons could be used to index decision-congruent evidence and therefore be used for confidence rating.…”

“…While this optimal 'probability correct' account of confidence has enjoyed significant empirical support, it seems difficult for it to account for cases where task performance and confidence dissociate [10][11][12][13][14][15][16][17][18]. Seemingly suboptimal behaviors have also been observed in post-decisional perceptual judgments other than confidence [19,20], leading these authors to hypothesize that these suboptimalities may stem from limitations on computational (i.e., neural) resources or a drive towards self-consistent behavior.…”