Priors and payoffs in confidence judgments

How do humans give confidence? A comprehensive comparison of process models of perceptual metacognition

Humans have the metacognitive ability to assess the accuracy of their decisions via confidence judgments. Several computational models of confidence have been developed but not enough has been done to compare these models, making it difficult to adjudicate between them. Here, we compare 14 popular models of confidence that make various assumptions, such as confidence being derived from postdecisional evidence, from positive (decision-congruent) evidence, from posterior probability computations, or from a separate decision-making system for metacognitive judgments. We fit all models to three large experiments in which subjects completed a basic perceptual task with confidence ratings. In Experiments 1 and 2, the best-fitting model was the lognormal meta noise (LogN) model, which postulates that confidence is selectively corrupted by signal-dependent noise. However, in Experiment 3, the positive evidence (PE) model provided the best fits. We evaluated a new model combining the two consistently best-performing models-LogN and the weighted evidence and visibility (WEV). The resulting model, which we call logWEV, outperformed its individual counterparts and the PE model across all data sets, offering a better, more generalizable explanation for these data. Parameter and model recovery analyses showed mostly good recoverability but with important exceptions carrying implications for our ability to discriminate between models. Finally, we evaluated each model's ability to explain different patterns in the data, which led to additional insight into their performances. These results comprehensively characterize the relative adequacy of current confidence models to fit data from basic perceptual tasks and highlight the most plausible mechanisms underlying confidence generation. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Neural Representations of Post-Decision Accuracy and Reward Expectation in the Caudate Nucleus and Frontal Eye Field

Performance monitoring that supports ongoing behavioral adjustments is often examined in the context of either choice confidence for perceptual decisions (i.e., "did I get it right?") or reward expectation for reward-based decisions (i.e., "what reward will I receive?"). However, our understanding of how the brain encodes these distinct evaluative signals remains limited because they are easily conflated, particularly in commonly used two-alternative tasks with symmetric rewards for correct choices. Previously we used a motion-discrimination task with asymmetric rewards to identify neural substrates of forming reward-biased perceptual decisions in the caudate nucleus (part of the striatum in the basal ganglia) and the frontal eye field (FEF, in prefrontal cortex). Here we leveraged this task design to partially decouple estimates of accuracy and reward expectation and examine their impacts on subsequent decisions and their representations in those two brain areas. We identified distinguishable representations of these two evaluative signals in individual caudate and FEF neurons, with regional differences in their distribution patterns and time courses. We observed that well-trained monkeys (both sexes) used both evaluative signals, infrequently but consistently, to adjust their subsequent decisions. We found further that these behavioral adjustments had reliable relationships with the neural representations of both evaluative signals in caudate, but not FEF. These results suggest that the cortico-striatal decision network may use diverse evaluative signals to monitor and adjust decision-making behaviors, adding to our understanding of the different roles that the FEF and caudate nucleus play in a diversity of decision-related computations.

Introspective inference counteracts perceptual distortion

Introspective agents can recognize the extent to which their internal perceptual experiences deviate from the actual states of the external world. This ability, also known as insight, is critically required for reality testing and is impaired in psychosis, yet little is known about its cognitive underpinnings. We develop a Bayesian modeling framework and a psychophysics paradigm to quantitatively characterize this type of insight while people experience a motion after-effect illusion. People can incorporate knowledge about the illusion into their decisions when judging the actual direction of a motion stimulus, compensating for the illusion (and often overcompensating). Furthermore, confidence, reaction-time, and pupil-dilation data all show signatures consistent with inferential adjustments in the Bayesian insight model. Our results suggest that people can question the veracity of what they see by making insightful inferences that incorporate introspective knowledge about internal distortions.

Toward 'Computational-Rationality' Approaches to Arbitrating Models of Cognition: A Case Study Using Perceptual Metacognition

Perceptual confidence results from a metacognitive process which evaluates how likely our percepts are to be correct. Many competing models of perceptual metacognition enjoy strong empirical support. Arbitrating these models traditionally proceeds via researchers conducting experiments and then fitting several models to the data collected. However, such a process often includes conditions or paradigms that may not best arbitrate competing models: Many models make similar predictions under typical experimental conditions. Consequently, many experiments are needed, collectively (sub-optimally) sampling the space of conditions to compare models. Here, instead, we introduce a variant of optimal experimental design which we call a computational-rationality approach to generative models of cognition, using perceptual metacognition as a case study. Instead of designing experiments and post-hoc specifying models, we began with comprehensive model comparison among four competing generative models for perceptual metacognition, drawn from literature. By simulating a simple experiment under each model, we identified conditions where these models made maximally diverging predictions for confidence. We then presented these conditions to human observers, and compared the models' capacity to predict choices and confidence. Results revealed two surprising findings: (1) two models previously reported to differently predict confidence to different degrees, with one predicting better than the other, appeared to predict confidence in a direction opposite to previous findings; and (2) two other models previously reported to equivalently predict confidence showed stark differences in the conditions tested here. Although preliminary with regards to which model is actually 'correct' for perceptual metacognition, our findings reveal the promise of this computational-rationality approach to maximizing experimental utility in model arbitration while minimizing the number of experiments necessary to reveal the winning model, both for perceptual metacognition and in other domains.

Prior information differentially affects discrimination decisions and subjective confidence reports

According to Bayesian models, both decisions and confidence are based on the same precision-weighted integration of prior expectations ("priors") and incoming information ("likelihoods"). This assumes that priors are integrated optimally and equally in decisions and confidence, which has not been tested. In three experiments, we quantify how priors inform decisions and confidence. With a dual-decision task we create pairs of conditions that are matched in posterior information, but differ on whether the prior or likelihood is more informative. We find that priors are underweighted in discrimination decisions, but are less underweighted in confidence about those decisions, and this is not due to differences in processing time. The same patterns remain with exogenous probabilistic cues as priors. With a Bayesian model we quantify the weighting parameters for the prior at both levels, and find converging evidence that priors are more optimally used in explicit confidence, even when underused in decisions.

Feeling lucky? prospective and retrospective cues for sensorimotor confidence

On a daily basis, humans interact with the outside world using judgments of sensorimotor confidence, constantly evaluating our actions for success. We ask, what sensory and motor-execution cues are used in making these judgements and when are they available? Two sources of temporally distinct information are prospective cues, available prior to the action (e.g., knowledge of motor noise and past performance), and retrospective cues specific to the action itself (e.g., proprioceptive measurements). We investigated the use of these two cues in two tasks, a secondary motor-awareness task and a main task in which participants reached toward a visual target with an unseen hand and then made a continuous judgment of confidence about the success of the reach. Confidence was reported by setting the size of a circle centered on the reach-target location, where a larger circle reflects lower confidence. Points were awarded if the confidence circle enclosed the true endpoint, with fewer points returned for larger circles. This incentivized accurate reaches and attentive reporting to maximize the score. We compared three Bayesian-inference models of sensorimotor confidence based on either prospective cues, retrospective cues, or both sources of information to maximize expected gain (i.e., an ideal-performance model). Our findings showed two distinct strategies: participants either performed as ideal observers, using both prospective and retrospective cues to make the confidence judgment, or relied solely on prospective information, ignoring retrospective cues. Thus, participants can make use of retrospective cues, evidenced by the behavior observed in our motor-awareness task, but these cues are not always included in the computation of sensorimotor confidence.

Confidence reflects a noisy decision reliability estimate

Decisions vary in difficulty. Humans know this and typically report more confidence in easy than in difficult decisions. However, confidence reports do not perfectly track decision accuracy, but also reflect response biases and difficulty misjudgements. To isolate the quality of confidence reports, we developed a model of the decision-making process underlying choice-confidence data. In this model, confidence reflects a subject's estimate of the reliability of their decision. The quality of this estimate is limited by the subject's uncertainty about the uncertainty of the variable that informs their decision ('meta-uncertainty'). This model provides an accurate account of choice-confidence data across a broad range of perceptual and cognitive tasks, investigated in six previous studies. We find meta-uncertainty varies across subjects, is stable over time, generalizes across some domains and can be manipulated experimentally. The model offers a parsimonious explanation for the computational processes that underlie and constrain the sense of confidence.

Suprathreshold perceptual decisions constrain models of confidence

Perceptual confidence is an important internal signal about the certainty of our decisions and there is a substantial debate on how it is computed. We highlight three confidence metric types from the literature: observers either use 1) the full probability distribution to compute probability correct (Probability metrics), 2) point estimates from the perceptual decision process to estimate uncertainty (Evidence-Strength metrics), or 3) heuristic confidence from stimulus-based cues to uncertainty (Heuristic metrics). These metrics are rarely tested against one another, so we examined models of all three types on a suprathreshold spatial discrimination task. Observers were shown a cloud of dots sampled from a dot generating distribution and judged if the mean of the distribution was left or right of centre. In addition to varying the horizontal position of the mean, there were two sensory uncertainty manipulations: the number of dots sampled and the spread of the generating distribution. After every two perceptual decisions, observers made a confidence forced-choice judgement whether they were more confident in the first or second decision. Model results showed that the majority of observers were best-fit by either: 1) the Heuristic model, which used dot cloud position, spread, and number of dots as cues; or 2) an Evidence-Strength model, which computed the distance between the sensory measurement and discrimination criterion, scaled according to sensory uncertainty. An accidental repetition of some sessions also allowed for the measurement of confidence agreement for identical pairs of stimuli. This N-pass analysis revealed that human observers were more consistent than their best-fitting model would predict, indicating there are still aspects of confidence that are not captured by our modelling. As such, we propose confidence agreement as a useful technique for computational studies of confidence. Taken together, these findings highlight the idiosyncratic nature of confidence computations for complex decision contexts and the need to consider different potential metrics and transformations in the confidence computation.

The feeling of confidence in what we perceive can influence our future behaviour and learning. Understanding how the brain computes confidence is an important goal of researchers. As such, researchers have identified a host of potential models. Yet, rarely are a wide range of models tested against each other to find those that best predict choice behaviour. Our study had human participants compare their confidence for pairs of easy perceptual decisions, reporting if they had higher confidence in the first or second decision. We tested twelve models, covering all three types of models proposed in previous studies, finding strong support for two models. The winning Heuristic model combines all three factors affecting choice uncertainty with an idiosyncratic weighting to compute confidence. The other winning model uses a transformation where the strength of the sensory signal is scaled according to sensory uncertainty. We also assessed the agreement of confidence reports in identical decision scenarios. Humans had higher agreement than almost all model predictions. We propose using confidence agreement intentionally as a second performance benchmark of model fit.

Perceptual confidence judgments reflect self-consistency

Each perceptual decision is commonly attached to a judgment of confidence in the uncertainty of that decision. Confidence is classically defined as the estimate of the posterior probability of the decision to be correct, given the evidence. Here we argue that correctness is neither a valid normative statement of what observers should be doing after their perceptual decision nor a proper descriptive statement of what they actually do. Instead, we propose that perceivers aim at being self-consistent with themselves. We present behavioral evidence obtained in two separate psychophysical experiments that human observers achieve that aim. In one experiment adaptation led to aftereffects, and in the other prior stimulus occurrences were manipulated. We show that confidence judgments perfectly follow changes in perceptual reports and response times, regardless of the nature of the bias. Although observers are able to judge the validity of their percepts, they are oblivious to how biased these percepts are. Focusing on self-consistency rather than correctness leads us to interpret confidence as an estimate of the reliability of one's perceptual decision rather than a distance to an unattainable truth.

Metacognitive resources for adaptive learning⋆

Biological organisms display remarkably flexible behaviours. This is an area of active investigation, in particular in the fields of artificial intelligence, computational and cognitive neuroscience. While inductive biases and broader cognitive functions are undoubtedly important, the ability to monitor and evaluate one's performance or oneself -- metacognition -- strikes as a powerful resource for efficient learning. Often measured as decision confidence in neuroscience and psychology experiments, metacognition appears to reflect a broad range of abstraction levels and downstream behavioural effects. Within this context, the formal investigation of how metacognition interacts with learning processes is a recent endeavour. Of special interest are the neural and computational underpinnings of confidence and reinforcement learning modules. This review discusses a general hierarchy of confidence functions and their neuro-computational relevance for adaptive behaviours. It then introduces novel ways to study the formation and use of meta-representations and nonconscious mental representations related to learning and confidence, and concludes with a discussion on outstanding questions and wider perspectives.

Affective Bias Through the Lens of Signal Detection Theory

Affective bias – a propensity to focus on negative information at the expense of positive information – is a core feature of many mental health problems. However, it can be caused by wide range of possible underlying cognitive mechanisms. Here we illustrate this by focusing on one particular behavioural signature of affective bias – increased tendency of anxious/depressed individuals to predict lower rewards – in the context of the Signal Detection Theory (SDT) modelling framework. Specifically, we show how to apply this framework to measure affective bias and compare it to the behaviour of an optimal observer. We also show how to extend the framework to make predictions about bias when the individual holds incorrect assumptions about the decision context. Building on this theoretical foundation, we propose five experiments to test five hypothetical sources of this affective bias: beliefs about prior probabilities, beliefs about performance, subjective value of reward, learning differences, and need for accuracy differences. We argue that greater precision about the mechanisms driving affective bias may eventually enable us to better understand the mechanisms underlying mood and anxiety disorders.