Self-evaluation of decision-making: A general Bayesian framework for metacognitive computation.

Stephen M. Fleming, Nathaniel D. DawPsychological 2016 · DOI

TL;DR

This paper introduces a novel Bayesian framework for understanding how we evaluate our own decision-making, proposing that self-evaluation is a "second-order" inference process. This unified model explains both confidence in our choices and our ability to detect errors, suggesting that our actions themselves can provide crucial information for this self-assessment.

What Problem Does This Paper Address?

Humans are remarkably good at assessing their own performance, reporting confidence levels that often align with their accuracy. This metacognitive ability is crucial for guiding behavior, especially when external feedback is scarce. However, existing computational models struggle to explain how both confidence and error detection arise from a single system, and often resort to ad-hoc explanations for discrepancies between performance and self-evaluation. The authors aim to bridge this gap by proposing a general Bayesian framework that can account for these phenomena in a unified way.

Methods

The core of the paper is a Bayesian ideal observer model. They propose a "second-order" inference framework where confidence is not directly derived from the same internal state that drives the decision (as in "first-order" models). Instead, confidence is an inference about the performance of a coupled but distinct decision system. This is computationally analogous to inferring another person's performance. They mathematically derive this framework and compare it to simpler "first-order" and "post-decisional" models. Simulations are used to explore the qualitative predictions of these models under various conditions, particularly focusing on how confidence relates to stimulus strength, accuracy, and error detection.

Datasets

Not discussed in the paper. The paper relies on simulations to test its theoretical framework.

Key Insights and Ideas

  • Self-evaluation as "second-order" inference: The central idea is that we don't just make a decision; we also make a judgment about that decision-making process. This is akin to observing another agent.
  • Unified account of confidence and error detection: The second-order framework naturally explains both how we feel confident and how we detect our own mistakes, unifying two previously disparate areas of metacognition research.
  • Actions as diagnostic information: Counterintuitively, the model suggests that our own actions can provide valuable information for self-evaluation, especially when the internal states driving decisions and confidence are not perfectly coupled.
  • Explaining hyper- and hypo-metacognitive sensitivity: The framework offers explanations for why metacognitive accuracy can sometimes be better (hyper-sensitive) or worse (hypo-sensitive) than actual task performance.
  • Nested model structure: The proposed second-order model encompasses simpler first-order and post-decisional models as special cases, providing a more general and flexible account.

Key Results

  • X-pattern reproduction: All three models (first-order, post-decisional, and second-order) can reproduce the "X-pattern" where confidence increases with stimulus strength for correct trials but decreases for error trials. However, the underlying mechanisms differ.
  • Error detection capability: Only the post-decisional and second-order models can account for error detection, where confidence can drop below 0.5, indicating a perceived error. The first-order model, by definition, cannot.
  • Influence of actions on confidence: The second-order model predicts that actions can influence confidence ratings, particularly when confidence and decision variables are weakly coupled. This is supported by empirical data from Siedlecka et al. (2016).
  • Metacognitive sensitivity and bias: The model demonstrates how changes in the covariance parameters (noise and correlation between decision and confidence variables) can lead to variations in metacognitive sensitivity and bias, explaining both hypo- and hyper-sensitivity.

Limitations and Future Work

  • Simplistic task: The current model is primarily derived for a two-choice perceptual discrimination task. Generalizing to other domains like memory or value-based choices is a key future direction.
  • Static model: The current framework is largely static, not fully incorporating temporal dynamics of evidence accumulation or learning processes for confidence and belief hyperparameters.
  • Neural implementation: While the paper discusses potential neural correlates (e.g., PFC, dmPFC), the specific algorithmic and mechanistic implementation remains an open question.
  • Action specification: The model simplifies actions to binary choices. Future work could explore how the kinematics and planning of more complex actions influence self-evaluation.
  • Interactions with others: The role of social interactions in shaping confidence and metacognition is not explored but is suggested as an avenue for future research.

Summary updated

Abstract

People are often aware of their mistakes, and report levels of confidence in their choices that correlate with objective performance. These metacognitive assessments of decision quality are important for the guidance of behavior, particularly when external feedback is absent or sporadic. However, a computational framework that accounts for both confidence and error detection is lacking. In addition, accounts of dissociations between performance and metacognition have often relied on ad hoc assumptions, precluding a unified account of intact and impaired self-evaluation. Here we present a general Bayesian framework in which self-evaluation is cast as a "second-order" inference on a coupled but distinct decision system, computationally equivalent to inferring the performance of another actor. Second-order computation may ensue whenever there is a separation between internal states supporting decisions and confidence estimates over space and/or time. We contrast second-order computation against simpler first-order models in which the same internal state supports both decisions and confidence estimates. Through simulations we show that second-order computation provides a unified account of different types of self-evaluation often considered in separate literatures, such as confidence and error detection, and generates novel predictions about the contribution of one's own actions to metacognitive judgments. In addition, the model provides insight into why subjects' metacognition may sometimes be better or worse than task performance. We suggest that second-order computation may underpin self-evaluative judgments across a range of domains. (PsycINFO Database Record
BibTeX
@inproceedings{fleming2016self,
  author = {Stephen M. Fleming and Nathaniel D. Daw},
  title = {{Self-evaluation of decision-making: A general Bayesian framework for metacognitive computation.}},
  booktitle = {Psychological},
  year = {2016},
  doi = {10.1037/rev0000045},
}

Loading related content...

Figure