Name: Medha Shekhar

Dissertation Proposal Meeting

Date: Monday, June 15th, 2020

Time: 11 AM

Location: https://bluejeans.com/750238989 (Meeting ID: 750238989)

Advisor: Dobromir Rahnev, Ph.D. (Georgia Tech)

Dissertation Committee Members:

Paul Verhaeghen, Ph.D. (Georgia Tech)

Christopher Rozell, Ph.D. (Georgia Tech)

Rani Moran, Ph.D. (UCL)

Rick Thomas, Ph.D. (Georgia Tech)

Title: Arbitrating between different models of metacognition 

Abstract: 

Metacognition refers to an observer’s ability to discriminate between their own correct and incorrect decisions using confidence ratings. Several theories have attempted to describe the metacognitive computations that generate confidence by instantiating process models. Yet, there is no consensus on the actual computations that give rise to metacognition. In this study, we aim to shed light on these computations by comparing various established theories of confidence. Traditionally, theories have assumed that confidence is derived entirely from perceptual evidence (signal detection theory; SDT). However, recent theories have modified traditional accounts to allow flexibility and independence between perceptual and confidence judgements. For instance, some models suggest that confidence accrues additional noise and may inherit a decaying perceptual signal (hierarchical SDT) or, in contrast, that confidence accumulates evidence after the perceptual decision (post-decisional models). Other theories put forth the possibility that perceptual and confidence decisions are the outcomes of two distinct decision-making systems. For instance, a model within this category casts the metacognitive system in the role of a Bayesian observer who infers confidence by observing the actions of the perceptual system (second order confidence model). On the opposite extreme to SDT is the dual-channel model which holds that the two decisions utilize information from completely independent sources. Lastly, the observed inefficiency of metacognition has also given rise to specific notions that confidence, unlike perceptual decisions, is processed heuristically. For example, it has been proposed that confidence neglects decision-incongruent evidence (positive evidence model) and automatically incorporates readily available, but task-irrelevant evidence (weighted evidence visibility model). In spite of an abundance of such diverse theories, studies have rarely performed extensive model comparisons to determine the plausibility of each theory relative to others. As a result, we do not yet understand how these theories stack up against each other in terms of empirical support. In this study, I propose to fit twelve process models (derived from the theories described above) to large datasets from two experiments in which participants completed a perceptual task with confidence ratings. Model performance will be mainly assessed in three ways – through a) quantitative assessment of their fits to data, b) qualitative assessment of their ability to predict basic empirical signatures of confidence and c) investigation of the properties of the models based on model and parameter recovery analyses. Particularly, model and parameter recovery can inform us about the internal consistency of model descriptions and their separability from other theories, which is crucial for making meaningful comparisons. Broadly, we can expect our analyses to falsify some of these theories while on the other hand, highlighting others that offer more plausible accounts of the process. These findings will provide insight into the computational mechanisms that drive confidence generation and help us understand to what degree each theory is congruent with the empirical phenomenon of confidence generation. In addition, through this study, I also aim to establish a framework for model evaluation that considers a diverse set of criteria, therefore increasing the reliability of our assessments.