Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Testing the Bayesian Confidence Hypothesis: Model Space, Linking Hypothesis and Empirical Fit

Xinyi Yuan¹, Ralf M Haefner¹; ¹University of Rochester

Presenter: Xinyi Yuan

How the brain computes confidence remains an open question in computational cognitive science. The Bayesian confidence hypothesis -- that confidence reflects the posterior probability of being correct -- has attracted both supporting and contradicting empirical evidence, yet model comparisons to alternatives are complicated by various degrees of freedom in model specification. Here, we first provide a systematic analysis of the free parameters involved in testing the Bayesian confidence hypothesis: (i) the choice of generative model used by the observer, and (ii) the link function mapping posterior beliefs to confidence reports. We propose a model space defined by two axes: how closely the generative model approximates the environment, and how much the model respects neural constraints. Second, we apply this framework to previously published data(Adler & Ma, 2018), who tested Bayesian and non-Bayesian models of confidence in a visual orientation-discrimination task. We show that a subtly modified link between the posterior and confidence judgments -- unspecified by the Bayesian framework -- substantially improves the model’s fit to empirical data. We further demonstrate how goodness-of-fit depends systematically on the location of the generative model in the proposed model space. Finally, we find that the more realistic and complete Bayesian models outperform a recently proposed signal-detection-theory-inspired two-stage process model(Boundy-Singer et al., 2023). Together, our results clarify the debate around the Bayesian confidence hypothesis and offer a principled approach to future model comparisons.

Topic Area: Methods, Tools, Theory & Neural Coding