Matthias Guggenmos: Computational models of confidence

Professor for Computational Cognitive Neuroscience, HMU Potsdam

Title: Computational models of confidence

Abstract:

Confidence is a central feature of human metacognition: we can evaluate our own thoughts, actions, and perceptions, communicate uncertainty, and use these evaluations to guide future behavior. Yet confidence remains computationally elusive: we lack a precise account of how it is generated, distorted, and utilized as a signal for decision-making and learning.

In the first part of this talk, I will introduce current computational models of confidence and show how they begin to address fundamental questions about human confidence computations. Do humans compute Bayes-optimal probabilities? Are humans truly overconfident? I will discuss key parameters – metacognitive biases and metacognitive noise – and how they relate to other trait-like features of human behavior.

In the second part, I will turn to the idea that confidence can reinforce behavior when external feedback is unavailable. Building on behavioral and neuroimaging evidence, I will discuss confidence prediction errors as neural and computational analogues of reward prediction errors, linking metacognitive evaluation to learning.

Together, these lines of work demonstrate how computational approaches can bridge the gap between low-dimensional behavioral outputs – choices and confidence ratings – and the underlying computational mechanisms of metacognition and learning.

 

Guests are welcome!

 

Organized by

Michael Brecht / Lisa Rosenblum