Reproduce “Inverse decision-making using neural amortized Bayesian actors” on OpenResearch

About this paper

Dominik Straub, Tobias F. Niehues, Jan Peters, Constantin A. Rothkopf

Researchers at the Technical University of Darmstadt developed a method utilizing neural networks to efficiently solve the inverse decision-making problem, inferring latent parameters like perceptual uncertainty, prior beliefs, and subjective cost functions directly from continuous behavioral data. This approach overcomes computational barriers, enabling the application of complex Bayesian actor models to naturalistic tasks and bridging the gap between normative theories and descriptive accounts of human behavior.

Key insights

• Unsupervised training of neural networks can effectively amortize the intractable Bayesian decision-making problem, eliminating the need for pre-computed optimal actions.
• Integrating a differentiable neural actor enables rapid, gradient-based Bayesian inference of subjective parameters like perceptual uncertainty, prior beliefs, and cost functions from continuous behavioral data.
• Carefully designed multi-condition experiments can resolve identifiability issues between confounding parameters, such as prior beliefs and specific cost function parameters.

Summary by alphaXiv.