Deep Active Inference and Scene Construction.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Adaptive agents operate in uncertain environments with complex structures.
• Perception and action are often modeled as approximate Bayesian inference.
• Hierarchical scene construction involves inferring higher-order patterns from ambiguous cues.

Purpose of the Study:

• To model visual foraging in a hierarchical context using active inference.
• To simulate agent decisions in categorizing scenes within a structured setting.
• To explain evidence accumulation phenomena through active inference principles.

Main Methods:

• Utilizing active inference to simulate agents inferring hierarchical visual patterns.
• Employing minimization of a free energy functional for perception and action.
• Analyzing the role of expected free energy in policy evaluation.

Main Results:

• Active inference in hierarchical scene construction reproduces empirical evidence accumulation phenomena.
• Observed behaviors include noise-sensitive reaction times and epistemic saccades.
• Novel agent behaviors provide new predictions for perceptual decision-making research.

Conclusions:

• Hierarchical active inference provides a principled framework for understanding planned information-gathering actions.
• The model offers insights into tasks requiring inference of compositional latent structure.
• This work bridges active inference with evidence accumulation models like drift-diffusion.