Deep temporal models and active inference.

Area of Science:

• Cognitive Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Understanding how humans navigate complex, structured environments is a fundamental challenge.
• Previous work in active inference provides a framework for understanding decision-making and perception.
• Hierarchical generative models offer a way to represent structured data and temporal dependencies.

Purpose of the Study:

• To extend active inference using deep temporal models to explain behavior in structured environments.
• To simulate and account for both behavioral and electrophysiological responses during sequential inference.
• To model the accumulation of evidence over nested timescales for narrative understanding.

Main Methods:

• Formulation of active inference based on deep temporal models with hierarchical generative structures.
• Simulation of sequential inference by inverting hierarchical models of state transitions.
• Integration of Bayesian belief updating and neuronal process theories to model epistemic foraging in reading.

Main Results:

• Simulations successfully reproduced empirical findings, including perisaccadic delay period activity and local field potentials during reading.
• The deep temporal models accounted for evidence accumulation across nested timescales, enabling inferences about temporal scenes.
• Model simulations also replicated electrophysiological responses like mismatch negativity and P300 to local and global violations.

Conclusions:

• Deep temporal models within the active inference framework can explain complex behaviors in structured environments.
• This approach provides a unified account of sequential inference, narrative understanding, and responses to contextual violations.
• The findings highlight the potential of active inference for understanding brain function in perception and cognition.