Deep active inference.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• The free energy principle offers a unified framework for understanding brain function and active inference.
• Advances in deep generative models and variational inference provide powerful tools for approximating complex probability distributions.
• Evolution strategies offer robust optimization methods for complex systems.

Purpose of the Study:

• To introduce a novel
• deep active inference
• agent by integrating the free energy principle with deep learning and evolution strategies.
• To demonstrate how this agent can achieve goal-directed behavior through learned generative models and active sampling of sensory input.
• To showcase the scalability and flexibility of this approach for complex cognitive tasks.

Main Methods:

• The deep active inference agent optimizes a variational free energy bound on sensory surprise.
• It employs deep and recurrent neural networks for internal dynamics and parameter optimization.
• Generative latent variable models and variational densities approximate environmental dynamics and posterior beliefs.

Main Results:

• Goal-directed behavior was successfully implemented in the mountain car problem by defining appropriate latent state priors.
• The agent demonstrated the ability to learn a generative model of its environment.
• Sampling from the learned model provided insights into the agent's beliefs and interactions.

Conclusions:

• Deep active inference provides a scalable and flexible framework for creating intelligent agents.
• This approach integrates perception, action, and learning under a unified principle.
• The framework holds promise for advancing artificial intelligence and understanding biological cognition.