Symmetry and complexity in object-centric deep active inference models.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Neuroscience

Background:

• Humans leverage object symmetries for efficient learning and skill generalization.
• Active inference models agents as learning through minimizing free energy (surprisal).
• Deep active inference aims to create agents with sophisticated generative models.

Purpose of the Study:

• To investigate how object symmetries emerge in the latent space of deep active inference models.
• To analyze the relationship between model complexity and symmetry exploitation.
• To demonstrate the utility of learned symmetries for improved generalization.

Main Methods:

• Training object-centric generative models using deep active inference.
• Analyzing latent state spaces for emergent symmetry properties.
• Employing principal component analysis to identify symmetry axes in latent representations.

Main Results:

• Object symmetries are found to emerge as symmetries within the learned latent state space.
• Model complexity correlates with the degree of symmetry exploitation.
• Principal component analysis reveals encoding of object symmetry axes.

Conclusions:

• Deep active inference can learn and represent object symmetries.
• Symmetrical latent representations enhance generalization capabilities, particularly in manipulation tasks.
• This work bridges principles of active inference with the importance of symmetry in perception and action.