Disentangled representations for causal cognition

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Complex adaptive agents solve problems requiring causal information about agent-environment systems.
• Causal cognition research describes learning and reasoning but lacks computational accounts for acquiring causal understanding without prior knowledge.
• Machine learning, particularly disentanglement, offers computational models for agents learning causality.

Purpose of the Study:

• To develop a unifying computational framework for causal cognition.
• To connect research in animal cognition and machine learning.
• To provide insights for developing new causal reinforcement learning algorithms.

Main Methods:

• Connecting psychological/behavioral research on causal cognition with machine learning approaches.
• Utilizing formal intervention-based models of causality (e.g., causal Bayesian networks).
• Investigating disentanglement as a process for building causal representations in artificial agents.

Main Results:

• The study proposes a novel computational perspective on causal cognition.
• It links the understanding of causal relationships in natural and artificial systems.
• It lays the groundwork for advanced AI algorithms capable of causal learning.

Conclusions:

• A unified framework for causal cognition is presented, integrating animal studies and AI.
• This framework offers a computational lens for understanding how agents learn causality from scratch.
• The research facilitates the development of more sophisticated artificial intelligence systems capable of causal reasoning.