Learning to Predict Requires Integrated Information
Explore the relationship between Integrated Information Theory and predictive learning in embodied agents through this 21-minute conference talk. Delve into the quantitative approach to consciousness applied to neural networks, examining how embodied agents interact with their environment through the sensorimotor loop. Discover various information theoretic measures quantifying different information flows, including Integrated Information and Morphological Computation. Investigate the antagonistic relationship between these concepts and the potential problem it creates for embodied intelligence and conscious experience. Learn about a proposed solution suggesting that high Integrated Information is necessary for agents to predict future sensory states. Examine the dynamics of these measures in a simple experimental setup using the em-algorithm for agent learning, supporting the hypothesis that integrated information is crucial for learning processes.