Sources of predictive information in dynamical neural networks.

Area of Science:

• Computational neuroscience
• Information theory
• Behavioral biology

Background:

• Adaptive behavior relies on organisms predicting future environmental stimuli.
• Two proposed mechanisms for acquiring predictive information: internal models vs. direct environmental extraction.
• Current methods cannot distinguish between these predictive information acquisition mechanisms.

Purpose of the Study:

• To develop a framework for identifying the source of predictive information in agent-environment systems.
• To differentiate between predictions generated internally versus those extracted from the environment.
• To understand the role of information flow in adaptive behavior.

Main Methods:

• Decomposition of information transfer within the organism-environment system.
• Tracking information flow over time using bivariate mutual information.
• Validation using computational models of idealized agent-environment systems.

Main Results:

• Predictive information from the environment can be present in any agent, regardless of task performance.
• Predictive information originating from the nervous system requires specific adaptive dynamics.
• The amount of predictive information can vary for the same task based on environmental structure changes.

Conclusions:

• A novel framework allows for the identification of predictive information sources in agent-environment interactions.
• Distinguishing between internal and external prediction generation is crucial for understanding adaptive behavior.
• Information flow dynamics are key to understanding how organisms adapt to their environments.