Inference in the Brain: Statistics Flowing in Redundant Population Codes.

Area of Science:

• Computational neuroscience
• Cognitive science
• Machine learning

Background:

• The brain infers causal variables from ambiguous sensory data using probabilistic inference.
• Understanding these computations requires analyzing information processing in nonlinear recurrent neural networks.

Purpose of the Study:

• To propose a framework where probabilistic computations are implemented via message-passing algorithms.
• To outline methods for identifying neural message-passing algorithms.

Main Methods:

• Reviewing concepts of graphical models, sufficient statistics, and message-passing.
• Describing implementation in recurrently connected probabilistic population codes.
• Developing an approach to identify neural message-passing algorithms.

Main Results:

• Information flow is interpretable at the population level, especially with redundant neural codes.
• A general approach to identify neural message-passing algorithms is outlined.

Conclusions:

• Probabilistic inference in the brain may operate through neural message-passing.
• Studying large-scale neural activity during naturalistic behaviors is crucial for understanding brain computations.