Learning and inference in the brain.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• The brain's functional architecture is shaped by centuries of anatomical and physiological study.
• Understanding how the brain processes sensory information is key to understanding cognition.
• Existing models often rely on assumptions not necessarily present in biological systems.

Purpose of the Study:

• To explore if functional brain anatomy principles can be predicted by representational learning theories.
• To investigate the role of hierarchical generative models and connection types in sensory processing.
• To assess the biological plausibility of learning mechanisms in the brain.

Main Methods:

• Reviewing hierarchical sensory cortex organization, focusing on forward and backward connections.
• Examining representational learning approaches as special cases of generative models.
• Analyzing learning based on empirical Bayes and its predictive power for brain function.

Main Results:

• Hierarchical generative models, particularly those using empirical Bayes, predict key brain features like hierarchical structure and top-down influences.
• Backward connections are crucial for recognition when input-generating processes are non-invertible, suggesting forward-only architectures are insufficient for perception.
• Non-linearities mediated by backward connections enable modulatory interactions between cortical levels, explaining functional asymmetries.

Conclusions:

• Hierarchical generative models offer a biologically plausible framework for understanding brain function and sensory processing.
• Backward connections play a critical, non-invertible role in perception and learning.
• The brain's architecture and processing align with predictions from empirical Bayes-based hierarchical learning models.