Neural implementations of Bayesian inference.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Decision Making

Background:

• Bayesian inference provides a normative framework for understanding decision-making under uncertainty.
• Experimental evidence reveals diverse neural mechanisms underlying probabilistic inference in the brain.
• Discrepancies exist between theoretical predictions and observed neural implementations of Bayesian computations.

Purpose of the Study:

• To categorize the neural implementations of Bayesian inference into distinct classes.
• To provide a framework for understanding how the brain performs probabilistic computations.
• To identify areas for future research and reconciliation between different theoretical and experimental approaches.

Main Methods:

• Literature review and synthesis of experimental neuroscience findings.
• Categorization of neural implementations based on computational principles.
• Analysis of theoretical models of Bayesian inference in neural systems.

Main Results:

• Two broad classes of neural implementations for Bayesian inference were identified: modular and transform.
• The modular class involves independent encoding of probabilistic components.
• The transform class utilizes latent processes to convert uncertain measurements into Bayesian estimates.

Conclusions:

• Current experimental findings on probabilistic inference largely fit within these two identified classes.
• Reconciliation between the modular and transform classes presents challenges and opportunities for future research.
• Distinguishing between neural implementation hypotheses requires interdisciplinary collaboration across scales, circuits, tasks, and species.