[Paper Review] How the brain might work: statistics flowing in redundant population codes

Xaq Pitkow, Dora E. Angelaki|arXiv (Cornell University)|Feb 12, 2017

Neural dynamics and brain function90 references3 citations

TL;DR

This paper proposes that the brain performs probabilistic inference through a message-passing algorithm operating on redundant population codes in recurrent neural networks. By framing neural computations in terms of sufficient statistics and graphical models, it shows how population-level activity can represent and transform statistical information, offering a biologically plausible mechanism for perception under uncertainty.

ABSTRACT

It is widely believed that the brain performs approximate probabilistic inference to estimate causal variables in the world from ambiguous sensory data. To understand these computations, we need to analyze how information is represented and transformed by the actions of nonlinear recurrent neural networks. We propose that these probabilistic computations function by a message-passing algorithm operating at the level of redundant neural populations. To explain this framework, we review its underlying concepts, including graphical models, sufficient statistics, and message-passing, and then describe how these concepts could be implemented by recurrently connected probabilistic population codes. The relevant information flow in these networks will be most interpretable at the population level, particularly for redundant neural codes. We therefore outline a general approach to identify the essential features of a neural message-passing algorithm. Finally, we argue that to reveal the most important aspects of these neural computations, we must study large-scale activity patterns during moderately complex, naturalistic behaviors.

Motivation & Objective

• To explain how the brain might perform approximate probabilistic inference from ambiguous sensory inputs.
• To bridge theoretical frameworks like graphical models and message-passing with biological neural circuits.
• To identify how redundant population codes in recurrent networks can represent and transmit sufficient statistics.
• To argue that large-scale neural activity patterns during naturalistic behaviors are essential for uncovering core computational principles.
• To propose a framework where population-level information flow reveals the brain's statistical computations.

Proposed method

• Adapts probabilistic graphical models and message-passing algorithms to neural population codes.
• Models neural populations as representing sufficient statistics of sensory inputs through nonlinear recurrent dynamics.
• Uses redundant coding to ensure robustness and interpretability of statistical information in neural activity.
• Applies the sum-product algorithm to population-level variables, treating neural ensembles as stochastic nodes.
• Analyzes how recurrent connectivity enables iterative refinement of beliefs about hidden causes in sensory data.
• Proposes that information flows as sufficient statistics across populations, not individual neurons.

Experimental results

Research questions

• RQ1 How can recurrent neural networks implement probabilistic inference using population codes?
• RQ2 What role do redundant neural codes play in making statistical computations interpretable at the population level?
• RQ3 How can message-passing algorithms be mapped onto neural circuit dynamics?
• RQ4 What kind of neural activity patterns reveal the most critical aspects of brain computation?
• RQ5 How do sufficient statistics emerge and flow through neural populations during perception?

Key findings

• Redundant population codes allow the brain to represent and transmit sufficient statistics of sensory inputs reliably.
• Message-passing algorithms operating on populations can perform approximate Bayesian inference in a biologically plausible manner.
• The most interpretable information flow occurs at the population level, especially in redundant codes.
• Recurrent connectivity enables iterative refinement of beliefs about hidden causes using sufficient statistics.
• Large-scale activity patterns during naturalistic behaviors are essential for uncovering the core computational logic of neural circuits.
• The framework provides a unifying perspective linking probabilistic inference, neural coding, and network dynamics.