The Redemption of Noise: Inference with Neural Populations

Rodrigo Echeveste, Máté LengyelTrends 2018 · DOI

TL;DR

This paper revisits the seminal 2006 work by Ma et al. that proposed "probabilistic population codes" (PPCs) as a framework for how neural populations represent uncertainty to perform Bayesian inference. It highlights how neural variability, traditionally seen as noise, is fundamental to this process, enabling the brain to encode probability distributions and perform computations with them.

What Problem Does This Paper Address?

The brain constantly makes sense of a noisy and incomplete world using sensory inputs. Traditionally, neural variability (the fact that the same stimulus elicits different neural responses each time) was considered a nuisance to be overcome. This paper, building on the 2006 Ma et al. paper, addresses how this variability is not a bug but a feature. It explores how neural populations can leverage this inherent noise to represent uncertainty and perform sophisticated Bayesian inference, a powerful framework for optimal decision-making under uncertainty. The core challenge is understanding how probability distributions, which inherently capture uncertainty, are encoded and manipulated by neural circuits.

Methods

The paper discusses the theoretical framework of Probabilistic Population Codes (PPCs) introduced by Ma et al. This framework posits that neural populations encode probability distributions over stimuli. The key insight is that the likelihood function, which describes the probability of a neural response given a stimulus (P(response | stimulus)), is central. This likelihood function inherently captures uncertainty. The paper highlights how a simple transformation of neural responses, specifically their summation, can implement the multiplication of likelihoods, which is the probabilistically correct way to combine independent sources of information (like different sensory cues) in Bayesian inference. The paper also touches upon the concept of linear decodability of stimulus likelihoods from neural responses, a crucial prediction of the PPC theory that has been experimentally supported.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Noise as a Feature: Neural variability is not a problem to be eliminated but a fundamental mechanism for representing uncertainty and performing probabilistic computations.
  • Bayesian Inference in the Brain: The brain likely employs Bayesian inference to make sense of ambiguous sensory information, and PPCs provide a theoretical framework for how this is achieved neurally.
  • Encoding Uncertainty: Neural populations encode uncertainty not by providing a single "best guess" but by representing a full probability distribution over possible stimulus values.
  • Optimal Cue Combination: The summation of neural responses in a downstream population can implement the multiplication of likelihoods from upstream populations, enabling optimal integration of information from multiple sources.
  • Linear Decodability: A key prediction of PPCs is that stimulus likelihoods should be linearly decodable from neural responses, a finding that has been experimentally validated.

Key Results

The paper doesn't present new experimental results but rather synthesizes and discusses the theoretical implications and experimental validation of the PPC framework. The key "result" is the conceptual shift: neural variability is essential for encoding uncertainty and performing Bayesian inference. The paper references experimental confirmations of the linear decodability of stimulus likelihoods from neural responses, suggesting that the brain does indeed operate according to PPC principles.

Limitations and Future Work

The paper acknowledges several limitations and areas for future research:

  • Nuisance Parameters: While the original PPC theory works well with a single nuisance parameter (like image contrast), real-world scenarios involve multiple, complex nuisance parameters. How PPCs handle this more general case is an active area of research.
  • Poisson-like Assumption: The assumption of Poisson-like neural responses (constant Fano factor) might not always hold true in biological systems, as response properties can change with stimulus onset, contrast, or task events.
  • Deterministic Output Layer: The idea of a strictly deterministic output layer in the PPC architecture might be difficult to reconcile with the inherently stochastic nature of cortical circuits.
  • Experimental Tests: While linear decodability has been confirmed, testing the full range of PPC predictions, especially in the presence of multiple nuisance parameters, remains challenging.
  • Alternative Theories: The paper notes that other theories exist for how neural variability supports probabilistic inference without necessarily relying on linear decodability. Future work will involve distinguishing between these competing models.

The authors suggest that continued exploration of these theories will lead to specific, testable predictions that will advance our understanding of the neural basis of probabilistic inference and how the brain makes sense of the world.

Summary updated

Abstract

In 2006, Ma et al. (Nat. Neurosci. 1006;9:1432-1438) presented an elegant theory for how populations of neurons might represent uncertainty to perform Bayesian inference. Critically, according to this theory, neural variability is no longer a nuisance, but rather a vital part of how the brain encodes probability distributions and performs computations with them.
BibTeX
@inproceedings{echeveste2018redemption,
  author = {Rodrigo Echeveste and Máté Lengyel},
  title = {{The Redemption of Noise: Inference with Neural Populations}},
  booktitle = {Trends},
  year = {2018},
  doi = {10.1016/j.tins.2018.09.003},
}
Figure