How the brain might work: statistics flowing in redundant population codes

Xaq Pitkow, Dora E. AngelakiarXiv 2017 · DOI · arXiv

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.

Beyond Simple Decisions

Neuroscientists have long hypothesized that the brain operates as a "Bayesian machine," continuously calculating the most likely state of the environment based on noisy and ambiguous sensory inputs. Despite this theoretical consensus, the specific "neural code" used to perform these complex statistical calculations remains elusive.

A primary obstacle in current neuroscience is the reliance on overly simplistic experimental paradigms, such as two-alternative forced-choice (2AFC) tasks. While these tasks—like asking a subject to identify the direction of moving dots—are easy to control, they are inherently low-dimensional. They do not require the brain to engage its full computational capacity for "untangling" task-relevant information from the sea of nuisance variables (such as lighting changes, shadows, or varying perspectives) encountered in the real world. To understand the brain's true software, we must bridge the gap between high-level behavioral goals and the low-level firing patterns of neurons.

A Unified Framework for Probabilistic Inference

This perspective proposes a synthesis of three major pillars of computational theory to explain neural computation: Probabilistic Graphical Models (PGMs), Message-Passing Algorithms, and Redundant Population Codes.

PGMs allow us to represent the world as a network of interacting variables. Rather than attempting the mathematically impossible feat of calculating every possible outcome, the brain likely tracks direct interactions between related variables. To reach a global consensus on the state of the world, different parts of the neural network exchange "messages"—statistical summaries—in an iterative process known as message-passing.

The proposed framework mapping probabilistic graphical models (A) to neural population dynamics and message-passing (B).

As illustrated in the schematic above, the framework maps these abstract mathematical concepts onto biological reality. The "representational level" is key: because neural codes are highly redundant, information is not held by single neurons but is spread across large populations. This redundancy makes the system robust to noise and individual neuron failure, while allowing for the complex, nonlinear transformations necessary to recover information hidden by nuisance variables.

Statistics as "Flow"

In this model, inference is not a static snapshot but a dynamic process. Summary statistics—such as the mean or the local uncertainty of a variable—literally "flow" through the brain's recurrent connections as an animal processes a scene. The paper outlines a unified analysis framework to test this hypothesis in neural data:

  1. Encoding: Identifying the "summary statistics" in neural activity that represent latent variables in a behavioral model.
  2. Recoding: Testing whether the dynamics of these neural statistics follow the mathematical rules of message-passing (e.g., how $\theta_i(t)$ updates to $\theta_i(t+1)$ based on neighbors).
  3. Decoding: Verifying that these population-level statistics actually predict the animal's subsequent behavior.

By focusing on the coarse-scale population statistics rather than the fine-scale wiring of every synapse, researchers can identify the essential features of the neural algorithm. The authors argue that "information-limiting correlations"—noise that mimics signal—actually make these population-level computations more interpretable to researchers, as they highlight the specific statistics the brain is prioritizing.

Future Directions and Limitations

While this framework provides a rigorous roadmap, several challenges remain. Exact probabilistic inference is computationally intractable for biological systems; therefore, the brain must utilize clever approximations of message-passing. Identifying the nature of these approximations is a major goal for future research. Additionally, while the current framework focuses on how the brain uses an internal model, it does not yet account for how that model is learned through experience.

Finally, the authors call for a shift in experimental design. To reveal the brain's capacity for probabilistic inference, we must move toward moderately complex, naturalistic behavioral datasets. These tasks should manipulate uncertainty and require animals to predict unseen variables, forcing the brain to engage the internal models and message-passing dynamics described here. Only by studying the brain in its natural computational element can we hope to decode the software of the mind.

Summary updated

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.
BibTeX
@misc{pitkow2017how,
  author = {Xaq Pitkow and Dora E. Angelaki},
  title = {{How the brain might work: statistics flowing in redundant population codes}},
  year = {2017},
  eprint = {1702.03492},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1702.03492},
}
Figure