Inferring Inference

Rajkumar Vasudeva Raju, Zhe Li, Scott Linderman, Xaq PitkowarXiv 2023 · DOI · arXiv

Abstract

Understanding how the brain processes information requires more than just recording neural activity; it requires identifying the underlying computations. This paper introduces "Inferring Inference," a mathematical framework that treats neural population activity as a distributed message-passing system. By applying this framework to synthetic datasets, the authors demonstrate that it is possible to recover both the brain's internal model of the world and the specific nonlinear rules it uses to perform probabilistic inference.

Introduction

Neuroscience has entered an era of "big data," with the ability to record thousands of neurons simultaneously. However, a significant gap remains between observing neural spikes and understanding the distributed computations they perform. While theories of canonical neural circuits exist, they often focus on local interactions. In reality, neural representations are highly multiplexed—information about a single variable is spread across many neurons, and a single neuron may contribute to multiple variables.

The core challenge addressed here is: if the brain acts as a "Bayesian machine" performing inference on an internal model of the world, can we deduce that model and the specific inference algorithm (e.g., belief propagation) directly from raw neural data?

The Three-Tiered Framework

The authors propose a hierarchical approach to bridge the gap between normative theories and mechanistic implementations:

  1. Normative Level: The brain is assumed to maintain a Probabilistic Graphical Model (PGM) to infer latent causes from sensory inputs.
  2. Algorithmic Level: Inference is implemented via nonlinear message-passing on a graph. These computations are assumed to be "canonical," meaning the same mathematical rules for information updates are reused throughout the network.
  3. Mechanistic Level: Neural population activity is viewed as a linear or nonlinear embedding of these latent message-passing states.
The framework relates the true world state and observations to the brain's internal inference model and neural representations.

Methodology: Inferring the Inference

To recover the hidden algorithms, the authors utilize a Hidden Markov Model (HMM) where the latent states represent the messages being passed within the brain. They employ a Particle Expectation-Maximization (Particle EM) algorithm:

  • E-step: A particle filter (Sequential Monte Carlo) estimates the most likely trajectories of hidden messages given observed neural data.
  • M-step: Gradient ascent updates the parameters of the neural encoding, the structure of the world model, and the nonlinear message-passing functions.

To address the high dimensionality and potential degeneracies in the parameter space, the authors use greedy backward optimization (pruning) to identify the most parsimonious set of rules that explain the observed data.

The mechanistic model of neural representation is contrasted with the brain's assumed probabilistic model of the world.

Key Insights and Results

The framework was validated using a "TAP Brain"—a model recurrent neural network trained to implement the Thouless-Anderson-Palmer (TAP) equations, a sophisticated form of mean-field inference.

  • Algorithmic Symmetry: The reuse of the same nonlinear functions across the graph creates a statistical symmetry that makes the algorithm identifiable even in noisy data.
  • Stimulus Design: Recovery is most successful within a "Goldilocks" zone of sensory input intensity. Inputs must be strong enough to reveal interactions but not so strong that the system becomes purely linear.
  • Successful Recovery: The model correctly identified the number of latent variables and successfully recovered the ground-truth coupling matrix and polynomial coefficients of the message-passing function.
  • Predictive Accuracy: The inferred model accurately predicted neural population responses to novel sensory stimuli, demonstrating its robustness and generalizability.

Discussion and Future Directions

While this framework provides a powerful tool for discovering distributed computations, several challenges remain. The current model assumes a static internal world model, whereas biological systems often perform temporal and causal inference. Future work will aim to scale the framework to continuous variables and higher-order interactions. The ultimate goal is to apply "Inferring Inference" to real electrophysiological data from animals, providing a direct link between functional algorithms and biological neural circuits.

Summary updated

Abstract

Patterns of microcircuitry suggest that the brain has an array of repeated canonical computational units. Yet neural representations are distributed, so the relevant computations may only be related indirectly to single-neuron transformations. It thus remains an open challenge how to define canonical distributed computations. We integrate normative and algorithmic theories of neural computation into a mathematical framework for inferring canonical distributed computations from large-scale neural activity patterns. At the normative level, we hypothesize that the brain creates a structured internal model of its environment, positing latent causes that explain its sensory inputs, and uses those sensory inputs to infer the latent causes. At the algorithmic level, we propose that this inference process is a nonlinear message-passing algorithm on a graph-structured model of the world. Given a time series of neural activity during a perceptual inference task, our framework finds (i) the neural representation of relevant latent variables, (ii) interactions between these variables that define the brain's internal model of the world, and (iii) message-functions specifying the inference algorithm. These targeted computational properties are then statistically distinguishable due to the symmetries inherent in any canonical computation, up to a global transformation. As a demonstration, we simulate recordings for a model brain that implicitly implements an approximate inference algorithm on a probabilistic graphical model. Given its external inputs and noisy neural activity, we recover the latent variables, their neural representation and dynamics, and canonical message-functions. We highlight features of experimental design needed to successfully extract canonical computations from neural data. Overall, this framework provides a new tool for discovering interpretable structure in neural recordings.
BibTeX
@misc{raju2023inferring,
  author = {Rajkumar Vasudeva Raju and Zhe Li and Scott Linderman and Xaq Pitkow},
  title = {{Inferring Inference}},
  year = {2023},
  eprint = {2310.03186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.03186},
}
Figure