A Spiking Neural Network Implementation of Gaussian Belief Propagation

Bayesian inference offers a principled account of information processing in natural agents. However, it remains an open question how neural mechanisms perform their abstract operations. We investigate a hypothesis where a distributed form of Bayesian inference, namely message passing on factor graphs, is performed by a simulated network of leaky-integrate-and-fire neurons. Specifically, we perform Gaussian belief propagation by encoding messages that come into factor nodes as spike-based signals, propagating these signals through a spiking neural network (SNN) and decoding the spike-based signal back to an outgoing message. Three core linear operations, equality (branching), addition, and multiplication, are realized in networks of leaky integrate-and-fire models. Validation against the standard sum-product algorithm shows accurate message updates, while applications to Kalman filtering and Bayesian linear regression demonstrate the framework’s potential for both static and dynamic inference tasks. Our results provide a step toward biologically grounded, neuromorphic implementations of probabilistic reasoning.

💡 Research Summary

This paper presents a novel framework for implementing Gaussian belief propagation, a core algorithm for efficient Bayesian inference, using biologically plausible spiking neural networks (SNNs). The central hypothesis is that the abstract operations of probabilistic message passing on factor graphs can be mapped onto the dynamics of networks composed of leaky-integrate-and-fire (LIF) neurons.

The authors begin by highlighting the gap between the Bayesian theory of brain function and the biophysical mechanisms of neural activity. To bridge this gap, they leverage Forney-style factor graphs (FFGs), which provide a graphical representation for factorized probabilistic models, enabling inference through localized message passing. The paper details how linear Gaussian models, such as those used in Kalman filtering, can be decomposed into fundamental building blocks: equality (branching), addition, and multiplication nodes. For each node type, the standard sum-product algorithm defines precise mathematical rules for updating Gaussian messages (represented by mean and variance parameters).

The core technical contribution lies in designing SNN modules that emulate these mathematical rules. The proposed method involves encoding the parameters of an incoming Gaussian message (e.g., its mean) into a spike-based signal, such as a firing rate pattern. This signal is then processed through a specifically configured network of LIF neurons. The connectivity and synaptic weights within each SNN module are engineered so that the collective spiking activity of its output neurons decodes back into the exact Gaussian message prescribed by the sum-product rule for that node. For instance, the SNN implementing an equality node performs the equivalent of summing incoming precision-weighted means and precisions.

The framework is validated through two main avenues. First, the individual SNN node implementations are shown to produce output messages that closely match those generated by the conventional sum-product algorithm, confirming their functional correctness. Second, the integrated system is applied to two classic inference problems: Bayesian linear regression and Kalman filtering. In the static regression task, the SNN successfully infers the posterior distribution over model parameters. In the dynamic filtering task, it accurately tracks a hidden state over time by sequentially processing new observations, demonstrating the approach’s capability for both static and dynamic inference.

The results provide a significant step towards neuromorphic implementations of probabilistic reasoning. By demonstrating that key Bayesian operations can be performed using spike-based computation, the work opens a path for developing energy-efficient, robust, and brain-inspired inference engines on specialized hardware like neuromorphic chips. Furthermore, it offers a concrete computational model for how neural populations might represent and manipulate uncertainty, thereby enriching the dialogue between machine learning, computational neuroscience, and signal processing.