Concurrent model evidence computation and posterior sampling in continuous attractor network subspaces
Extensive studies suggest the brain performs Bayesian inference to infer the latent world states. It is a fundamental neuroscience question that how canonical recurrent neural circuits in the brain implement Bayesian inference. Many existing theoretical studies focused on how the recurrent circuits compute the posterior, while largely overlooking how the circuits compute the model evidence…
Recent studies indicate that the brain employs Bayesian inference to deduce latent world states. A key question in neuroscience is how recurrent neural circuits within the brain carry out Bayesian inference. While many theoretical works have explored how these circuits compute the posterior, the computation of the model evidence, a critical component in Bayes' theorem, has remained largely unexplored.
This study investigates continuous attractor networks, a well-known recurrent circuit model, and provides substantial theoretical insights. It demonstrates that nonlinear dynamics within the circuit can perform both posterior sampling and model evidence computation in the first two significant subspaces of the circuit dynamics. The stimulus feature subspace facilitates Langevin posterior sampling, while the total neuronal activity subspace calculates the model evidence, similar to the evidence lower bound in stochastic variational inference.
The researchers further expanded the model to accommodate multiple inputs, and simulations confirmed the computations occurring within the network. This groundbreaking work reveals, for the first time, the coexistence of model evidence and posterior sampling in subspaces within continuous attractor networks, substantially enhancing our understanding of the computational algorithms utilized by neural circuits.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
