A Geometric Free-Energy Framework for Pragmatic Neural Population Codes

This record contains the preprint "A Geometric Free-Energy Framework for Pragmatic Neural Population Codes" by Ryotaro Saito (2025).

Biological systems must act and communicate under pervasive uncertainty. The Free Energy Principle (FEP) and active inference describe such systems as minimising variational and expected free energy, but it remains unclear how this view constrains pragmatic communication and concrete neural population codes. This work develops a geometric framework that links an expected free energy (EFE)–based control objective, probabilistic speaker–listener models from pragmatic language theory, and information-theoretic measures of communication quality that can be evaluated on large-scale neural data.

On the "speaker" side, the paper posits a message-level EFE G(u; w, c) for triples of message u, world state w, and context c, and shows how an EFE-based softmax over messages induces a pragmatic speaker compatible with Rational Speech Act–style models. A behaviour-preserving reparameterisation separates pure generative semantics, normalisation-adjusted production cost, and expected information gain, clarifying which components are in principle identifiable from behaviour.

On the "listener" side, neural population responses are embedded into a latent metric space equipped with a Gaussian prototype geometry. An ideal listener combines latent likelihoods with task priors via Bayes’ rule, while actual listeners are implemented as logistic decoders on the same latent codes. From their posteriors, the paper derives two unitful diagnostics: a single-trial information gain and a KL-based pragmatic loss that quantify, in nats, how effectively a given listener exploits information encoded in a sender population.

The framework is instantiated on Neuropixels recordings from the International Brain Laboratory (IBL) visual decision-making task. Across 13 probe insertions (8,476 trials, 2,344 neurons), the latent geometry supports robust ideal decoding, whereas logistic listeners systematically underuse available information. A four-pattern decomposition of pragmatic loss disentangles contributions from task priors, latent geometry, and decoder class, showing that much of the apparent loss under biased priors is attributable to deliberately withholding block priors from the main decoder.

Although the empirical application focuses on cortical spike trains and a binary decision task, the construction is substrate-agnostic: wherever world states, contexts, messages, and an appropriate geometric representation can be defined, the same pair of metrics (information gain and pragmatic loss) can be used to quantify communication quality under uncertainty. Code and summary data for the analyses are available in a separate Zenodo repository (DOI: 10.5281/zenodo.17718560).