Polysemanticity in human hippocampal neurons

Abstract

To comprehend language, the brain must navigate a high-dimensional semantic landscape while seamlessly contextualizing meaning. Inspired by recent advances in the mechanistic interpretability of large language models (LLMs), we hypothesized that the brain utilizes polysemanticity, a coding strategy wherein individual neurons represent multiple semantically unrelated features through high-dimensional superposition (Elhage et al., 2022; Olah et al., 2020). We recorded single-unit activity from the human hippocampus during podcast listening. We found that hippocampal neurons exhibit dense semantic codes characterized by multiple tuning peaks with an overdispersed, isotropic geometry. This geometry satisfies the theoretical requirements for interference minimization in superimposed codes. Furthermore, semantic responses are strongly modulated by lexical and speaker-identity context; nonetheless, the underlying population geometry remains stable. This coding strategy permits rapid contextualization without requiring specialized, context-specific neurons. Indeed, we show clear pattern separation of similar terms, along with pattern completion for held-out words. Together, these results demonstrate that the human brain leverages superposition to solve a universal computational problem: maximizing semantic capacity within a constrained representational space.