New math method classifies neural computation by dynamical archetypes

In a new paper on arXiv (2507.05505), researchers Sagodi and Park tackle the challenge of understanding neural computation by proposing Dynamical Archetype Analysis. They argue that complex neural systems should be abstracted by their 'ideal representative'—an archetype defined by asymptotic dynamical structure. To group systems by effective behavior, they introduce a novel dissimilarity measure that explicitly considers both deformations breaking topological conjugacy and diffeomorphisms preserving it. This measure can be estimated from observed trajectories, making it practical for real data.

Numerical experiments on high-dimensional recurrent neural networks (RNNs) for working memory tasks demonstrate the method's ability to overcome previously reported fragility in existing similarity measures for approximate continuous attractors. While the study focuses on working memory, the theoretical framework naturally extends to general mechanistic interpretation of recurrent dynamics in both biological neural circuits and artificial neural networks. The authors argue that abstract dynamical archetypes, rather than detailed system parameters, offer a more useful vocabulary for describing neural computation—potentially transforming how we compare neural systems across species and architectures.