Abstract
A foundational challenge in neuroscience and AI is understanding how physical space is mapped into neural representations. While artificial neural networks can learn brain-like spatial representations, their ``black-box'' nature obscures whether these representations reflect general principles or artifacts of specific architectures, objectives or optimization choices, offering no guarantee that solutions for tasks like path integration generalize beyond training data. Here, we use a first-principles framework based on the matrix exponential, in which generator matrices map physical locations to neural population vectors without gradient-based optimization. Within this formalism, navigational properties follow directly from algebraic constraints on the generators: commuting generators enable trajectory-invariant, exact path integration, while skew-symmetry yields equinorm representations with translation-invariant similarities. Preserving the flat Euclidean metric further constrains the generator spectra to balanced wavevector configurations, such as the roots-of-unity constellations that generate periodic grid fields. Similarity preservation further produces localized place-like fields, and extending the generators to a context signal models context-dependent remapping. Beyond its algebraic relation to continuous attractor and oscillatory interference models, the framework natively supports goal-oriented multi-map navigation through hyperdimensional bundling and binding, fits recorded grid and place cell data, extrapolates their rate maps far beyond the fitted region, and makes testable predictions for entorhinal circuits. By grounding spatial representations in exact algebraic conditions, this work offers a transparent alternative to empirical representation learning, revealing the conditions under which a coherent neural map of space arises.
Funding Information
  • Sure-AI, Simula Research Laboratory, the Norwegian research council (Grants 357482 and 300504)