Abstract: Neural circuits and artificial neural networks (ANNs) process information by constructing and manipulating highly distributed representations [1, 2, 3, 4]. Patterns of activity in these systems, across either neurons or units, correspond to manifold-like representations (Box 1) — lines [5], surfaces [6,7], trajectories [8•, 9••, 10••], subspaces [11], and clouds of points [12,13] — in a high dimensional ‘neural state space’, where coordinates represent the activities of individual neurons or units. Approaches focused on studying geometric properties of these manifolds are becoming more widely used as advances in experimental neuroscience expand our ability to probe large neural populations [14], and advances in ANNs [15,16] introduce new challenges of interpretation.

Authors: SueYeon Chung abd L.F. Abbott