Neural manifolds - The Geometry of Behaviour

Neural activity across populations of neurons doesn’t wander through a high-dimensional space at random. Instead, the firing-rate patterns trace out a lower-dimensional “neural manifold” whose intrinsic geometry—especially its dimension and topology—can reveal what latent variables a brain circuit is encoding. That matters because it offers a concrete way to move from raw spike trains to testable claims about the variables behind behavior, without needing to guess the encoding scheme from first principles.

The path starts with how population activity is represented. With multi-electrode arrays, researchers can record from up to a few hundred neurons at once, marking when each neuron spikes. By binning time into short intervals and counting spikes per bin (often smoothing the result), each moment in time becomes an n-dimensional vector: one coordinate per neuron’s instantaneous firing rate. As an animal forages or performs a task, that vector changes, producing a trajectory through this n-dimensional “empirical neural activity space.” But physiological constraints—such as limits on maximum firing rates—already rule out arbitrary points, and the fact that neurons interact means the trajectory is confined to a subspace rather than filling the whole ambient space.

The key mathematical idea is that the relevant structure can be captured by manifold geometry. Manifolds are shapes that look locally like lower-dimensional Euclidean space, even if they sit inside a higher-dimensional ambient space. Algebraic topology formalizes which global properties survive continuous deformation: two shapes can be homeomorphic (deformable without tearing or gluing), yet still differ in crucial ways—like whether a surface has a hole. The discussion emphasizes two invariants that are especially useful for neural data: intrinsic dimension (the number of independent degrees of freedom needed to specify a point on the manifold) and genus/topology (how many “holes” the manifold has).

Dimension helps translate geometry into coding capacity. If a neural population’s activity lives on a two-dimensional manifold, it can’t independently represent four independent variables (for example, x and y position plus two independent velocity components) because the dynamics would require at least four degrees of freedom. Topology adds a second constraint: holes correspond to variables with circular or constrained structure. A torus and a sphere can look locally identical from the surface, yet differ globally because one contains a “hole” that prevents certain deformations.

A concrete example comes from head-direction cells in the mouse. Recording from neurons in the thalamus that participate in the head-direction system, researchers obtain a cloud of points in high-dimensional firing-rate space and reconstruct its shape under the hypothesis that the circuit encodes head angle. If the circuit truly represents only facing direction, the manifold should be one-dimensional (one variable) and have a hole consistent with angular wraparound. The analysis finds activity localized to a one-dimensional ring. Moreover, equal arc lengths along the ring correspond to equal differences in facing angle, yielding a direct one-to-one mapping between manifold position and behavioral direction.

The broader takeaway is that topological data analysis and computational neuroscience are beginning to converge. By treating neural population activity as geometric data—recovering intrinsic dimension and topology—researchers can infer what latent variables are being represented in circuits involved in navigation, movement, spatial mapping, and potentially higher-level abstractions. The approach reframes “what the brain encodes” as a question about the geometry of population dynamics.