NeuroMASS v4: Geometry-Aware Neural Population Dynamics via Learnable Hyperbolic Latent Representations and Biologically-Calibrated Diagonal State Space Models
In the authors' words
The geometry of neural population activity is a fundamental challenge in computational neuroscience. We present NeuroMASS v4, a novel architecture that introduces learnable hyperbolic latent geometry to spike-train co-smoothing on the Neural Latents Benchmark (NLB'21). Our central scientific claim is that the curvature c of a Poincaré-ball--learned end-to-end from extracellular spike recordings--constitutes a biologically interpretable geometric signature: significantly positive for ring-manifold motor circuits (MC_Maze, MC_RTT) and near-zero for topologically unstructured frontal circuits (DMFC_RSG). The architecture unifies five principled components: (i) discrete spike-count embed- dings with learned neuron-identity vectors, (ii) permutation-equivariant attention-weighted neuron pooling, (iii) a diagonal S4D state space model initialised with a log-uniform frequency prior spanning 2-200 Hz, (iv) a Poincaré-ball projection with learnable scalar curvature, and (v) a Frobenius-bounded session adapter. Training uses masked self-prediction of spike counts with an auxiliary behaviour decoding objective; all experiments use exclusively real DANDI/NLB recordings under a strict fail-stop data policy. We report results across four NLB datasets spanning motor, somatosensory, and cognitive cortices in rhesus macaque, and demonstrate that learned curvature statistically predicts circuit topology.
Appeared: Tuesday, November 17. bioRxiv. Preprint, not yet peer-reviewed.