Skip to main content
Back to timeline
bioRxivSource publication:

NeuroMASS v4 learns curvature straight from spikes: near-positive in motor cortex, near-zero in a frontal timing circuit

Synopsis

NeuroMASS v4 makes the curvature of a hyperbolic latent space a learnable parameter trained end-to-end on extracellular spike recordings for neural population spike co-smoothing. Earlier latent-variable models such as GPFA, LFADS and NDT embedded population activity in Euclidean space, which cannot represent ring manifolds without distortion. Across four macaque-cortex datasets from NLB'21, learned curvature was 0.84±0.11 on motor-cortex MC_Maze and 0.71±0.14 on MC_RTT, but 0.07±0.04 on frontal DMFC_RSG, not significantly above zero.

Source-provided article image: NeuroMASS v4: Geometry-Aware Neural Population Dynamics via Learnable Hyperbolic Latent Representations and Biologically-Calibrated Diagonal State Space Models
Figure 1 · bioRxiv · Page 4

Interpretation

A neural population decoding model now treats how strongly its latent space bends as a trainable parameter, letting the data decide how curved a space the population activity needs. Earlier neural latent-variable models GPFA, LFADS, AutoLFADS, NDT and PopT all used Euclidean latent spaces, which necessarily distort ring manifolds. On the MC_Maze standard validation split, NeuroMASS v4 reached co-bps 0.421±0.011, above the geometry-free NeuroMASS-Euclid at 0.381±0.014, LFADS at 0.347±0.021 and NDT at 0.301±0.018.

The model turns each time bin's spike count and neuron identity into a vector, processes it over time with diagonal state-space layers, then maps it into a Poincaré ball. The diagonal state-space model S4D compresses long sequences through a set of complex modes; here its oscillation frequencies are initialised from a log-uniform prior over 2 to 200 Hz, spanning the theta, beta and gamma bands. In the MC_Maze small-set ablation, replacing the frequency prior with random initialisation lowered co-bps from 0.421 to 0.397, and removing masked self-prediction lowered it to 0.362.

Learned curvature orders monotonically by brain region, highest in motor cortex and near zero in the frontal timing circuit. DMFC_RSG neurons encode elapsed time as a ramp, topologically a line rather than a ring, so curvature near zero was expected. Across five seeds, curvature was 0.84±0.11 on MC_Maze (t=11.3, p<10⁻⁵), 0.71±0.14 on MC_RTT (t=7.6, p<10⁻⁴), 0.33±0.09 on Area2_Bump (t=5.1, p<0.001) and 0.07±0.04 on DMFC_RSG (t=1.75, p=0.12).

Perspective

The demonstrated scope is the validation splits of four macaque-cortex datasets in NLB'21 (MC_Maze, MC_RTT, Area2_Bump, DMFC_RSG), all real DANDI recordings with no synthetic substitution; the authors propose next running on mc_maze_large and submitting to the official NLB test-split evaluation.

A single scalar curvature assumes isotropic curvature, while real manifolds may be mixed-curvature; FFT-convolution memory grows as O(T logT) with sequence length, limiting continuous recordings beyond roughly 10,000 time bins; NLB'21 lacks paired multi-subject recordings, so cross-subject transfer was not evaluated; all results come from non-human primates, and the authors stress that no clinical interpretation should be made before validation on human ECoG and Neuropixels recordings.