Mamba forecaster decodes mouse behavior from neural spike forecasts

A team led by John Minnick et al. has shown that a single Mamba sequence model, trained solely to predict the next time-step's spike counts from population-scale neural recordings, can simultaneously decode behavioral state. The model is applied to the Steinmetz visual-discrimination benchmark (39 sessions, ~27,000 neurons, 1,994 held-out trials). By feeding the model's predicted firing rates into a per-session linear classifier, the system decodes mouse choice at 75.7±0.2% trial-vote accuracy (roughly 2.3x chance) and stimulus side at 66.1±0.6% (about 2x chance). This outperforms a matched linear decoder that uses raw spike counts with 500 ms of context by 4–6 percentage points on both metrics.

The approach is remarkably practical: a calibration block of just 100–150 trials at session start brings the readout within 1–2 percentage points of asymptotic performance. The entire pipeline—from spike binning to classification—runs within the 50 ms bin budget typical of tethered chronic Neuropixels recordings when executed on workstation-class GPUs. This suggests that implicit forecasting of neural population dynamics can serve as a drop-in replacement for traditional feature engineering in closed-loop brain-computer interfaces, reducing calibration time and improving decoding accuracy without bespoke per-session tuning.