Stabilizing BCI through alignment of latent dynamics

• Alignment of latent state spaces can stabilize neural decoding over time and is critical for real-world BCI use and Braingate-style implants.¹
• The approach is computational and applicable to existing iEEG/ECoG pipelines.^{1 1}

Weekly enrichment (2026-07-20)

• The underlying method is NoMAD (Nonlinear Manifold Alignment with Dynamics), an unsupervised platform that stabilizes intracortical BCI decoding by aligning nonstationary neural activity to a fixed recurrent-neural-network model of latent dynamics built on the LFADS (latent factor analysis via dynamical systems) architecture.²
• On a two-dimensional isometric wrist-force task recorded from a monkey’s primary motor cortex (M1) with a 96-channel array across 20 sessions spanning 95 days, NoMAD achieved a median across-session force-decoding R² of 0.91, versus 0.65 for ADAN, 0.59 for aligned factor analysis (Aligned FA), and 0.14 for a static (unadapted) decoder.²
• NoMAD’s decoding “half-life” was 208.7 days compared with 45.1 days for Aligned FA, and it produced zero decoding failures across 380 session pairs, versus 223 failures for the static decoder and 51 for Aligned FA.²
• Alignment is unsupervised: the Day-0 LFADS generator (the dynamics model) and Wiener-filter decoder are frozen, and only a feedforward alignment network plus low-dimensional read-in/rate-readout are updated by minimizing the Kullback-Leibler divergence between Day-0 and Day-K generator-state distributions plus a Poisson negative-log-likelihood reconstruction term.²
• Behavioral decoding used a Wiener filter mapping manifold activity to 2-D forces at 20 ms bins, and inference was run with a causal sliding-window approach to emulate real-time online BCI use.²
• Within-session (Day-0) baseline performance for NoMAD’s generator-state decoder was a median R² of 0.971 (Q1–Q3: 0.965–0.974), the highest of all methods tested.²
• A second, dynamically distinct center-out reaching task (12 sessions over 38 days, 96 M1 channels) showed NoMAD again outperforming ADAN, Aligned FA, and static decoders on cursor-velocity decoding, demonstrating generalization beyond the isometric task.²
• The peer-reviewed version appeared in Nature Communications (2025); an earlier bioRxiv preprint reported the same 95-day isometric and 38-day reaching results, indicating stable feasibility of long-term unsupervised iBCI use with less frequent recalibration.³
• The approach builds on prior unsupervised manifold-alignment work using adversarial networks—ADAN and Cycle-GAN—where Cycle-GAN aligns full-dimensional recordings and was reported as more robust and easier to train, maintaining fixed-decoder performance over months to years despite neuron turnover.⁴

Footnotes

1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE5BZUhsQWRHcnlvTllUYTA2bXJyNVpyai1nZW1HMVpteFZxcGpmdXZDbzFXQnhkb0hxOVlDbXcyVWp5Mmd0MnVqb3pnOFllZDhhVGtueWNrSzVnOVk1ZWlZ?oc=5 ↩ ↩² ↩³

2. https://www.nature.com/articles/s41467-025-59652-y ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. https://www.biorxiv.org/content/10.1101/2022.04.06.487388v2.full-text ↩

4. https://elifesciences.org/articles/84296 ↩