Neural Manifold Regularization: Aligning 2D Latent Dynamics with Stereotyped, Natural, and Attempted Movements

ICLR 2025 Conference Submission1722 Authors

Keywords: dimensionality reduction, brain-machine interfaces, motor control, neural coding, self-supervised learning

TL;DR: A dimensionality reduction method that maps high-dimensional neural dynamics to two-dimensional, behavior-aligned latent dynamics.

Abstract:

Mapping neural activity to behavior is a fundamental goal in both neuroscience and brain-machine interfaces. Traditionally, at least three-dimensional (3D) latent dynamics have been required to represent two-dimensional (2D) movement trajectories. In this work, we introduce Neural Manifold Regularization (NMR), a method that embeds neural dynamics into a 2D latent space and regularizes the manifold based on the distances and densities of continuous movement labels. NMR pulls together positive pairs of neural embeddings (corresponding to closer labels) and pushes apart negative pairs (representing more distant labels). Additionally, NMR applies greater force to infrequent labels to prevent them from collapsing into dominant labels.
We evaluated NMR across four modalities of neural signals and three types of movements. When combined with a linear regression decoder, NMR outperformed other dimensionality reduction methods by over 50\% across 68 sessions. The highly consistent neural manifolds extracted by NMR enable robust motor decoding across sessions, years, and subjects using a simple linear regression decoder.
Our code is uploaded.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1722