Dynamical Latent Flow Matching for High-Performance Few-Shot Neural Adaptation

ICLR 2026 Conference Submission15592 Authors

Keywords: Dynamical Latent Space, Flow Matching, Neural Adaptation, Brain-Computer Interface

Abstract:

The primary objective of brain–computer interfaces (BCIs) is to establish a direct connection between neural activity and external devices. However, variability in neural recordings poses significant challenges to maintaining stable neural decoding with minimal recalibration.
Existing neural decoding frameworks often fail to enable efficient few-shot adaptation, typically due to the constraints imposed by prior assumptions on latent variables or issues with training instability. Motivated by the flexibility and tractability of diffusion models, we propose the novel Dynamical Latent Flow Matching (DLFM) framework for high-performance few-shot neural adaptation. Our DLFM performs flow matching in dynamical latent spaces, leveraging preserved neural dynamics within the neural manifold. The probabilistic flexibility of DLFM effectively captures intrinsic features of dynamical patterns across heterogeneous sessions, significantly enhancing few-shot neural adaptation.
The efficiency of DLFM for few-shot adaptation is validated on the Falcon benchmark, achieving competitive performance with only 60\% calibration trials.
Further experimental evaluation on the Neural Latents Benchmark 2021 demonstrates that DLFM ranks among the top two for forward prediction tasks across all web submissions.
Additional interpretability analysis on the Lorenz attractor model and the Falcon dataset confirms that DLFM precisely identifies the intrinsic features of neural dynamics, thus facilitating efficient few-shot neural adaptation for neural decoding.
Our DLFM framework emerges as a promising candidate for superior few-shot neural adaptation, advancing the practicality of real-world BCI systems.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 15592