Research map / Probabilistic and scientific modeling / Dynamical systems
Flow Matching for Few-Trial Neural Adaptation with Stable Latent Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FDA, a flow-matching framework for BCI neural decoding that learns stable latent dynamics in pre-training and then aligns sessions with fewer than five target trials via source-free likelihood maximization or MMD.
Problem
Neural decoders trained on one day degrade on later days because neural signals are nonstationary; existing alignment methods (NoMAD, Cycle-GAN, ERDiff) become unstable or perform poorly when only a few target trials are available.
Approach
Conditional flow matching on short-term neural context windows learns a flexible latent representation with provably stable (Lyapunov-based) dynamics, and alignment is done by fine-tuning only the feature extractor to maximize target likelihood (source-free) or minimize MMD, enabling few-trial adaptation.
Key results
- With target ratio r=0.02 (≤5 trials), FDA-MMD reaches average R2 of 45.59% on CO-M and 42.08% on RT-M, versus NoMAD and Cycle-GAN which perform much worse in few-trial settings.
- Nearly all maximum Lyapunov exponents of FDA are non-positive (CO-M 55/55, RT-M 52/55), indicating greater latent-dynamics stability than ERDiff and NoMAD.
- Ablating activation functions or scale coefficients (FDA-al, FDA-sc) produces frequent positive MLEs and substantially degraded transfer, e.g. CO-M r=0 R2 of -9.34% and -18.20%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FDA-MMD instead of NoMAD or Cycle-GAN | cross-session neural decoding · CO-M (r=0.02) | R2 (%): −121.47±77.80 -> 45.59±5.15; −126.84±23.82 -> 45.59±5.15 | Helped |
| use FDA-MMD instead of NoMAD or Cycle-GAN | cross-session neural decoding · RT-M (r=0.02) | R2 (%): −74.06±49.94 -> 42.08±6.31; −3.42±5.55 -> 42.08±6.31 | Helped |
| add activation functions (ablation) | FDA latent dynamics · CO-M cross-session (r=0.02) | R2 (%): 14.51±16.37 -> 45.59±5.15 | Helped |
| add scale coefficients (ablation) | FDA latent dynamics · CO-M cross-session (r=0.02) | R2 (%): 13.35±19.01 -> 45.59±5.15 | Helped |
| add activation functions (ablation) | FDA latent dynamics · RT-M cross-session (r=0.02) | R2 (%): 1.23±4.76 -> 42.08±6.31 | Helped |
| add scale coefficients (ablation) | FDA latent dynamics · RT-M cross-session (r=0.02) | R2 (%): 1.78±3.89 -> 42.08±6.31 | Helped |
| retain flow-based learning rather than using transformer features alone (ablation) | FDA · CO-M cross-session (r=0.02) | R2 (%): −82.67±90.02 -> 45.59±5.16 | Helped |
| retain flow-based learning rather than using transformer features alone (ablation) | FDA · RT-M cross-session (r=0.02) | R2 (%): 0.15±23.78 -> 42.08±6.31 | Helped |
| replace flow alignment with Cycle-GAN alignment (ablation) | FDA · CO-M cross-session (r=0.02) | R2 (%): 45.59±5.16 -> 35.57±6.46 | Hurt |
| align conditional feature distribution c instead of final latent state z(1) (ablation) | FDA · CO-M cross-session (r=0.02) | R2 (%): 45.59±5.16 -> 42.88±4.73 | Hurt |
| align conditional feature distribution c instead of final latent state z(1) (ablation) | FDA · RT-M cross-session (r=0.02) | R2 (%): 42.08±6.31 -> 42.28±6.29 | No clear effect |
| add a reconstruction loss term (ablation) | FDA · CO-M and RT-M cross-session | R2 (%) | No clear effect |
Surprising findings
- Cycle-GAN and NoMAD perform far worse with few target trials than in their original papers, reaching performance comparable to FDA only at r≈0.3 (~60 trials).
- Zero-shot NDT-2 performance degrades substantially across sessions (inter-session R2 of -34.70% on CO-C and -52.31% on CO-M).
Limitations
- Target sessions come from the same subject with substantial overlap in recorded neuronal populations; cross-task and cross-subject generalization is untested.
- Results are limited to non-human primate motor cortex recordings; clinical human data and Neuropixels are not evaluated.
- FDA-MLA performs worse than FDA-MMD and shows abnormal NLL increases on some sessions, with outlier-driven instability in Monkey C data.
Experimental setup
Datasets: CO-C CO-M RT-M
Benchmarks: CO-C CO-M RT-M Lorenz synthetic spiking data
Metrics: R2 maximum Lyapunov exponent negative log-likelihood
Baselines: LSTM CEBRA ERDiff NoMAD Cycle-GAN NDT-2
Builds on: flow matching ERDiff NoMAD Cycle-GAN CEBRA
Closest papers by meaning
- Extracting Semantic-Dynamic Features for Long-Term Stable Brain Computer Interface (AAAI 2023 · similarity 0.88)
- B[FM]$^2$: Brain Foundation Model via Flow Matching with SplitUNet (NeurIPS 2026 · similarity 0.86)
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.84)
- Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark (NeurIPS 2024 · similarity 0.84)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.83)
- Trajectory Flow Matching with Applications to Clinical Time Series Modelling (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.