Research map / Probabilistic and scientific modeling / Dynamical systems

Quasi-Linear ICA for Motor Unit Decomposition during Dynamic Contractions

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Synaptic plasticity learning rules · 1 citations (OpenAlex)

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 HarmonICA, a quasi-linear ICA method that pairs a static linear separator with a learned low-rank time-varying Woodbury compensator and per-spike neural peel-off to decompose dynamic HD-EMG into motor unit spike trains.

Problem

Decomposing surface EMG into individual motor neuron spike trains fails during movement because volume-conductor deformation and electrode shift make the neuron-to-electrode mixing time-varying, which violates the stationarity assumption of linear ICA and convolutive BSS.

Approach

Keep source extraction linear (identifiable via classical ICA) and place a learned low-rank time-varying invertible Woodbury transform in front of it, trained with a stationarity (GMM cluster-spread) loss with decoupled gradients; the closed-form inverse enables per-spike time-varying template subtraction during peel-off.

Key results

• At recall >= 0.95 HarmonICA recovers 2.5x (320-channel) and 2.2x (70-channel) as many motor units as the unrefined FastICA seed.
• At F1 >= 0.95 on ch320, HarmonICA (PE) recovers 515 MUs vs 249 for ICA/cBSS and 269 for the strongest adaptive baseline (Chen 2020); angle-conditioned variant reaches 571.
• Ablation: replacing the Woodbury compensator with an additive correction drops yield by 39%/41%; kurtosis loss instead of GMM drops yield by 78%/81%.

Surprising findings

• Kurtosis-based contrast in the stationarity objective collapses yield by roughly 80%, indicating outlier-driven solutions.
• Within-pool MUAP near-duplicates are so common in the simulator that at T=0.75 only 13.5% of motor units survive a crosstalk filter, so raw yield ceilings are far below nominal pool size.

Limitations

• Benchmark is fully simulated; no paired real-recording validation.
• Offline by construction since positional features span the full recording window.
• Recovery bounded by upstream FastICA seed; high within-pool crosstalk can cause false positives.

Experimental setup

Datasets: muniverse-benchmark-dynamic MUniverse NeuroMotion BioMime

Benchmarks: muniverse-benchmark-dynamic

Metrics: F1 yield at F1 >= 0.95 precision recall

Baselines: ICA / cBSS (FastICA seed) Chen 2020 Kramberger 2021 Yeung 2024 Glaser 2018 Mendez Guerra 2024

Builds on: FastICA Convolution Kernel Compensation Woodbury identity layer (normalizing flow) NeRF positional encoding MUniverse

Closest papers by meaning

• SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.85)
• Learning Cortico-Muscular Dependence through Orthonormal Decomposition of Density Ratios (NeurIPS 2024 · similarity 0.84)
• EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography (NeurIPS 2024 · similarity 0.84)
• Learning Options for Compositional Motor Control with Adapter Banks (NeurIPS 2026 · similarity 0.84)
• Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark (NeurIPS 2024 · similarity 0.84)
• emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (NeurIPS 2024 · similarity 0.83)