Every muscle is driven by motor units, each a single spinal neuron plus the fibers it controls. A grid of electrodes on the skin records the summed electrical activity of dozens of them, and "decomposition" recovers each neuron's individual firing times from that mixture. Current decomposition tools work offline on steady, held contractions. When a person moves, sweats or tires, the mixture changes. An online decomposer then has to keep updating its model of each unit, and it can only do so from its own detections. Its mistakes become its training labels. A unit's filter slowly slides toward a neighboring unit or toward noise, and nothing inside the loop notices, because the drifted filter still reports confident spikes.
We propose to break this loop with a second, independent opinion. We train a small detector that answers only "is a motor unit action potential happening here, now, on these electrodes?" and never "which unit is it?". This is the same split that open-world object detection makes between finding objects and naming them. The per-unit trackers may update only on spikes that this detector confirms in the right place on the grid. Detector events that no tracked unit explains are treated as evidence of a newly recruited unit, not as drift. The bet is that a detector blind to unit identity makes errors that are largely uncorrelated with the trackers' errors. If so, it can act as the missing verification signal that continual test-time adaptation lacks.
Core Claim
In self-updating decomposition, error accumulation is self-confirming. Take a separation vector, the per-unit linear filter that turns the multichannel recording into one unit's pulse train. Its spikes are both its prediction and its label. An identity-free event detector, trained only on simulated physics of how action potentials spread across the grid, gives a second view. That view can fail, but it fails differently. We claim three things. Gating tracker updates on agreement with this detector extends how long each unit is tracked correctly under non-stationary conditions, compared with self-confidence gating. Residual, unexplained proposals let the system tell recruitment apart from drift. Both effects come from the identity-free design, and not merely from having a learned detector.
The core mechanism comes from the decomposition literature. The application regime is online adaptation without labels. Class-agnostic proposal is the lens that supplies the independent check.
Prior Work and Where This Sits
The offline decomposition pipeline we build on is formalized in Tutorial on MUedit. MUedit whitens and time-extends the signal, optimizes sparse separation vectors, and lets an expert edit pulse trains, after which it recomputes each vector from the corrected spikes. That recompute step is exactly the update we automate. The paper states that the method assumes stationary contractions and needs a human in the loop. Our gate is meant to replace the human editor. Two other methods improve offline yield and accuracy. Swarm-Contrastive Decomposition tunes the contrast function per unit with particle swarm search and peels off found units. EMUsort adapts template spike sorting to intramuscular arrays. Both are batch methods with fixed templates. They motivate our work by showing that superposition and dominant-unit bias remain the main error sources. EMUsort's accuracy falls from 94.9% to 85.1% on overlapping potentials in simulated rat data. On the learned side, Toward a generalizable deep CNN for neural drive estimation regresses the pooled cumulative spike train across sessions 20 months apart. It never adapts online and gives up on single units, which is precisely the resolution we want to keep.
Class-agnostic proposal supplies the method. Deep learning for class-generic object detection showed that localization learned without category labels transfers to held-out categories. OpenInst showed that localization-only queries separate objects from clutter without a classification head. Exploring Orthogonality in Open World Object Detection showed that decorrelating "is there an object" from "which class" raises recall of unknown objects. Our detector is the electrode-grid analogue. Its "objects" are spatio-temporal footprints, which Muceli et al. found to be highly localized, decaying below 13% of peak within millimeters for most units. Proposal networks have moved to one-dimensional streams before, as in Speaker Diarization with Region Proposal Network, but not into convolutive neural mixtures. RandBox constrains our design. Its random proposals avoid training bias, but they need 10,000 boxes per image at inference, which cannot work at 2,048 Hz.
The continual test-time adaptation literature both motivates and constrains the design. PCoTTA anchors adaptation to source prototypes. DPLOT restricts which layers update and builds pseudo-labels from paired views. DynaPrompt documents collapse after a single confident error. Sequence-level training for tracking shows that models never exposed to their own errors cannot recover from them. All of these rely on statistical self-consistency within one model. Test-time adaptation for open-vocabulary segmentation extends adaptation to open-set images, but still has no channel that separates drift from genuinely new sources. Our contribution is that channel, built for a domain where physics makes an identity-free check credible.
Research Program
The system starts from a short stationary calibration segment. Running the MUedit or Swarm-Contrastive pipeline on it gives $N$ initial separation vectors. For each unit it also stores a prototype: a spike-triggered average waveform and its electrode footprint, meaning the set of channels where the waveform exceeds a fixed fraction of its peak.
The proposal detector is a lightweight causal convolutional network that runs over a sliding window of the grid. It outputs, for each time step, a heatmap over electrode positions together with a footprint extent. It is trained entirely on simulated recordings from the NeuroMotion and MUsim simulators, sampled across many anatomies, fat thicknesses, electrode shifts and force levels. The labels say only "some unit fired at time $t$ with footprint $F$" and never which unit. To mirror the lesson of sequence-level training, training includes heavy superposition and injected artifacts such as powerline hum, motion bursts and baseline wander as negatives. At deployment the detector stays frozen apart from normalization statistics. Its target, MUAP-ness, does not change when units are recruited or drift, so freezing it is what keeps its errors from coupling to the trackers' errors.
Each unit's tracker updates its vector online by recursive re-whitening and spike-triggered re-estimation over a sliding buffer. The update uses only gated spikes. Let $S_k$ be the spikes detected by unit $k$'s current vector, $P$ the detector's proposals, each with time $t_p$ and footprint $F_p$, and $F_k$ the unit's current prototype footprint. The gated set is
$$\tilde S_k = \{\, t \in S_k : \exists\, p \in P,\ |t - t_p| \le \delta,\ \mathrm{IoU}(F_p, F_k) \ge \tau \,\}.$$
Here $\delta$ is a timing tolerance of a few milliseconds and $\tau$ is an overlap threshold. The prototype footprint itself moves by a slow exponential average, and the system limits how far it may move from its calibration value, in the spirit of PCoTTA's source anchoring.
Discovery runs on the residual stream. Proposals claimed by no unit are accumulated and clustered by footprint. A new unit is created from a cluster's spike-triggered average when the cluster shows a stable footprint and physiological inter-spike intervals. This replaces the random-restart global search of classical methods with search seeded by unexplained events.
Baselines
The naive reference is MUedit's offline decomposition from the calibration segment, frozen for the whole recording. It shows how fast stationary filters decay under the conditions we create. Its stationary accuracy anchors the ceiling: MUedit reports about 95% rate of agreement on simulated surface EMG and 94% agreement against intramuscular units in tibialis anterior (Tutorial on MUedit).
The strongest practical competitor is the same pipeline with periodic offline recalibration. It is re-run every $T$ seconds, with $T$ swept, and the new units are matched to the previous ones. It is strong because it uses the most accurate offline method available. We give it the same signal, with the computation used for recalibration counted separately, and we also report it with oracle recalibration timing.
The strongest adaptive competitor is ungated online re-estimation, where MUedit's recompute step runs automatically on every detected spike. The most important ablation keeps the same online tracker but gates on each unit's own pulse-train confidence, in the style of the silhouette measure. That isolates whether an independent view matters, or merely selective updating. A second ablation replaces the learned detector with a multichannel energy-burst detector. A third gives the gate oracle ground-truth events, which sets an upper bound. For single-unit online tracking under dynamic conditions, no reported result was found on Lacuna, so all adaptive numbers will be produced here under one shared tuning budget. Each method gets the same grid search over its two or three thresholds on held-out simulated sequences.
Beating these baselines means two things. First, a higher median unit-tracking survival time, the time until a unit's rate of agreement drops below 80%. Second, a mean rate of agreement over the recording that matches recalibration at its best $T$, without the pauses.
Why not just…?
Why not just recalibrate offline every few minutes? This is a real competitor for slow drift such as fatigue. The deep CNN paper describes repeated recalibration as impractical for users. That is a usability argument, though, not evidence of accuracy. Our sweep over $T$ settles the question. If recalibration every 60 seconds matches our survival times on fatigue sequences, we will report that the gate only helps under faster non-stationarity, namely movement and electrode shift.
Why not just gate on the unit's own confidence? This is the most serious threat, because it is cheap and already standard in decomposition. Our hypothesis is that a vector drifting toward a neighboring unit keeps producing a sparse, confident pulse train, so self-confidence cannot see the failure. The self-confidence ablation tests this directly. So does a measurement of the correlation between tracker false spikes and detector false events. If that correlation is high, or the self-gated version matches ours, the independence claim fails.
Why not just threshold multichannel energy instead of learning a detector? At low force, energy bursts may suffice. At high force, the dense interference that Swarm-Contrastive Decomposition targets makes nearly every window energetic. The energy-detector ablation, stratified by force level, shows where the learned detector earns its cost.
Why not just apply entropy-minimization test-time adaptation or train an end-to-end decoder? PCoTTA reports that unanchored continual adaptation reaches a Chamfer distance of 79.7, against 6.3 with anchoring. That suggests unanchored self-training is the failure mode we are studying, not a cure for it. End-to-end decoders such as emg2qwerty report 51.8% character error rate zero-shot on new users, and they do not resolve single units, which clinical uses like circuit mapping require.
Proposed Experiments
The first test is controlled drift in simulation. NeuroMotion and MUsim sequences are generated with three kinds of change, each applied separately and then combined: gradual waveform change mimicking fatigue, electrode-grid translation of 0 to 10 millimeters over a minute, and force ramps that recruit new units. All methods are scored on per-unit rate of agreement over time, tracking survival time, false merges where one vector switches to another unit, precision and latency of new-unit discovery, and compute per second of signal.
The decisive check is a crossed design in which recruitment and electrode shift happen together. Here a self-referential system cannot tell a new source from a moved old one. Our claim predicts that residual-proposal discovery creates new units, and that gated trackers do not absorb those units. Failure looks like new units being merged into drifting old ones at the same rate as in the self-gated ablation.
The test on real data uses dual surface and intramuscular recordings of tibialis anterior and forearm, following the protocols of MUedit and Swarm-Contrastive Decomposition. They are extended to slow dynamic ankle movements and sustained fatiguing contractions, with intramuscular units as partial ground truth. Because the detector is trained only on simulation, this is also where the simulation-to-reality gap will show. We will report the detector's own event precision and recall against intramuscular spikes. That way a failure can be traced either to the detector or to the gating logic.
The emg2qwerty dataset, with its low-density 32 channels, is too underdetermined for single-unit claims. We will use it only to measure the detector's false-proposal rate on real wrist artifacts.
References
- Tutorial on MUedit: An open-source software for identifying and analysing the discharge timing of motor units from electromyographic signals
- Unlocking the full potential of high‐density surface EMG: novel non‐invasive high‐yield motor unit decomposition
- High performance sorting of motor unit action potentials with EMUsort
- Toward a generalizable deep CNN for neural drive estimation across muscles and participants
- Deep learning for class-generic object detection
- OpenInst: A simple query-based method for open-world instance segmentation
- Exploring Orthogonality in Open World Object Detection
- Decoding motor neuron activity from epimysial thin-film electrode recordings following targeted muscle reinnervation
- Speaker Diarization with Region Proposal Network
- Random Boxes Are Open-world Object Detectors
- PCoTTA: Continual Test-Time Adaptation for Multi-Task Point Cloud Understanding
- Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation
- DynaPrompt: Dynamic Test-Time Prompt Tuning
- Towards Sequence-Level Training for Visual Tracking
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation
- emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography