Cross-session and cross-subject transfer

"I scored 98 percent yesterday. Then you took the armband off, slept, put it back on two millimeters to the left, and now I think every gesture is a fist. I did not forget your muscles. I memorized your electrodes."

An Overfitted AI Agent

This section builds directly on the feature representations of Section 32.2 and the classifiers of Section 32.3. It is, at heart, a domain-shift problem, so it leans on the distribution-shift and test-time adaptation machinery of Chapter 66 and the leakage-safe splitting discipline of Chapter 5. Both are load-bearing here. We stay on transfer; the underlying decoding tasks live elsewhere in the chapter.

What actually shifts, and why

Frame the problem as covariate shift: the mapping from muscle intent to gesture label is roughly stable, but the distribution of the input features \(p(x)\) drifts between the training domain and the deployment domain while the conditional \(p(y \mid x)\) we want stays fixed. Three drivers dominate. Electrode shift is geometric: donning an sEMG array a few millimeters rotated or translated changes which motor units sit under each channel, permuting and rescaling the channel pattern for a given gesture. Physiological session drift is electrical: overnight changes in skin-electrode impedance, hydration, temperature, sweat, and the slow onset of fatigue (Section 32.4) all scale amplitudes and tilt spectra even for the same electrode placement. Cross-subject variation is anatomical and the harshest of the three: subcutaneous fat thickness attenuates and blurs the signal, muscle size and fiber orientation set the baseline amplitude, and limb circumference changes the spacing between a fixed array and the active tissue. A model that latched onto one person's absolute channel amplitudes has learned a fingerprint, not a physiology.

The cheap first line: normalization and calibration

Before any learned adaptation, remove the shifts you can remove by construction. The most effective single step is amplitude normalization: rather than feeding raw envelope magnitudes, normalize each channel so that its scale is comparable across sessions and people. Classic clinical practice normalizes to a maximum voluntary contraction (MVC), but a lighter and label-free alternative is to standardize each channel by its per-session mean and standard deviation, or by the RMS of a short reference contraction. This directly cancels the multiplicative part of both session drift and cross-subject amplitude difference. Pair it with features that are already scale-robust or that emphasize spectral shape over absolute power (the frequency-domain descriptors of Section 32.2 and the reductions of Chapter 8).

When normalization is not enough, add a short per-session calibration: ask the user to perform each gesture briefly at setup, then update the decoder. Even a few seconds per class lets you refit a lightweight classifier head or recenter the feature space, which is why nearly every commercial myoelectric armband ships with a guided calibration ritual. The design tension is user burden: every second of calibration is friction, so the research goal is to push calibration toward zero while keeping cross-session accuracy high.

import numpy as np

def per_channel_standardize(X, stats=None):
    """Cancel multiplicative session/subject amplitude shift.
    X: (n_windows, n_channels) feature magnitudes for one session."""
    if stats is None:                       # fit on this session (calibration data)
        mu = X.mean(axis=0)
        sd = X.std(axis=0) + 1e-8
        stats = (mu, sd)
    mu, sd = stats
    return (X - mu) / sd, stats             # reuse stats on later windows

def coral_align(Xs, Xt):
    """CORAL: recolor source features to match target covariance.
    Xs = source (training subjects), Xt = target (new subject) features."""
    Cs = np.cov(Xs, rowvar=False) + np.eye(Xs.shape[1])
    Ct = np.cov(Xt, rowvar=False) + np.eye(Xt.shape[1])
    Ws = np.linalg.cholesky(np.linalg.inv(Cs))     # whiten source
    Wt = np.linalg.cholesky(Ct)                    # recolor to target
    return (Xs - Xs.mean(0)) @ Ws @ Wt.T + Xt.mean(0)

The coral_align function above is the bridge from hand normalization to genuine domain adaptation: it needs no labels from the new user, only a handful of unlabeled windows, and it aligns second-order statistics rather than just per-channel scale.

Learned transfer: adaptation and pretraining

When cheap alignment plateaus, three learned strategies climb higher. Domain adaptation makes the model's internal features invariant to which session or subject produced them: adversarial methods (a domain-adversarial network, DANN) add a discriminator that tries to guess the source domain from the features while the encoder is trained to fool it, so the surviving representation encodes gesture but not identity; statistical methods (CORAL, maximum-mean-discrepancy alignment) match feature distributions directly. These need unlabeled target data but no target labels, which fits the deployment reality where a new user will not hand-label their own gestures. Transfer learning pretrains a decoder on many subjects, then fine-tunes on a few calibration examples from the target: the shared backbone learns cross-subject physiology and the small head absorbs the individual's idiosyncrasy, exactly the recipe that made rare-diagnosis ECG models work in Chapter 29. Self-supervised pretraining (Chapter 17) goes further, learning representations from large pools of unlabeled EMG so that a new user needs only a linear probe. A complementary trick specific to array EMG is electrode-shift augmentation: synthetically rotate or translate the channel grid during training so the model sees many virtual placements and stops depending on any one. At inference, the streaming test-time adaptation of Chapter 66 can keep recentering the decoder as the session drifts under fatigue.

Evaluating transfer without fooling yourself

None of the above is trustworthy without the right split. The gold standard for cross-subject claims is leave-one-subject-out (LOSO): hold every window from one subject entirely out of training, evaluate on that subject, rotate through all subjects, and report the distribution not just the mean. For cross-session claims, split by recording session so no window from a test session ever appears in training, mirroring the re-donning that happens in the field. The cardinal sin, splitting windows randomly, lets adjacent overlapping windows and shared electrode placement leak across the boundary and inflates accuracy by tens of points; it is the EMG face of the leakage discipline in Chapter 5 and the benchmarking protocols of Chapter 65. Report calibration cost alongside accuracy: a decoder that needs thirty seconds of per-class calibration to reach 90 percent is a different product from one that reaches 85 percent with zero calibration, and the tradeoff is the design decision.

What's Next

In Section 32.6, we confront the noise that transfer cannot fix by normalization alone: motion artifacts, power-line interference, ECG contamination of trunk EMG, and crosstalk between neighboring muscles. Clean signals make every transfer method in this section work harder for you, so we turn to detecting and suppressing the artifacts that corrupt EMG at the source.