Project objective

This project investigates why an EMG gesture classifier succeeds or fails when the acquisition day changes. High predictive accuracy on a familiar split is not enough: the explanation should remain stable, localise plausible temporal and channel-level evidence, and avoid relying on acquisition artefacts.

Current methodology

The work combines a compact neural classifier with a case-based explanation audit. Rather than averaging explanations across the entire test set, we inspect distinct behavioural regimes:

• canonical correct predictions;
• borderline confusions;
• persistent hard cases;
• representative errors under cross-day shift.

Attribution methods, including Integrated Gradients, are compared with temporal attention and signal context. Positive and negative contributions are kept separate so that an explanation does not hide cancellation effects behind a single absolute-importance score.

Research question

The central question is not merely whether an attribution map is visually smooth. It is whether the explanation identifies signal regions that are physiologically plausible, stable across days, and causally relevant to the model’s decision, rather than unstable background or preprocessing artefacts.

Status

This is ongoing work with Anany. Results are treated as preliminary until the cross-day evaluation, explanation stability tests, and case-selection protocol are fully validated.