Incremental Learning for Cross-User sEMG Gesture Recognition: Overcoming User Forgetting

Surface Electromyography (sEMG) based gesture recognition has gained significant attention in the fields of medical rehabilitation and entertainment. However, sEMG exhibits strong user-dependent properties among users with different physiological states, limiting the generalization and personalization of existing models. Although incremental learning methods can help the model balance the knowledge learned from old and new users, the existing incremental learning literature lacks attempts at sEMG gesture recognition. This paper fills these gaps and proposes the incremental learning-based cross-user sEMG gesture recognition framework (IncreEGR), to explore the catastrophic forgetting challenges in cross-user sEMG data. IncreEGR consists of two parts: a sampling method based on the center of features (CoF) and discrepancy-based distillation learning (DDL). CoF samples and saves representative user data and replays them during subsequent training processes to ensure that the model can review previous knowledge. DDL aligns the features extracted from the old and new models using a Gaussian kernel-based distance constraint to help the model overcome catastrophic forgetting. We conduct experiments on the benchmark sEMG gesture recognition dataset (i.e., the Ninapro dataset) to verify the effectiveness of IncreEGR. The experimental results demonstrate that our proposed IncreEGR method successfully balances the knowledge between new and old users and achieves the best recognition performance.