Dynamic Distillation Boosts Single-IMU Activity Recognition Performance.

This research explores how to enhance activity recognition using a single Inertial Measurement Unit (IMU) at inference time, even when multiple IMUs are available during training. The core challenge is to transfer the rich information from a multi-sensor setup to a more constrained single-sensor deployment without increasing the inference burden. The paper proposes Dynamic Influence Weighting (DIW), a novel knowledge distillation technique. DIW improves upon traditional fixed-weight knowledge distillation, which applies the same strength to all training samples, regardless of their individual benefit to the student model. Instead, DIW dynamically assigns separate sample-wise gates to logit and feature losses, testing candidate updates on internal training participants to determine optimal weighting. This allows the single-IMU student model to selectively learn more effectively from the multi-IMU teacher. Evaluated on the WEAR dataset for 19 activity labels, DIW achieved a macro-F1 score of 0.638451, representing significant gains of 7.66 and 6.68 percentage points over supervised learning and fixed-weight knowledge distillation, respectively. Crucially, these improvements were realized without any changes to the deployed single-IMU sensor or the student model's forward graph, making it a highly practical advancement for resource-constrained applications.