Bio-Inspired Self-Supervised Learning for Wrist-worn IMU Signals
Published in arXiv preprint, 2026
Wearable accelerometers have enabled large-scale health and wellness monitoring, yet learning robust human-activity representations has been constrained by the scarcity of labeled data. While self-supervised learning offers a potential remedy, existing approaches treat sensor streams as unstructured time series, overlooking the underlying biological structure of human movement.
We introduce a novel tokenization strategy grounded in the submovement theory of motor control, which posits that continuous wrist motion is composed of superposed elementary basis functions called submovements. By treating movement segments as tokens, we pretrain a Transformer encoder via masked movement-segment reconstruction to model temporal dependencies — shifting the learning focus beyond local waveform morphology.
Pretrained on the NHANES corpus (~28k hours; ~11k participants; ~10M windows), our representations outperform strong wearable SSL baselines across six subject-disjoint HAR benchmarks and demonstrate stronger data efficiency in data-scarce settings.
Recommended citation: Tarale, P., Chu, K., Varghese, A., Liu, K., Xu, M. A., Iyyer, M., & Lee, S. I. (2026). "Bio-Inspired Self-Supervised Learning for Wrist-worn IMU Signals." arXiv:2603.10961.
Download Paper