Bio-PM: Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Summary

UMass Amherst / Google Health / UMD team argues that wearable accelerometer SSL has been leaving performance on the table by treating motion as an unstructured time series and windowing it arbitrarily. They ground a new tokenization in the submovement theory of motor control: continuous wrist motion decomposes into bell-shaped kinematic primitives ("submovements"), and the practically extractable unit — the "movement segment" bounded by acceleration zero-crossings — functions like a word made of phoneme-like submovements. Code and pretrained weights are released.