LSM-2: Learning from Incomplete Wearable Sensor Data

Summary

LSM-2 is the second-generation "Large Sensor Model" from Google Research (Google Research, Google Health, Google DeepMind, plus UIUC/UW interns; co-first authors Maxwell A. Xu and Girish Narayanswamy; co-last authors Xin Liu and Daniel McDuff). It introduces Adaptive and Inherited Masking (AIM), a masked-autoencoder (MAE) variant designed to learn directly from incomplete, day-long wearable sensor recordings without any imputation step, either during pretraining or at inference.