Variable Input Masked Autoencoders for Wrist-Worn Accelerometry (VarMAE)

Anonymous (ICLR 2027 double-blind submission)

Summary

VarMAE learns a single representation across sampling rates and temporal contexts: random crop + downsample creates complementary bottlenecks, and a cross-resolution masked reconstruction loss ties the views. Pretrained on NHANES; a single 0.7M-param encoder matches or outperforms controlled FM baselines on eight benchmark datasets.