Assessing the State of Self-Supervised Human Activity Recognition Using Wearables

The emergence of self-supervised learning in the field wearables-based human activity recognition (HAR) has opened up opportunities to tackle most pressing challenges field, namely exploit unlabeled data derive reliable systems for scenarios where only small amounts labeled training samples can be collected. As such, self-supervision, i.e., paradigm 'pretrain-then-finetune' potential become a strong alternative predominant end-to-end approaches, let alone hand-crafted features classic chain. Recently number contributions have been made that introduced into HAR, including, Multi-task Masked Reconstruction, CPC, and SimCLR, name but few. With initial success these methods, time come systematic inventory analysis field. This paper provides exactly that. We assess progress HAR research by introducing framework performs multi-faceted exploration model performance. organize three dimensions, each containing constituent criteria, such dimension captures specific aspects performance, including robustness differing source target conditions, influence dataset characteristics, feature space characteristics. utilize this seven state-of-the-art methods leading formulation insights properties techniques establish their value towards representations diverse scenarios.