SimPer: Simple Self-Supervised Learning of Periodic Targets

Summary

SimPer introduces a self-supervised contrastive learning framework tailored for signals with periodic structure — specifically heart rate from PPG/rPPG, respiratory rate from video, and pedometer signals. The key insight is that periodic targets are fundamentally different from categorical labels: neighboring frequencies are semantically similar, so standard SimCLR-style augmentation misses the continuity of the label space. SimPer defines "period-based" augmentations (speed perturbation that changes the signal's dominant frequency in a controlled way) and a generalised contrastive loss that treats augmented views with close frequencies as positives and distant ones as negatives.