Title: SimPer: Self-Supervised Learning of Periodic Targets
Learning from periodic data is essential for various applications, such as weather monitoring and healthcare. Traditional approaches require labeled data, which can be challenging and time-consuming to obtain. Alternatively, self-supervised learning methods like SimPer leverage unlabeled data to capture periodic information. In this article, we introduce SimPer and discuss its features and advantages.
SimPer is a self-supervised contrastive framework for learning periodic information. It uses temporal self-contrastive learning, where positive and negative samples are obtained through periodicity-invariant and periodicity-variant augmentations from the same input. The framework includes a periodic feature similarity measure and a generalized contrastive loss for contrasting over continuous labels (frequency).
How SimPer Works:
SimPer transforms an input sequence to create different negative views by altering the speed or frequency. Pseudo speed or frequency labels are assigned to these unlabeled inputs. Periodicity-invariant augmentations are applied to create positive views of the sample. These augmented views are then sent to the encoder to extract features.
SimPer outperforms state-of-the-art self-supervised learning schemes on various real-world tasks, including heart rate measurement and exercise repetition counting from video data. It demonstrates superior performance in terms of data efficiency, robustness to spurious correlations, and generalization to unseen targets.
Conclusion and Applications:
SimPer provides an intuitive and flexible approach for learning strong feature representations of periodic signals. It can be applied to a wide range of fields, including environmental remote sensing and healthcare.
Overall, SimPer offers a promising solution for learning from periodic data without the need for extensive labeling. Its self-supervised learning framework and advanced features make it a valuable tool in the field of artificial intelligence.