Capturing Static, Short-Term, and Long-Term Dynamics Through Self-Supervised Time Series Learning: CHRONOS
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Self-Supervised Learning, Time Series, Physiological Signals.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: CHRONOS , a novel Self-supervised Learning (SSL) method , effectively captures patterns from distinct temporal natures in time series through separate spaces and contrasting heads. It consistently performs across various downstream tasks.
Abstract: Time series data presents a rich tapestry of temporal patterns, encompassing both enduring static trends that persist throughout the entire temporal sequence and dynamic patterns that define its evolving nature. To advance the field of Self-Supervised Learning (SSL) in time series analysis, it is essential to adopt a comprehensive approach that considers these distinct temporal facets. In this paper, we introduce Contrasting Heads Represent Opposed Natures of Signals (CHRONOS), a novel SSL methodology which drives the model to understand three distinct temporal attributes – static, short-term dynamics, and long-term dynamics. This is achieved by projecting the representations into two separate spaces, employing contrasting heads. Furthermore, a selective optimization leads distinct model units to be specialized in different temporal natures. To evaluate the effectiveness of CHRONOS, we applied this methodology to the analysis of electrocardiagram (ECG) signals across four distinct downstream tasks, utilizing four independent datasets. Our study demonstrates the consistent performance of CHRONOS across all tasks, surpassing state-of-the-art methods for time series data analysis. CHRONOS serves as a testament to the importance of capturing diverse temporal aspects of time series data for driving versatile models capable of consistently excelling in a wide spectrum of downstream tasks
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3585
Loading