Debiased Contrastive Learning with multi-resolution Kolmogorov-Arnold Network for Gravitational Wave Glitch Detection

Henry Han, Jie Ren

Keywords: self-supervised learning, Debiased Contrastive Learning, Gravitational Wave, Glitch detection, deep learning

TL;DR: Debiased Contrastive Learning with Kolmogorov-Arnold Network for Gravitational Wave Glitch Detection

Abstract: Time-series gravitational wave glitch detection presents significant challenges for machine learning due to the complexity of the data, limited labeled examples, and data imbalance. To address these issues, we introduce Debiased Contrastive Learning with Multi-Resolution Kolmogorov-Arnold Network(dcMltR-KAN), a novel self-supervised learning (SSL) approach that enhances glitch detection, robustness, explainability, and generalization. dcMltR-KAN consists of three key novel components: Wasserstein Debiased Contrastive Learning (wDCL), a CNN-based encoder, and a Multi-Resolution KAN (MltR-KAN). The wDCL improves the model’s sensitivity to data imbalance and geometric structure. The CNN-based encoder eliminates false negatives during training, refines feature representations through similarity-based weighting (SBW), and reduces data complexity within the embedding. Additionally, MltR-KAN enhances explainability, generalization, and efficiency by adaptively learning parameters. Our model outperforms widely used baselines on O1, O2, and O3 data, demonstrating its effectiveness. Extending dcMltR-KAN to other time-series benchmarks underscores its novelty and efficiency, marking it as the first model of its kind and paving the way for future SSL and astrophysics research.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 8626