Speaker
Description
This project builds on DeepClean, a machine learning framework for subtracting instrumental and environmental noise from gravitational-wave detector data using auxiliary “witness” channels. While effective, DeepClean’s convolutional architecture limits its ability to scale across varying channel configurations and frequency bands.
We propose a hybrid CNN–transformer architecture that maintains permutation invariance across channels. The model combines a per-channel CNN for independent feature extraction with a transformer that treats time steps as tokens and models cross-channel interactions via self-attention, while preserving channel-specific representations. This design aims to reduce retraining requirements under changing detector configurations, such as varying channel sets or frequency bands.
Preliminary results show performance comparable to DeepClean on standard denoising benchmarks, indicating that this added flexibility does not compromise accuracy. Ongoing work explores improved performance across broader frequency bands and increased robustness to imperfect or noisy channel selection. Ultimately, this project aims to move toward a more automated version of DeepClean, incorporating learning-based channel selection tailored to the denoising task.