Exploring LIGO Glitch Morphologies through Transformer Embeddings and t-SNE Visualization
ORAL
Abstract
Alongside astrophysical signals, LIGO continuously records a wide range of non-astrophysical noise transients (glitches) arising from instrumental and environmental disturbances. These glitches can mimic or obscure real signals, posing a challenge for confident detection and interpretation.
We use MIT's Audio Spectrogram Transformer (AST) for glitch analysis in gravitational-wave data. By leveraging the power of a large pre-trained transformer, AST offers a way to extract rich features without extensive labeled datasets or training from scratch. Instead of full retraining, we fine-tune the model and apply t-SNE (t-Distributed Stochastic Neighbor Embedding) to its embedding space for dimensionality reduction into a 3D representation, enabling label-agnostic clustering of glitches.
This approach will provide a path towards rapid and adaptive glitch characterization. It would make it possible to discover new glitch morphologies as emergent clusters, suggest potential substructures within established glitch classes, and reveal how glitches may evolve across observing runs. It can also provide a quantitative way to assess similarity between glitch families, offering new insights into the complex noise environment of gravitational-wave detectors. Beyond glitches, the same embedding space can cluster gravitational-wave events, opening the door to label-agnostic exploration of signal populations, such as distinguishing between high- and low-mass events.
We use MIT's Audio Spectrogram Transformer (AST) for glitch analysis in gravitational-wave data. By leveraging the power of a large pre-trained transformer, AST offers a way to extract rich features without extensive labeled datasets or training from scratch. Instead of full retraining, we fine-tune the model and apply t-SNE (t-Distributed Stochastic Neighbor Embedding) to its embedding space for dimensionality reduction into a 3D representation, enabling label-agnostic clustering of glitches.
This approach will provide a path towards rapid and adaptive glitch characterization. It would make it possible to discover new glitch morphologies as emergent clusters, suggest potential substructures within established glitch classes, and reveal how glitches may evolve across observing runs. It can also provide a quantitative way to assess similarity between glitch families, offering new insights into the complex noise environment of gravitational-wave detectors. Beyond glitches, the same embedding space can cluster gravitational-wave events, opening the door to label-agnostic exploration of signal populations, such as distinguishing between high- and low-mass events.
–
Presenters
-
Suyash Deshmukh
- Vanderbilt University