Scalable Variational Gaussian Processes for Crowdsourcing: Glitch Detection in LIGO

Area of Science:

• Machine Learning
• Astrophysics
• Data Science

Background:

• Crowdsourcing is revolutionizing data labeling for tasks like gravitational-wave detection.
• Traditional Gaussian Processes (GPs) struggle with large datasets common in real-world applications such as LIGO.
• Deep learning methods offer scalability but often compromise on uncertainty quantification.

Purpose of the Study:

• To develop a scalable Gaussian Process (GP) based crowdsourcing method for improved uncertainty quantification.
• To address the limitations of existing methods in handling large datasets and heterogeneous annotator expertise.
• To enhance the accuracy of glitch detection systems for gravitational-wave observatories.

Main Methods:

• Leveraged sparse GP approximation (SVGP) to create a mini-batch factorized GP model.
• Developed a novel method named Scalable Variational Gaussian Processes for Crowdsourcing (SVGPCR).
• Adapted recent GP inference techniques for crowdsourcing applications and evaluated them experimentally.

Main Results:

• SVGPCR demonstrated state-of-the-art performance in crowdsourcing tasks, particularly for large datasets.
• The method significantly outperformed deep learning and previous probabilistic approaches in analyzing LIGO data.
• SVGPCR excels in uncertainty quantification, a critical factor for astrophysical data analysis.

Conclusions:

• SVGPCR effectively scales GP-based methods to previously prohibitive dataset sizes.
• The proposed method offers a superior balance between scalability and accurate uncertainty estimation.
• SVGPCR represents a significant advancement for probabilistic crowdsourcing in scientific applications like gravitational-wave astronomy.