Active Learning approach to Gravitational Waves Classification Algorithms

Abstract

This project explores the integration of Bayesian Optimization (BO) algorithms into a base machine learning model, specifically Convolutional Neural Networks (CNNs), for classifying gravitational waves among background noise. The primary objective is to evaluate whether optimizing hyperparameters using Bayesian Optimization enhances the performance of the base model. For this purpose, a Kaggle [1] dataset that comprises real background noise (labeled 0) and simulated gravitational wave signals with noise (labeled 1) is used. Data with real noise is collected from three detectors: LIGO Livingston, LIGO Hanford, and Virgo. Through data preprocessing and training, the models effectively classify testing data, predicting the presence of gravitational wave signals with a remarkable score, 83.61%. The BO model demonstrates comparable accuracy to the base model, but its performance improvement is not very significant (84.34%). However, it is worth noting that the BO model needs additional computational resources and time due to the iterations required for hyperparameter optimization, requiring an additional training on the entire dataset. For this reason, the BO model is less efficient in terms of resources compared to the base model in gravitational wave classification