Speaker
Description
Our work identifies the sources of 11 interconnected machine learning (ML) biases that hinder the generalisation of supervised learning models in the context of gravitational wave (GW) detection. We use GW domain knowledge to propose a set of mitigation tactics and training strategies for ML algorithms that aim to address these biases concurrently and improve detection sensitivity. We empirically prove that our approach leads to [i] detection sensitivities that rival current matched filtering pipelines in real noise at low false alarm rates, [ii] the ability to handle out-of-distribution noise power spectral densities, [iii] the ability to strongly reject non-Gaussian transient noise artefacts, and [iv] data efficient learning of the detection problem.
Via the injection study introduced in the Machine Learning Gravitational-Wave Search Challenge, we show that our search pipeline (Sage) detects ~11.2% more signals than the benchmark PyCBC search at a false alarm rate of 1 per month and >48% than previous machine learning based detection pipelines. In light of the identified biases, we demonstrate that existing detection sensitivity metrics are unreliable for machine-learning pipelines and discuss the trustworthiness of ML results. By studying machine-learning biases and conducting empirical investigations to understand the reasons for performance improvement/degradation, we aim to address the need for interpretability of machine-learning methods in GW science. Link to the paper
| AI keywords | AI Interpretability; Machine Learning Biases; Data Efficient Learning; Detection; |
|---|
