Identifying and mitigating machine-learning biases for the gravitational-wave detection problem
识别并缓解引力波探测问题中的机器学习偏差
Matched filtering is a long-standing technique for the optimal detection of known signals in stationary Gaussian noise. However, it has known departures from optimality when operating on unknown signals in real noise and suffers from computational inefficiencies in its pursuit of near-optimality. A compelling alternative that has emerged in recent years to address this problem is deep learning. Although it has shown significant promise when applied to the search for gravitational-waves in detector noise, we demonstrate the existence of a multitude of learning biases that hinder generalization and lead to significant loss in detection sensitivity. Our work identifies the sources of a set of 11 interconnected biases present in the supervised learning of the gravitational-wave detection problem and contributes mitigation tactics and training strategies to concurrently address them. In light of the identified biases, we demonstrate that existing detection sensitivity metrics are not reliable for machine-learning pipelines and discuss the trustworthiness of previous results. We use gravitational-wave domain knowledge to build a bespoke machine-learning-based binary black hole search pipeline called Sage that addresses these biases. Via the injection study presented in the Machine-Learning Gravitational-Wave Search Challenge, we show that Sage detects <inline-formula><mml:math><mml:mo>≈</mml:mo><mml:mn>11.2</mml:mn><mml:mo>%</mml:mo></mml:math></inline-formula> more signals than the benchmark PyCBC analysis at a false-alarm rate of one per month in O3a noise. Moreover, we also show that it can detect <inline-formula><mml:math><mml:mo>≈</mml:mo><mml:mn>48.29</mml:mn><mml:mo>%</mml:mo></mml:math></inline-formula> more signals than the previous best-performing machine-learning pipeline on the same dataset. We empirically prove that Sage can (i) effectively handle out-of-distribution noise power spectral densities, (ii) strongly reject non-Gaussian transient noise artifacts, and (iii) achieve higher detection sensitivities using less data than network architectures of a similar size. By studying machine-learning biases and conducting empirical investigations to understand the reasons for performance improvement or degradation, we aim to address the need for interpretability of machine-learning methods for gravitational-wave science. All code and implementations are available.
展开 ▾首次全面归纳出11种学习偏差并提供成套缓解策略;利用领域知识构建Sage管道,融合多速率采样、大核多尺度卷积与注意力机制,实现优于匹配滤波的检测灵敏度;揭示常用敏感距离指标会掩盖参数偏差,强调公平评估的必要性。
此前利用机器学习进行引力波探测的工作在理想条件下曾声称超越匹配滤波,但在真实噪声中表现大幅下降,MLGWSC-1 挑战显示 ML 方法在 O3a 噪声下的灵敏度远不及 PyCBC。后续 AresGW 虽宣称反超,但其增益可能来源于训练分布导致的高啁啾质量偏差。本文在全面识别偏差的基础上,通过训练分布重调、实时信号与噪声增强、归纳偏置注入等手段显著降低了系统偏差,使 Sage 在相同参数空间检测到更多信号。未来 ML 引力波探测应致力于偏差的协同缓解、无偏参数分布的探索以及更公平的评估指标,同时证明恰当的架构偏置和噪声多样性对提升泛化能力至关重要,有望在 O5 等实际观测中稳健运行。
预印本 2025-01-23 · 刊出 2025-11-04 · 收录 2026-10-12