# Training Strategies for Deep Learning Gravitational-Wave Searches

**Authors:** Marlin B. Schäfer, Ondřej Zelenka, Alexander H. Nitz, Frank Ohme, Bernd Brügmann  
**arXiv:** [2106.03741](https://arxiv.org/abs/2106.03741)  
**Date:** 2021-06-07  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2106.03741v1/

## Key points

- Authors explore potential of deep learning in detecting gravitational waves emitted by compact binary systems
- Deep learning algorithms can generalize low signal-to-noise ratio (SNR) signals to high SNR ones but not vice versa
- Providing high SNR signals during training does not offer significant benefits; fastest convergence achieved with low SNR samples early on
- Networks sometimes struggle to recover signals when false alarm probability of less than $10^{-3}$ is required
- Introduction of unbounded Softmax replacement (USR) after training significantly improves performance, retaining over 97.5% sensitivity down to a false-alarm rate of 1 per month

## Summary

, , , , 

In their paper "Training Strategies for Deep Learning Gravitational-Wave Searches," Marlin B. Schäfer, Ondřej Zelenka, Alexander H. Nitz, Frank Ohme, and Bernd Brügmann explore the potential of deep learning in detecting gravitational waves emitted by compact binary systems. The authors focus on signals from non-spinning binary black holes and investigate various training strategies for presenting data to neural networks. The study reveals that deep learning algorithms have the capability to generalize low signal-to-noise ratio (SNR) signals to high SNR ones but not vice versa. Interestingly, providing high SNR signals during training does not offer any significant benefits, as the fastest convergence is achieved when low SNR samples are presented early on in the training process. During testing, it was observed that the networks sometimes struggle to recover signals when a false alarm probability of less than $10^{-3}$ is required. To address this limitation, the authors introduce a modification called unbounded Softmax replacement (USR) after training. This alteration significantly improves the performance of the machine learning search, retaining over 97.5% of the sensitivity of a matched-filter search down to a false-alarm rate of 1 per month. Overall, this research highlights the potential of deep learning in enhancing gravitational-wave searches and provides valuable insights into optimizing training strategies for improved detection capabilities in compact binary systems. The findings contribute to advancing our understanding of gravitational wave astronomy and pave the way for more efficient signal extraction techniques in future studies.