应用于引力波探测的深度学习网络结构复杂度研究

Research on the Influence of Network Structure Complexity on Deep Learning for Gravitational Wave Detection

深度学习用于引力波探测是近几年的研究热点。匹配滤波法可以看作模板存储于卷积核参数中的单卷积层的神经网络,通过加深模型的深度可以在参数大大减少的同时得到相似的探测效果。对不同的卷积核大小、卷积核的数量(模型的宽度)、卷积层的数量(模型的深度)的深度学习引力波探测模型进行了讨论。另外,对全连接层前采用批量归一优化(batch normalization,BN)模型的探测效果进行了研究,发现单卷积层的模型在加入BN后的探测精度由50%左右提高到了90%以上。研究结果为匹配滤波模板数量的压缩提供了潜在的新方法,匹配滤波后通过BN层和全连接层也许能够大大减少匹配模板数量。

Deep learning for gravitational wave detection has been a hot spot in recent years.The matched filtering method can be regarded as a neural network with a single convolutional layer,and the template is stored in the convolution kernel.By increasing the depth of the model,similar detection effects can be obtained while the parameter number is greatly reduced.We study the deep learning gravitational wave detection models with different convolution kernel sizes,convolution kernel number,and convolution layer number.In addition,we investigate the detection effect of the model with batch normalization before the fully connected layer,and find that the detection accuracy of the model with a single convolutional layer increase from about 50%to more than 90%when the batch normalization is applied.A potential new method is provided for the compression of the number of matched filtering templates.By adding BN and full connection layer after matching filter,the number of matching templates may be greatly reduced.The generalization ability of the optimized CNN models on different background noise is also investigated.We find that the model trained by the data with H1 background noise can be used to detect the data with L1 background noise,but the accuracy is slightly reduced.