Variable selection with false discovery rate control in deep neural networks

Deep neural networks are famous for their high prediction accuracy, but they also known black-box nature and poor interpretability. We consider the problem of variable selection, that is, selecting input variables have significant predictive power on output, in deep networks. Most existing selection methods only applicable to shallow or computationally infeasible large datasets; moreover, lack a control quality selected variables. Here we propose backward elimination procedure called SurvNet, which is based new measure importance applies wide variety More importantly, SurvNet able estimate false discovery rate empirically. Further, adaptively determines how many eliminate at each step order maximize efficiency. The validity efficiency shown various simulated real datasets, its performance compared with other methods. Especially, systematic comparison knockoff-based shows although more rigorous data strong correlation, usually has higher power. Identifying salient features can be challenge authors developed works classification regression problems, one multiple output neurons,