This paper describes an approach for feature selection for deep learning based on two separate networks. All reviewers have found some issues, but tend to accepting the paper, and the authors have promised to make the necessary improvements in the presentation.