Reviews: DeepPINK: reproducible feature selection in deep neural networks

The paper proposes a method for feature selection in neural networks using a method for a controlled error rate, quantified through the False Discovery Rate. To control FDR the paper is using the model-X knockoffs framework [2, 3, 10]: construct random features that obey distributional properties with respect to the true features, and extract statistics (filters) of pairwise importance measures between true-knockoff dimensions. The choice of the importance function and the knockoff filters is flexible. The novelty of this paper lies in using a neural network (MLP) to get the importance measure through a linear layer that couples the true and knockoff features pairwise. The final statistic depends both on the trainable linear layer weights and the rest of the network weights.