We consider the variable selection problem for two-sample tests, aiming to select the most informative features to best distinguish samples from two groups. We propose a kernel maximum mean discrepancy (MMD) framework to solve this problem and further derive its equivalent mixed-integer programming formulations for linear, quadratic, and Gaussian types of kernel functions. Our proposed framework admits advantages of both computational efficiency and nice statistical properties: (i) A closed-form solution is provided for the linear kernel case. Despite NP-hardness, we provide an exact mixed-integer semi-definite programming formulation for the quadratic kernel case, which further motivates the development of exact and approximation algorithms. We propose a convex-concave procedure that finds critical points for the Gaussian kernel case. (ii) We provide non-asymptotic uncertainty quantification of our proposed formulation under null and alternative scenarios. Experimental results demonstrate good performance of our framework.
In this paper, we explore some basic questions on the complexity of training Neural networks with ReLU activation function. We show that it is NP-hard to train a two- hidden layer feedforward ReLU neural network. If dimension d of the data is fixed then we show that there exists a polynomial time algorithm for the same training problem. We also show that if sufficient over-parameterization is provided in the first hidden layer of ReLU neural network then there is a polynomial time algorithm which finds weights such that output of the over-parameterized ReLU neural network matches with the output of the given data