Neural Networks are Convex Regularizers: Exact Polynomial-time Convex Optimization Formulations for Two-Layer Networks

We develop exact representations of two layer neural networks with rectified linear units in terms of a single convex program with number of variables polynomial in the number of training samples and number of hidden neurons. Our theory utilizes semi-infinite duality and minimum norm regularization. Moreover, we show that certain standard multi-layer convolutional neural networks are equivalent to L1 regularized linear models in a polynomial sized discrete Fourier feature space. We also introduce exact semi-definite programming representations of convolutional and fully connected linear multi-layer networks which are polynomial size in both the sample size and dimension.