Learning ReLU Networks via Alternating Minimization

We propose and analyze a new family of algorithms for training neural networks with ReLU activations. Our algorithms are based on the technique of alternating minimization: estimating the activation patterns of each ReLU for all given samples, interleaved with weight updates via a least-squares step. The main focus of our paper are 1-hidden layer networks with k hidden neurons and ReLU activation. We show that under standard distributional assumptions on the d-dimensional input data, our algorithm provably recovers the true `ground truth' parameters in a linearly convergent fashion. This holds as long as the weights are sufficiently well initialized; furthermore, our method requires only n=widetildeO(dk2) samples. We also analyze the special case of 1-hidden layer networks with skipped connections, commonly used in ResNet-type architectures, and propose a novel initialization strategy for the same. For ReLU based ResNet type networks, we provide the first linear convergence guarantee with an end-to-end algorithm. We also extend this framework to deeper networks and empirically demonstrate its convergence to a global minimum.