- Abstract: We study model recovery for data classification, where the training labels are generated from a one-hidden-layer fully -connected neural network with sigmoid activations, and the goal is to recover the weight vectors of the neural network. We prove that under Gaussian inputs, the empirical risk function using cross entropy exhibits strong convexity and smoothness uniformly in a local neighborhood of the ground truth, as soon as the sample complexity is sufficiently large. This implies that if initialized in this neighborhood, which can be achieved via the tensor method, gradient descent converges linearly to a critical point that is provably close to the ground truth without requiring a fresh set of samples at each iteration. To the best of our knowledge, this is the first global convergence guarantee established for the empirical risk minimization using cross entropy via gradient descent for learning one-hidden-layer neural networks, at the near-optimal sample and computational complexity with respect to the network input dimension.
- Keywords: cross entropy, neural networks, parameter recovery
- TL;DR: We provide the first theoretical analysis of guaranteed recovery of one-hidden-layer neural networks under cross entropy loss for classification problems.