By specializing our result to two-layer (i.e., one-hidden-layer) neural networks, it also provides a milder

In a seminal paper, Szegedy et al. [Sze+14] observed that deep networks are extremely vulnerable to adversarial examples, namely, very small perturbations to the inputs

In a seminal paper, Szegedy et al. [Sze+14] observed that deep networks are extremely vulnerable to adversarial examples, namely, very small perturbations to the inputs

Neural Information Processing Systems http://nips.cc/

We study the optimization problem associated with fitting two-layer ReLU neural networks with respect to the squared loss, where labels are generated by a target network. Use is made of the rich symmetry structure to develop a novel set of tools for studying the mechanism by which over-parameterization annihilates spurious minima through. Sharp analytic estimates are obtained for the loss and the Hessian spectrum at different minima, and it is shown that adding neurons can turn symmetric spurious minima into saddles through a local mechanism that does not generate new spurious minima; minima of smaller symmetry require more neurons. Using Cauchy's interlacing theorem, we prove the existence of descent directions in certain subspaces arising from the symmetry structure of the loss function. This analytic approach uses techniques, new to the field, from algebraic geometry, representation theory and symmetry breaking, and confirms rigorously the effectiveness of over-parameterization in making the associated loss landscape accessible to gradient-based methods.

A neural network with one hidden layer or a two-layer network (regardless of the input layer) is the simplest feedforward neural network, whose mechanism may be the basis of more general network architectures. However, even to this type of simple architecture, it is also a ``black box''; that is, it remains unclear how to interpret the mechanism of its solutions obtained by the back-propagation algorithm and how to control the training process through a deterministic way. This paper systematically studies the first problem by constructing universal function-approximation solutions. It is shown that, both theoretically and experimentally, the training solution for the one-dimensional input could be completely understood, and that for a higher-dimensional input can also be well interpreted to some extent. Those results pave the way for thoroughly revealing the black box of two-layer ReLU networks and advance the understanding of deep ReLU networks.