Alexandr Andoni, Rina Panigrahy, Gregory Valiant, and Li Zhang
We study the effectiveness of learning low degree polynomials using neural networks by the gradient descent method. While neural networks have been shown to have great expressive power, and gradient descent has been widely used in practice for learning neural networks, few theoretical guarantees are known for such methods. In particular, it is well known that gradient descent can get stuck at local minima, even for simple classes of target functions. In this paper, we present several positive theoretical results to support the effectiveness of neural networks. We focus on twolayer neural networks where the bottom layer is a set of non-linear hidden nodes, and the top layer node is a linear function, similar to Barron (1993). First we show that for a randomly initialized neural network with sufficiently many hidden units, the generic gradient descent algorithm learns any low degree polynomial, assuming we initialize the weights randomly. Secondly, we show that if we use complex-valued weights (the target function can still be real), then under suitable conditions, there are no “robust local minima”: the neural network can always escape a local minimum by performing a random perturbation. This property does not hold for real-valued weights. Thirdly, we discuss whether sparse polynomials can be learned with small neural networks, with the size dependent on the sparsity of the target function.