Recovery Guarantees for One-hidden-layer Neural Networks
Kai Zhong, UT Austin
10 – 11AM
Friday Nov 3, 2017
POB 6.304
Abstract
Neural Networks (NNs) have achieved great empirical success recently in various applications, including computer vision, natural language processing and reinforcement learning. However, due to the non-convexity of the NNs, the theoretical understanding of NNs is still limited. In this presentation, I will show that when inputs are sampled from Gaussian distribution and the activation function satisfies some properties, one-hidden-layer fully-connected NNs and one-hidden-layer convolutional NNs can be recovered in polynomial time. Specifically, I will first give a brief overview of recent theoretical progress on neural networks. Then we show gradient descent with tensor method initialization is guaranteed to converge to the ground truth parameters of the NNs with polynomial sample complexity and computational complexity.