|Date||May 31, 2022|
In this thesis, we study the loss surface of deep neural networks. Does the loss function of deep neural network have no bad local minimum like the convex function? Although it is well known for piece-wise linear activations, not much is known for the general smooth activations. We explore that a bad local minimum also exists for general smooth activations. In addition, we characterize the types of such local minima. This provides a partial explanation for the understanding of the loss surface of deep neural networks. Additionally, we present several applications of deep neural networks in learning theory, private machine learning, and computer vision.