B1373
Title: Adaptive variational deep learning
Authors: Ilsang Ohn - University of Notre Dame (Korea, South) [presenting]
Lizhen Lin - The University of Notre Dame (United States)
Abstract: A novel variational Bayes method is introduced for deep neural networks, which we call adaptive variational deep learning. The proposed method first obtains individual variational posterior over deep neural network models with varying network width (i.e., number of hidden nodes per layer) and combines them with certain weights to obtain a variational posterior over the entire deep neural network model. We show that the resulting variational posterior can obtain adaptive optimal contraction rates in a large number of statistical problems. Simulation results demonstrate that the strong empirical performance of the adaptive variational deep learning.
