Events Calendar

Beating the Perils of Non-convexity: Guaranteed Training of Neural Networks Using Tensor Methods

Speaker: Majid Janzamin

Title: Beating the Perils of Non-convexity: Guaranteed Training of Neural Networks Using Tensor Methods

Series: AI Seminar

Abstract:
Training neural networks is a highly non-convex problem and in general is NP-hard. Local search methods such as gradient descent get stuck in spurious local optima, especially in high dimensions. We present a novel method based on tensor decomposition that trains a two-layer neural network with guaranteed risk bounds with polynomial sample and computational complexity. We also demonstrate how unsupervised learning can help in supervised tasks. In our context, we estimate probabilistic score functions via unsupervised learning which are then employed for training neural networks using tensor methods.

Bio:
Majid Janzamin is a sixth year PhD student at the EECS Dept. at UC Irvine. He received his BSc and MSc in Electrical Engineering, from Sharif University of Technology, Tehran, Iran in 2007 and 2010, respectively. He has also visited and has done internship at Microsoft research labs at New England and Silicon Valley. His research interests are in the area of large-scale machine learning and high-dimensional statistics, and probabilistic modeling. In particular, he has worked on optimization methods for learning graphical models, and tensor methods for latent variable models.

Location: Information Science Institute (ISI) - 11th floor large conference room

WebCast Link: http://webcasterms1.isi.edu/mediasite/Viewer/?peid=07a00eec98a44b81ab87fdfd8a6368151d

Audiences: Everyone Is Invited

Contact: Kary LAU