Visit ShanghaiTech University | 中文 | How to find us
HOME > News and Events > Seminars
Global Optimality and Regularization in Deep Learning
Date: 2017/6/5             Browse: 268

Speaker:  Ben Haeffele 

Time:       June 5, 15:00-16:00.

Location: Room 1A-200, SIST Building

Inviter:    Prof.Shenghua Gao


Many commonly used methods in machine learning-including matrix factorization, tens or factorization, and deep neural networks-seek to learn meaningful representations directly from data. These techniques have achieved considerable empirical success in many fields, but common to a vast majority of these approaches are the disadvantages that the associated optimization problems are typically non-convex and one is forced to specify the size of the learned representation a priori. This talk will present a general framework which allows for the analysis of a wide range of non-convex representation learning problems. The framework allows the derivation of sufficient conditions to guarantee that a local minimizer of the non-convex optimization problem is a global minimizer and that from any initialization it is possible to reach a global minimizer using a local descent algorithm. Further, the flexibility of the framework allows for a wide range of regularization functions to be incorporated into the model to capture known prior information and to adaptively fit the size of the learned representation to the data instead of defining it a priori. Implications of this work will be discussed as they relate to modern practices in deep learning.


Ben Haeffele is an Associate Research Scientist in the Center for Imaging Science at Johns Hopkins University.  His research interests include multiple topics in representation learning, matrix/tensor factorization, sparse and low-rank methods, optimization, phase recovery, subspace clustering, and applications of machine learning in medicine, neuroscience, and microscopy.  He received his Ph.D. in Biomedical Engineering at Johns Hopkins University and his B.S. in Electrical Engineering from the Georgia Institute of Technology.



SIST-Seminar 17018