EE Seminar: On Expressiveness and Optimization in Deep Learning

24 בדצמבר 2018, 15:00 
חדר 011, בניין כיתות-חשמל 

 (The talk will be given in English)

 

Speaker:     Dr. Nadav Cohen
                   School of Mathematics, Institute for Advanced Study, Princeton NJ

 

Monday, December 24th, 2018
15:00 - 16:00

Room 011, Kitot Bldg., Faculty of Engineering

 

On Expressiveness and Optimization in Deep Learning

 

Abstract

Understanding deep learning calls for addressing three fundamental questions: expressiveness, optimization and generalization.  Expressiveness refers to the ability of compactly sized deep neural networks to represent functions capable of solving real-world problems.  Optimization concerns the effectiveness of simple gradient-based algorithms in solving non-convex neural network training programs.  Generalization treats the phenomenon of deep learning models not overfitting despite having much more parameters than examples to learn from.  This talk will describe a series of works aimed at unraveling some of the mysteries behind expressiveness and optimization.  I will begin by establishing an equivalence between convolutional and recurrent networks --- the most successful deep learning architectures to date --- and hierarchical tensor decompositions.  The equivalence will be used to answer various questions concerning expressiveness, resulting in new theoretically-backed tools for deep network design.  I will then turn to discuss a recent line of work analyzing optimization of deep linear neural networks.  By studying the trajectories of gradient descent, we will derive the most general guarantee to date for efficient convergence to global minimum of a gradient-based algorithm training a deep network.  Moreover, in stark contrast with conventional wisdom, we will see that sometimes, gradient descent can train a deep linear network faster than a classic linear model.  In other words, depth can accelerate optimization, even without any gain in expressiveness, and despite introducing non-convexity to a formerly convex problem.

 

Works covered in this talk were in collaboration with Amnon Shashua, Sanjeev Arora, Elad Hazan, Or Sharir, Yoav Levine, Noah Golowich, Wei Hu, Ronen Tamari and David Yakira. 

 

Short Bio

Nadav Cohen is a postdoctoral member at the School of Mathematics in the Institute for Advanced Study.  His research focuses on the theoretical and algorithmic foundations of deep learning.  In particular, he is interested in mathematically analyzing aspects of expressiveness, optimization and generalization, with the goal of deriving theoretically founded procedures and algorithms that will improve practical performance.  Nadav earned his PhD at the School of Computer Science and Engineering in the Hebrew University of Jerusalem, under the supervision of Prof. Amnon Shashua. Prior to that, he obtained a BSc in electrical engineering and a BSc in mathematics (both summa cum laude) at the Technion Excellence Program for distinguished undergraduates. For his contributions to the theoretical understanding of deep learning, Nadav received a number of awards, including the Google Doctoral Fellowship in Machine Learning, the Rothschild Postdoctoral Fellowship, and the Zuckerman Postdoctoral Fellowship.

אוניברסיטת תל אביב עושה כל מאמץ לכבד זכויות יוצרים. אם בבעלותך זכויות יוצרים בתכנים שנמצאים פה ו/או השימוש
שנעשה בתכנים אלה לדעתך מפר זכויות, נא לפנות בהקדם לכתובת שכאן >>