*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Duality Between Deep Learning And Algorithm Design
Date: 04/13/2022
Time: 11:00 am EST
Location (BlueJeans meeting link): https://bluejeans.com/788150008/3007
Xinshi Chen
Machine Learning PhD Student
School of Mathematics
Georgia Institute of Technology
Committee
1 Dr. Le Song (Advisor)
2 Dr. Koltchinskii Vladimir (Advisor)
3 Dr. Guanghui Lan
4 Dr. Xiuwei Zhang
5 Dr. Molei Tao
Abstract
Deep learning is a data-driven method, whereas conventional algorithm design is a knowledge-driven method. Based on their connections and complementary features, this thesis introduces new methods to combine the merits of both and, in turn, improve both. Specifically:
The development of deep neural networks is hindered by their lack of interpretability and the need for very large training sets. To eliminate these issues, this thesis introduces the use of algorithms as modeling priors to integrate specialized knowledge of domain experts into deep learning models. From both the empirical and theoretical perspective, this thesis explains how such algorithm inspired deep learning models can achieve improved interpretability and sample efficiency.
In conventional algorithm design, domain experts will first develop a model to describe the mechanism behind it and then establish a mathematical algorithm to find the solution. Notwithstanding its interpretability, this model-based method is inferior in terms of its limited effective range and accuracy. This is mostly due to the simplifying assumptions of the models which often deviate from real-world problems. To address these issues, this thesis investigates the potential of deep learning based methods for discovering data-driven algorithms that adapt better to the interested problem distribution, from both the empirical and theoretical perspectives.