Learning to Generalize with Deep Neural Networks

Event Sponsor: 
Mathmatics and Computer Science Division Candidate Presentation
Start Date: 
Jul 19 2019 - 10:30am
Building 240/Room 4301
Argonne National Laboratory
Krishnan Raghavan
Speaker(s) Title: 
Missouri University of Science & Technology
Prasanna Balaprakash

The human brain is fascinating. Many philosophers have contemplated the incredible feats it can accomplish and have attempted to replicate the inner workings of the human brain.  One of the first steps in this regard was made by McCulloch and Pitts in 1943. Since then the research in the field of artificial intelligence has prospered for decades. However, we are still far away from achieving true general intelligence. For instance, consider generalization that refers to the idea of learning a concept from a small amount of experience and then extrapolating it to the larger world. Humans are inherently capable of this whereas generalizing is still beyond the reach of the current state of the art in artificial intelligence.

In this talk, we will consider the problem of supervised learning through Deep Neural Networks (DNN) and we will review why poor generalization is observed with DNNs. To mitigate this issue, a data-augmentation-based approach will be introduced. Furthermore, we will discuss a novel learning framework where the cost due to generalization is approximated and then minimized during learning. To efficiently optimize the DNN for classification, a distributed learning regime will be presented. Simulation and theoretical results will be reviewed to validate the approach.

Miscellaneous Information: 

The link to the recorded presentation will post to the CELS Seminars website after the seminar.