A Survey on Fault Localization for Deep Learning Systems

PhD Qualifying Examination


Title: "A Survey on Fault Localization for Deep Learning Systems"

by

Miss Jialun CAO


Abstract:

Deep Neural Networks (DNNs) have been actively deployed for 
mission-critical applications such as fraud detection, medical diagnosis, 
and autonomous driving. It motivates intensive studies to understand and 
detect the potential faults in DNN systems. However, debugging DNN systems 
is a non-trivial problem. Unlike traditional software systems, the 
behavior of a DNN is not explicitly encoded by the program's control flow. 
Instead, the underlying program defines only the configurations of a DNN 
model (e.g., network structures, training strategies, and 
hyperparameters), while the DNN model learns the parameters itself under 
these configurations from the training data. In addition, the densely 
inter-dependent neural network, massive trainable parameters, and the 
stochastic behavior arising from DNN training further increase the 
challenges of debugging DNN programs. In this survey, we conduct a 
systematic literature review on fault localization techniques for DNN 
systems and categorize them according to the debugging components they 
targeted on This survey introduces the general ideas, workflows, required 
techniques, and potential limitations as well as challenges for each 
category. It also outlines the promising research directions worthy of 
exploring in the future.


Date:			Thursday, 15 July 2021

Time:                  	10:00am - 12:00noon

Zoom meeting:
https://hkust.zoom.us/j/96994112085?pwd=UW1TaytUYjZFQkEvTDlDbWtuTGFQdz09

Committee Members:	Prof. Shing-Chi Cheung (Supervisor)
 			Dr. Sunghun Kim (Chairperson)
 			Dr. Yangqiu Song
 			Prof. Raymond Wong


**** ALL are Welcome ****