A Survey of Deep Learning Framework Testing

PhD Qualifying Examination


Title: "A Survey of Deep Learning Framework Testing"

by

Mr. Meiziniu LI


Abstract:

Deep Learning (DL) frameworks such as TensorFlow provide a compilation of 
interfaces and implementations for deep learning algorithms. Recent DL 
applications are mostly built on top of DL frameworks. Thus, ensuring the 
quality of these systems is critical to the safety and reliability of DL 
applications. However, testing DL frameworks is a non-trivial problem. 
Functions of DL frameworks commonly require DL-specific input constraints 
such as variable type or tensor dimension. Therefore, valid test cases can 
not be easily generated without knowing these constraints. Moreover, 
uncertain factors such as randomness in deep learning algorithms or 
floating-point computation deviation will also affect the test oracle 
design.

In recent years, assuring the quality of DL frameworks is becoming an 
emerging topic. This survey provides a systematic literature review of 
these research works. It concludes the characteristics of bugs inside DL 
frameworks, including the distributions, symptoms, and root causes of DL 
framework bugs. Besides, it summarizes the general workflows of existing 
DL framework testing techniques, categorizes them based on their testing 
method, and reveals their limitations. Finally, it points out some 
possible research directions worthy of exploring in the future with 
preliminary results.


Date:  			Friday, 8 July 2022

Time:                  	3:30pm - 5:30pm

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

Committee Members:	Prof. Shing-Chi Cheung (Supervisor)
 			Dr. Wei Wang (Chairperson)
 			Dr. Shuai Wang
 			Prof. Raymond Wong


**** ALL are Welcome ****