A Survey on Visualization for Explainable Representation Learning

PhD Qualifying Examination


Title: "A Survey on Visualization for Explainable Representation Learning"

by

Mr. Furui CHENG


Abstract:

Representation-learning methods, such as deep-learning methods, have 
become very successful as these methods can identify and summarize the 
underlying factors of variation behind data better than conventional 
feature engineering. However, current representation-learning methods can 
hardly disentangle the underlying factors, which makes them hard to 
interpret. The lack of interpretability limits their usage in 
security-critical areas and increases the difficulties for evaluation. To 
explain the representation-learning methods, researchers have exploited 
visualization techniques.

In this survey, we summarize the visualization approaches from three 
facets: Why, What, and How, to explain the 
representation-learning methods.  From the perspective of Why, we classify 
the problems into the representation-understanding problems and 
the inference-understanding problems. From the perspective of What, we 
classify the explanations into symbolic explanations and semantic 
explanations. Then we introduce How the visualization techniques are 
designed to generate different explanations for the two problems. In the 
end, we discuss the challenges and opportunities for future research in 
explainable representation learning.


Date:			Thursday, 3 October 2019

Time:                  	10:00am - 12:00noon

Venue:                  Room 2408
                         Lifts 17/18

Committee Members:	Prof. Huamin Qu (Supervisor)
 			Dr. Sunghun Kim (Chairperson)
 			Dr. Qifeng Chen
 			Dr. Raymond Wong


**** ALL are Welcome ****