Few-shot Learning and Zero-shot Learning

PhD Qualifying Examination


Title: "Few-shot Learning and Zero-shot Learning"

By

Miss Yaqing WANG


Abstract:

Artificial intelligence has gained much attention recently owing to the 
development of machine learning, especially deep learning. However, the 
success of deep models relies on large-scale data and time-consuming 
learning. The ability of rapid learning from few samples remains a huge 
challenge. Few-shot learning (FSL) is a research topic which deals with 
this problem. It learns models that can generalize well for new classes 
with few labeled samples, which mimics human's ability to acquire 
knowledge from few examples through generalization and analogy. In this 
survey, we first introduce the development of FSL, then we give a 
literature review on existing works with a detailed comparison. We also 
briefly review zero-shot learning (ZSL) which can benefit FSL. ZSL builds 
models for new classes in testing by using semantic information learned in 
other domain, which refers to human's ability to associate different 
knowledge sources. FSL and ZSL are intertwined. For example, we can build 
a model for the new class using ZSL methods and then fine tune it using 
the few labeled samples to do FSL. So we include ZSL to provide a 
comprehensive understanding. Finally, we conclude the survey with a 
discussion on potential future work.


Date:			Monday, 8 May 2017

Time:                  	3:00pm - 5:00pm

Venue:                  Room 4475
                        Lifts 25/26

Committee Members:	Prof. James Kwok (Supervisor)
                        Dr. Brian Mak (Chairperson)
 			Prof. Lionel Ni
 			Prof. Yangqiu Song


**** ALL are Welcome ****