A survey on few-shot classification

PhD Qualifying Examination


Title: "A survey on few-shot classification"

by

Miss Shuhan LI


Abstract:

Deep learning has shown its success in various applications when there are 
sufficient data during training, but it suffers a lot when the data set is 
small. People also realize that the learning process for humans is much 
simple than a deep learning model: we can learn new knowledge with only a 
few examples. To simulate the learning process of us, few-shot learning 
(FSL) targets to deal with this challenge by leveraging the prior 
knowledge to quickly adapt to the unseen categories with limited labeled 
data, and it has made some progress recently. In this survey, we provide a 
thorough analysis of the few-shot image classification problem in the 
following aspects: Firstly, we give the definition and the common settings 
(including problem definition and the benchmark datasets) of FSL over 
image classification problem. Secondly, based on the three main categories 
of few-shot learning methods built upon the concept of meta-learning, 
which are optimization-based, distance-metric-based, and 
hallucination-based methods specifically, we summarize the general 
frameworks for each group by introducing several typical algorithms. 
Additionally, a new baseline model of FSL was proposed by some researchers 
recently which consists of a standard supervised classifier trained on the 
base dataset and a fine-tuning stage on novel classes episodically. This 
so-called transfer-based method empirically achieves competitive accuracy 
comparing to meta-learning-based ways with less computational cost. We try 
to analyze the advantages of this latest framework and explore some 
state-of-the-art approaches built on it. In the end, we conclude the 
article with several concerns about current methods and try to provide 
potential research directions in the future.


Date:			Friday, 29 October 2021

Time:                  	11:45am - 1:45pm

Zoom meeting:		https://hkust.zoom.us/j/2526320493

Committee Members:	Prof. Tim Cheng (Supervisor)
 			Dr. Xiaomeng Li (Supervisor)
 			Prof. Pedro Sander (Chairperson)
 			Dr. Dan Xu


**** ALL are Welcome ****