Loading…

A hybrid deep model with cumulative learning for few-shot learning

Few-shot learning (FSL) aims to recognize unseen classes with only a few samples for each class. This challenging research endeavors to narrow the gap between the computer vision technology and the human visual system. Recently, mainstream approaches for FSL can be grouped into meta-learning and cla...

Full description

Saved in:
Bibliographic Details
Published in:Multimedia tools and applications 2023-05, Vol.82 (13), p.19901-19922
Main Authors: Liu, Jiehao, Yang, Zhao, Luo, Liufei, Luo, Mingkai, Hu, Luyu, Li, Jiahao
Format: Article
Language:English
Subjects:
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Few-shot learning (FSL) aims to recognize unseen classes with only a few samples for each class. This challenging research endeavors to narrow the gap between the computer vision technology and the human visual system. Recently, mainstream approaches for FSL can be grouped into meta-learning and classification learning. These two methods train the FSL model from local and global classification viewpoints respectively. In our work, we find the former method can effectively learn transferable knowledge (generalization capacity) with an episodic training paradigm but encounters the problem of slow convergence. The latter method can build an essential classification ability quickly (classification capacity) with a mini-batch training paradigm but easily causes an over-fitting problem. In light of this issue, we propose a hybrid deep model with cumulative learning to tackle the FSL problem by absorbing the advantages of the both methods. The proposed hybrid deep model innovatively integrates meta-learning and classification learning (IMC) in a unified two-branch network framework in which a meta-learning branch and a classification learning branch can work simultaneously. Besides, by considering the different characteristics of the two branches, we propose a cumulative learning strategy to take care of both generalization capacity learning and classification capacity learning in our IMC model training. With the proposed method, the model can quickly build the basic classification capability at the initial stage and continually mine discriminative class information during the remaining training for better generalization. Extensive experiments on CIFAR-FS, FC100, mini -ImageNet and tiered -ImageNet datasets are implemented to demonstrate the promising performance of our method.
ISSN:1380-7501
1573-7721
DOI:10.1007/s11042-022-14218-8