Representative Multi-Domain Feature Selection Based Cross-Domain Few-Shot Classification

Typical few-shot learning methods implicitly assume that the meta-training dataset and the meta-test dataset come from the same domain, which greatly limits the application of few-shot learning methods. To deal with this limitation, cross-domain few-shot classification has been proposed, in which th...

Full description

Saved in:
Bibliographic Details
Main Authors: Weng, Zhewei, Feng, Chunyan, Zhang, Tiankui, Zhu, Yutao, Chen, Zeren
Format: Conference Proceeding
Language:English
Subjects:
Benchmark testing
Classification algorithms
Conferences
Cross-domain
Feature extraction
Feature Selection
Few-shot Learning
Learning systems
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Typical few-shot learning methods implicitly assume that the meta-training dataset and the meta-test dataset come from the same domain, which greatly limits the application of few-shot learning methods. To deal with this limitation, cross-domain few-shot classification has been proposed, in which there is a significant difference between the meta-training set as the source domain and the meta-test set as the target domain. To address this problem, we introduce the idea of multi-domain feature selection and propose representative multi-domain feature selection (RMFS) algorithm, which optimizes the multi-domain feature extraction stage and the multi-domain feature selection stage. The effectiveness of the proposed algorithm is demonstrated by experiments on the benchmark dataset Meta-Dataset.
ISSN:2575-4955
DOI:10.1109/IC-NIDC54101.2021.9660577