Ling Yue, Lin Feng, Qiuping Shuai, Zihao Li, Lingxiao Xu
{"title":"Multi-level adaptive feature representation based on task augmentation for Cross-Domain Few-Shot learning","authors":"Ling Yue, Lin Feng, Qiuping Shuai, Zihao Li, Lingxiao Xu","doi":"10.1007/s10489-024-06110-9","DOIUrl":null,"url":null,"abstract":"<div><p>Cross-Domain Few-Shot Learning (CDFSL) is one of the most cutting-edge fields in machine learning. It not only addresses the traditional few-shot problem but also allows for different distributions between base classes and novel classes. However, most current CDFSL models only focus on the generalization performance of high-level features during training and testing, which hinders their ability to generalize well to domains with significant gaps. To overcome this problem, we propose a CDFSL method based on Task Augmentation and Multi-Level Adaptive features representation(TA-MLA). At the feature representation level, we introduce a meta-learning strategy for multi-level features and adaptive features. The former come from different layers of network. They jointly participate in image prediction to fully explore transferable features suitable for cross-domain scenarios. The latter is based on a feature adaptation module of feed-forward attention, aiming to learn domain-adaptive features to improve the generalization of the model. At the training task level, we employ a plug-and-play Task Augmentation(TA) module to generate challenging tasks with adaptive inductive biases, thereby expanding the distribution of the source domain and further bridging domain gaps. Extensive experiments conducted on multiple datasets. The results demonstrate that our method based on meta-learning can effectively improves few-shot classification performance, especially in cases with significant domain shift.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 4","pages":""},"PeriodicalIF":3.4000,"publicationDate":"2025-01-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Applied Intelligence","FirstCategoryId":"94","ListUrlMain":"https://link.springer.com/article/10.1007/s10489-024-06110-9","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
Cross-Domain Few-Shot Learning (CDFSL) is one of the most cutting-edge fields in machine learning. It not only addresses the traditional few-shot problem but also allows for different distributions between base classes and novel classes. However, most current CDFSL models only focus on the generalization performance of high-level features during training and testing, which hinders their ability to generalize well to domains with significant gaps. To overcome this problem, we propose a CDFSL method based on Task Augmentation and Multi-Level Adaptive features representation(TA-MLA). At the feature representation level, we introduce a meta-learning strategy for multi-level features and adaptive features. The former come from different layers of network. They jointly participate in image prediction to fully explore transferable features suitable for cross-domain scenarios. The latter is based on a feature adaptation module of feed-forward attention, aiming to learn domain-adaptive features to improve the generalization of the model. At the training task level, we employ a plug-and-play Task Augmentation(TA) module to generate challenging tasks with adaptive inductive biases, thereby expanding the distribution of the source domain and further bridging domain gaps. Extensive experiments conducted on multiple datasets. The results demonstrate that our method based on meta-learning can effectively improves few-shot classification performance, especially in cases with significant domain shift.
期刊介绍:
With a focus on research in artificial intelligence and neural networks, this journal addresses issues involving solutions of real-life manufacturing, defense, management, government and industrial problems which are too complex to be solved through conventional approaches and require the simulation of intelligent thought processes, heuristics, applications of knowledge, and distributed and parallel processing. The integration of these multiple approaches in solving complex problems is of particular importance.
The journal presents new and original research and technological developments, addressing real and complex issues applicable to difficult problems. It provides a medium for exchanging scientific research and technological achievements accomplished by the international community.