Journal
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION
Volume 26, Issue 3, Pages 555-569Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TEVC.2021.3100576
Keywords
Training; Task analysis; Feature extraction; Genetic programming; Transfer learning; Genetic algorithms; Knowledge transfer; Few-shot learning (FSL); fitness evaluation; genetic programming (GP); image classification; representation
Funding
- Marsden Fund of New Zealand Government [VUW1913, VUW1914]
- Science for Technological Innovation Challenge (SfTI) Fund [E3603/2903]
- University Research Fund at Victoria University of Wellington [223805/3986]
- MBIE Data Science SSIF Fund [RTVU1914]
- National Natural Science Foundation of China (NSFC) [61876169]
- New Zealand Ministry of Business, Innovation & Employment (MBIE) [RTVU1914] Funding Source: New Zealand Ministry of Business, Innovation & Employment (MBIE)
Ask authors/readers for more resources
This article proposes a GP-based approach with a dual-tree representation and a new fitness function to automatically learn image features for FSIC. The results show that the proposed approach achieves significantly better performance than a large number of state-of-the-art methods on various types of FSIC tasks.
Few-shot image classification (FSIC) is an important but challenging task due to high variations across images and a small number of training instances. A learning system often has poor generalization performance due to the lack of sufficient training data. Genetic programming (GP) has been successfully applied to image classification and achieved promising performance. This article proposes a GP-based approach with a dual-tree representation and a new fitness function to automatically learn image features for FSIC. The dual-tree representation allows the proposed approach to have better searchability and learn richer features than a single-tree representation when the number of training instances is very small. The fitness function based on the classification accuracy and the distances of the training instances to the class centroids aims to improve the generalization performance. The proposed approach can deal with different types of FSIC tasks with various numbers of classes and different image sizes. The results show that the proposed approach achieves significantly better performance than a large number of state-of-the-art methods on nine 3-shot and 5-shot image classification datasets. Further analysis shows the effectiveness of the new components of the proposed approach, its good searchability, and the high interpretability of the evolved solutions.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available