Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration
Entropy · 10.05.2023 · 25 (5) · 776
DOI: 10.3390/e25050776 ↗Kurz gefasst
Fachartikel in der Zeitschrift Entropy, online erschienen am 10.05.2023. Verfasst von Wei Zhang und 1 weiteren Autor:innen. Beteiligte Einrichtung: Department of Electronic Engineering, Fudan University, Shanghai 200438, China. Die Arbeit wurde bisher 14-mal zitiert (Crossref) und nennt 54 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Few shot class incremental learning (FSCIL) is an extremely challenging but valuable problem in real-world applications. When faced with novel few shot tasks in each incremental stage, it should take into account both catastrophic forgetting of old knowledge and overfitting of new categories with limited training data. In this paper, we propose an efficient prototype replay and calibration (EPRC) method with three stages to improve classification performance. We first perform effective pre-training with rotation and mix-up augmentations in order to obtain a strong backbone. Then a series of pseudo few shot tasks are sampled to perform meta-training, which enhances the generalization ability of both the feature extractor and projection layer and then helps mitigate the over-fitting problem of few shot learning. Furthermore, an even nonlinear transformation function is incorporated into the similarity computation to implicitly calibrate the generated prototypes of different categories and alleviate correlations among them. Finally, we replay the stored prototypes to relieve catastrophic forgetting and rectify prototypes to be more discriminative in the incremental-training stage via an explicit regularization within the loss function. The experimental results on CIFAR-100 and miniImageNet demonstrate that our EPRC significantly boosts the classification performance compared with existing mainstream FSCIL methods.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Electronic Engineering, Fudan University, Shanghai 200438, China.
ORCID-Profile: Xiaodong Gu
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Zhang, W., & Gu, X. (2023). Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration. Entropy, 25(5), 776. https://doi.org/10.3390/e25050776
Vancouver
Zhang W, Gu X. Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration. Entropy. 2023;25(5):776. doi:10.3390/e25050776
BibTeX
@article{Zhang2023,
author = {Zhang, Wei and Gu, Xiaodong},
title = {Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration},
journal = {Entropy},
year = {2023},
volume = {25},
number = {5},
pages = {776},
doi = {10.3390/e25050776}
}
EUVIMED