Overcoming Catastrophic Forgetting for Continual Learning via Feature Propagation
Xuejun Han (Carleton University), Yuhong Guo (Carleton University)*
The 33rd British Machine Vision Conference
Abstract
Classical machine learners are designed only to tackle one task and suffer catastrophic forgetting as new tasks or classes emerge. To address this shortcoming, continual machine learners are elaborated to commendably learn a stream of tasks with domain and class shifts among different tasks. In this paper, we propose a general feature-propagation based contrastive continual learning method for image recognition in an online fashion which is capable of handling multiple continual learning scenarios. Specifically, we align the current and previous representation spaces by means of feature propagation and contrastive representation learning to bridge the domain shifts among distinct tasks. To further mitigate the class-wise shifts of the feature representation, a supervised contrastive loss is exploited to make the image embeddings of the same class closer than those of different classes. The extensive experimental results demonstrate the outstanding performance of the proposed method in multiple image classification tasks (MNIST, CIFAR-10/100 and Tiny ImageNet) compared to other cutting-edge continual learning methods.Video
Citation
@inproceedings{Han_2022_BMVC,
author = {Xuejun Han and Yuhong Guo},
title = {Overcoming Catastrophic Forgetting for Continual Learning via Feature Propagation},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/1011.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by