More Classifiers, Less Forgetting: A Generic Multi-classifier Paradigm for Incremental Learning (ECCV 2020)

• Exploit the classifier ensemble for reducing forgetting on learning tasks incrementally.
• Extend two regularization methods (MAS and LwF) focusing on parameter and activation regularization.
• Obtain consistent improvements over the single-classifier paradigm.

Dependencies

• PyTorch
• Python
• Numpy
• scipy

Data

• Download the dataset (CIFAR-100, Tiny-ImageNet, SVHN) and save them to the 'data' directory.
• SVHN is used as an out-of-distribution dataset for training additional side classifiers.

Experiment on CIFAR-100 incremental benchmark

• Run cifar100_MUC_MAS.py to train the MUC-MAS method.

• Run cifar100_MUC_LwF.py to train the MUC-LwF method.

Experiment on Tiny-ImageNet incremental benchmark

• Run tinyimagenet_MUC_MAS.py to train the MUC-MAS method.

• Run tinyimagenet_MUC_LwF.py to train the MUC-LwF method.

Notes

• Some codes are based on the codebase of the repository.
• More instructions will be provided later.

Citation

Please cite the following paper if it is helpful for your research:

@InProceedings{MUC_ECCV2020,
author = {Liu, Yu and Parisot, Sarah and Slabaugh, Gregory and Jia, Xu and Leonardis,Ales and Tuytelaars, Tinne}
title = {More Classifiers, Less Forgetting: A Generic Multi-classifier Paradigm for Incremental Learning},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2020}
}