Feature Expansion and Enhanced Compression for Class Incremental Learning

14 Pages Posted: 25 Apr 2024

See all articles by Quentin Ferdinand

Quentin Ferdinand

affiliation not provided to SSRN

Benoit Clement

affiliation not provided to SSRN

Panagiotis Papadakis

affiliation not provided to SSRN

Quentin Oliveau

affiliation not provided to SSRN

Gilles Le Chenadec

affiliation not provided to SSRN

Abstract

Class incremental learning consists in training discriminative models to classify an increasing number of classes over time. However, doing so using only the newly added class data leads to the known problem of catastrophic forgetting of the previous classes. Recently, dynamic deep learning architectures have been shown to exhibit a better stability-plasticity trade-off by dynamically adding new feature extractors to the model in order to learn new classes followed by a compression step to scale the model back to its original size, thus avoiding a growing number of parameters. In this context, we propose a new algorithm that enhances the compression of previous class knowledge by cutting and mixing patches of previous class samples with the new images during compression using our Rehearsal-CutMix method. We show that this new data augmentation reduces catastrophic forgetting by specifically targeting past class information and improving its compression. Extensive experiments performed on the CIFAR and ImageNet datasets under diverse incremental learning evaluation protocols demonstrate that our approach consistently outperforms the state-of-the-art . The code will be made available upon publication of our work.

Keywords: Class incremental learning, Catastrophic forgetting, Knowledge distillation, CutMix, Dynamic networks, Convolutional neural networks

Suggested Citation

Ferdinand, Quentin and Clement, Benoit and Papadakis, Panagiotis and Oliveau, Quentin and Le Chenadec, Gilles, Feature Expansion and Enhanced Compression for Class Incremental Learning. Available at SSRN: https://ssrn.com/abstract=4806921 or http://dx.doi.org/10.2139/ssrn.4806921

Quentin Ferdinand (Contact Author)

affiliation not provided to SSRN ( email )

No Address Available

Benoit Clement

affiliation not provided to SSRN ( email )

No Address Available

Panagiotis Papadakis

affiliation not provided to SSRN ( email )

No Address Available

Quentin Oliveau

affiliation not provided to SSRN ( email )

No Address Available

Gilles Le Chenadec

affiliation not provided to SSRN ( email )

No Address Available

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
5
Abstract Views
46
PlumX Metrics