Overcoming challenges in leveraging GANs for few-shot data augmentation

Christopher Beckham, Issam H. Laradji, Pau Rodriguez, David Vazquez, Derek Nowrouzezahrai, Christopher Pal
Proceedings of The 1st Conference on Lifelong Learning Agents, PMLR 199:255-280, 2022.

Abstract

In this paper, we explore the use of GAN-based few-shot data augmentation as a method to improve few-shot classification performance. We perform an exploration into how a GAN can be fine-tuned for such a task (one of which is in a \emph{class-incremental} manner), as well as a rigorous empirical investigation into how well these models can perform to improve few-shot classification. We identify issues related to the difficulty of training and applying such generative models under a purely supervised regime with very few examples, as well as issues regarding the evaluation protocols of existing works. We also find that in this regime, classification accuracy is highly sensitive to how the classes of the dataset are randomly split. To address difficulties in applying these generative models under the few-shot regime, we propose a simple and pragmatic semi-supervised fine-tuning approach, and demonstrate gains in FID and precision-recall metrics as well as classification performance.

Cite this Paper

BibTeX
@InProceedings{pmlr-v199-beckham22a, title = {Overcoming challenges in leveraging GANs for few-shot data augmentation}, author = {Beckham, Christopher and Laradji, Issam H. and Rodriguez, Pau and Vazquez, David and Nowrouzezahrai, Derek and Pal, Christopher}, booktitle = {Proceedings of The 1st Conference on Lifelong Learning Agents}, pages = {255--280}, year = {2022}, editor = {Chandar, Sarath and Pascanu, Razvan and Precup, Doina}, volume = {199}, series = {Proceedings of Machine Learning Research}, month = {22--24 Aug}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v199/beckham22a/beckham22a.pdf}, url = {https://proceedings.mlr.press/v199/beckham22a.html}, abstract = {In this paper, we explore the use of GAN-based few-shot data augmentation as a method to improve few-shot classification performance. We perform an exploration into how a GAN can be fine-tuned for such a task (one of which is in a \emph{class-incremental} manner), as well as a rigorous empirical investigation into how well these models can perform to improve few-shot classification. We identify issues related to the difficulty of training and applying such generative models under a purely supervised regime with very few examples, as well as issues regarding the evaluation protocols of existing works. We also find that in this regime, classification accuracy is highly sensitive to how the classes of the dataset are randomly split. To address difficulties in applying these generative models under the few-shot regime, we propose a simple and pragmatic semi-supervised fine-tuning approach, and demonstrate gains in FID and precision-recall metrics as well as classification performance.} }
Endnote
%0 Conference Paper %T Overcoming challenges in leveraging GANs for few-shot data augmentation %A Christopher Beckham %A Issam H. Laradji %A Pau Rodriguez %A David Vazquez %A Derek Nowrouzezahrai %A Christopher Pal %B Proceedings of The 1st Conference on Lifelong Learning Agents %C Proceedings of Machine Learning Research %D 2022 %E Sarath Chandar %E Razvan Pascanu %E Doina Precup %F pmlr-v199-beckham22a %I PMLR %P 255--280 %U https://proceedings.mlr.press/v199/beckham22a.html %V 199 %X In this paper, we explore the use of GAN-based few-shot data augmentation as a method to improve few-shot classification performance. We perform an exploration into how a GAN can be fine-tuned for such a task (one of which is in a \emph{class-incremental} manner), as well as a rigorous empirical investigation into how well these models can perform to improve few-shot classification. We identify issues related to the difficulty of training and applying such generative models under a purely supervised regime with very few examples, as well as issues regarding the evaluation protocols of existing works. We also find that in this regime, classification accuracy is highly sensitive to how the classes of the dataset are randomly split. To address difficulties in applying these generative models under the few-shot regime, we propose a simple and pragmatic semi-supervised fine-tuning approach, and demonstrate gains in FID and precision-recall metrics as well as classification performance.
APA
Beckham, C., Laradji, I.H., Rodriguez, P., Vazquez, D., Nowrouzezahrai, D. & Pal, C.. (2022). Overcoming challenges in leveraging GANs for few-shot data augmentation. Proceedings of The 1st Conference on Lifelong Learning Agents, in Proceedings of Machine Learning Research 199:255-280 Available from https://proceedings.mlr.press/v199/beckham22a.html.

Related Material