[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Can Synthetic Images Conquer Forgetting? Beyond Unexplored Doubts in Few-Shot Class-Incremental Learning

Created by
  • Haebom

Author

Junsu Kim, Yunhoe Ku, Seungryul Baek

Outline

In this paper, we propose a Diffusion-FSCIL method that uses a pre-trained text-to-image diffusion model as a fixed backbone to solve the FSCIL problem, which suffers from very limited training data. We aim to solve the FSCIL problem by leveraging the advantages of large-scale generative models, such as generative power obtained through large-scale pre-training, multi-scale representation, and flexibility of representation through text encoders. We extract multiple complementary diffusion features to act as latent replay, and slightly utilize feature distillation to prevent generative bias. We achieve efficiency by using a fixed backbone, minimal trainable components, and batch processing of multiple feature extractions. Experimental results on CUB-200, miniImageNet, and CIFAR-100 datasets show that Diffusion-FSCIL outperforms existing state-of-the-art methods and effectively adapts to new classes while maintaining performance on previously learned classes.

Takeaways, Limitations

•
Takeaways:
◦
We present a novel approach to effectively address the FSCIL problem by leveraging large-scale pre-trained generative models.
◦
Implementing an efficient learning process using a fixed backbone and minimal trainable components.
◦
Demonstrated performance superiority through experimental results that outperform existing best-performing methods.
◦
Exploiting representational flexibility through multi-scale representations and text encoders.
•
Limitations:
◦
High reliance on pre-trained text-to-image diffusion models. The performance of the model may be affected by the quality of the pre-trained model.
◦
There may be limitations to preventing production bias by using only a small amount of feature distillation.
◦
Further research is needed on the generalization performance of the proposed method. Performance evaluations on various datasets and situations are needed.
👍
Made with Slashpage