Reproduce “Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingzhuo Li, Guang Li, Jiafeng Mao, Takahiro Ogawa, Miki Haseyama
This research presents a diversity-driven generative dataset distillation method, leveraging a diffusion model with a self-adaptive memory mechanism to create smaller, more representative datasets. The approach consistently achieves superior downstream classification accuracy compared to existing state-of-the-art methods across various ImageNet subsets and images-per-class settings by generating synthetic data with enhanced diversity.
Key insights
- Actively guiding the generative process towards diversity, rather than merely resemblance, is crucial for improving the performance of models trained on distilled datasets.
- A self-adaptive memory system, which efficiently prunes redundant elements, can effectively guide a diffusion model to generate diverse and representative data without direct comparison to the entire large original dataset.
- Combining objectives that both encourage covering the original data distribution (representativeness) and ensuring variety within the generated samples (diversity) leads to more effective dataset distillation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.