Reproduce “Steering Diffusion Models via Class-Contrastive Influence for Few-Shot Medical Classification” 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
Jeeyung Kim, Erfan Esmaeili, Qiang Qiu
The authors develop C2I, which fine-tunes a Stable Diffusion 2.1 generator with reinforcement learning, rewarding generated batches whose classifier-loss gradients align more with same-class validation examples than with other classes. In few-shot MedMNIST classification, C2I augmentation gives ViT-B/16 a mean AUC of 0.894 versus 0.885 for RandomErasing across three tasks. Gains do not establish real-world robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.