Reproduce “Diffusion Model as Representation Learner” 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

Xingyi Yang, Xinchao Wang

This paper investigates the capacity of Diffusion Probabilistic Models (DPMs) to learn meaningful representations for recognition tasks by establishing their theoretical connection to denoising autoencoders. It proposes RepFusion, a knowledge transfer method that leverages reinforcement learning to dynamically select optimal time steps for feature extraction, demonstrating improved performance in image classification, semantic segmentation, and facial landmark detection.

Key insights

  • Diffusion Probabilistic Models (DPMs) can be formally viewed as specialized denoising autoencoders (DAEs), providing a theoretical basis for their representation learning capabilities.
  • Intermediate time steps within the diffusion process often yield the most informative and discriminative representations for downstream tasks, suggesting an optimal balance between noise and data structure.
  • A reinforcement learning approach can adaptively select the optimal time step for feature extraction from a diffusion model, leading to improved performance compared to fixed or random selection strategies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.