扩散概率模型：理论与应用-英-43页_1mb

Diffusion Probabilistic Models Summary

1. Models and Theory

• Foundational Works:
  • Ho et al. (Neurips 2020): Denoising Diffusion Probabilistic Models (DDPM)
  • Song et al. (ICLR 2021): Score-based generative modeling through stochastic differential equations
• Analytic Solutions:
  • Bao et al. (ICLR 2022): Announce analytic estimate for optimal reverse variance
  • Bao et al. (ICML 2022): Propose method for optimal covariance estimation with imperfect mean

2. Mathematical Foundations

• Diffusion Process: Describes data through a Markov chain with gradual noise injection.
  • Formula: Data evolves as: ρ₀, ρ₁, ..., ρₜ₋₁ || ρₜ₁₂...₁₂... |ρ<ēₛ₊₁
• Reverse Process: Achieved by minimizing Kullback-Leibler divergence via maximizing Evidence Lower Bound (ELBO).
• Optimal Solutions:
  • For scalar variance and diagonal covariance, specific analytical forms for optimum parameters Σ* and σ*_t are derived.

3. Training and Implementation

• Training involves optimizing noise prediction networks:
  • Initial focus on predicting noise ∇logp(x)
  • Extended to predict squared noise and noise residuals for optimized covariance estimation.
• Application in conditional models and discriminative techniques.

4. Applications

• Conditional Generation: Paired data scenarios, e.g., image synthesis from text or labels.
• Discriminative Guidance: Scaling influence of discriminators within diffusion processes.
• Energy Guidance: Incorporating domain-specific energy functions (e.g., CLIP-based similarity for text-image alignment).
• Unpaired Data: Constructing conditional distributions via energy functions.

5. Other Domains and Science

• Extensions: Text-to-speech (Grad-TTS), video generation.
• Scientific Applications: Anomaly detection in cyber-physical systems, molecular dynamics, Alzheimer’s classification.

6. Downstream Tasks

• Label-Efficient Segmentation: Use DPM features with minimal labeled data.
• Domain-Specific DPMs: Adapt DPM features for tasks like 3D point cloud completion.

7. Conclusion

Bao's work presents advances in optimizing and applying Diffusion Probabilistic Models across diverse applications and domains, combining theory with practical implementations for tasks ranging from unconditional generation to scientific applications.