Learning Generative Models from Corrupted Data
summary
Proposes training diffusion generative models from corrupted examples by learning denoisers via the score–conditional expectation identity, starting with additive Gaussian math then mapping real corruptions into noisy regimes; introduces a consistency objective to recover denoisers for unseen noise times and shows a small clean + large noisy mix often matches full-clean performance. Demonstrations cut memorization, recover latent diffusion quality, and on protein folding a 17M-parameter model beats a 400M baseline with ~30–40% higher diversity on long proteins.
AI-written summary · about
chapters (14)
- 0:00Generative models with corrupted data
- 2:23Training data quality matters
- 8:10Low-quality data challenges
- 11:44Diffusion models primer
- 15:21Denoisers from Gaussian noise
- 23:36Training with varying noise levels
- 31:45Self-consistency for denoisers
- 34:17Clean and noisy data mix
- 39:16Generation and memorization
- 47:33Framework for arbitrary corruptions
- 51:57Protein structure generation
- 59:51Protein generation evaluation
- 1:03:10Method for large language models
- 1:08:32Summary
related talks
01h 21mGenerative AI with Diffusion Models
01h 04mGeneralization in Diffusion Models Arises from Geometry-Adaptive Harmonic Rrepresentations
23mEmbracing Noise- How Data Corruption Can Make Models Smarter
41mBuilding Generative Image & Video models at Scale
53mGeneralization in Robotic Behavior Cloning: Pitfalls and Promises
40mTraining generative models from locally privatized data via entropic optimal transport
53mDesigning Robust Learners
30mRepetita Non Iuvant: Why Generative AI Models Cannot Feed Themselves