Can Humans’ Verification of Synthetic Data Save Models from Collapse?
summary
Shows that retraining only on human-verified synthetic examples can avoid the known collapse from unfiltered self-generated data by proving in a Gaussian linear regression model that filtered retraining converges toward the human knowledge center and can reduce MSE for a few rounds; experiments include VAE on MNIST (500 seed images improved after 40 rounds) and fine-tuning SmolLM2-135M on XSUM, both confirming practical gains when filtering quality is sufficiently high.
AI-written summary · about
chapters (12)
- 0:00Synthetic data filtering research
- 1:01AI-generated content future
- 3:57Model collapse from training
- 7:08Human feedback for filtering
- 10:21Gaussian linear regression setup
- 16:24Human knowledge filtering criteria
- 23:50First retraining improves estimation
- 29:53Retraining converges to human knowledge
- 35:31VAEs on MNIST experiments
- 37:42Filtered summaries language generation
- 40:31Filtering synthetic data conclusion
- 40:52Q&A
related talks
30mRepetita Non Iuvant: Why Generative AI Models Cannot Feed Themselves
30m🤯 No data? No problem! Synthetic data to the rescue
45mSynthetic Data for Machine Learning Applications
31mReliable Data for Large ML Models: Principles and Practices
20mSynthetic Data: The new Super Power for AI
28m🧼 From GPU-poor to data-rich: data quality practices for LLM fine-tuning
28mWhy we do ML model retraining wrong, and how to do better
25mWhen Your Dataset Has Blind Spots LLM-Based Data Augmentation