Modern Bayesian Experimental Design
summary
Presents concrete methods: learn neural surrogates to estimate the intractable expected information gain (using nested Monte Carlo and variational upper/lower bounds), train an offline policy network called deep adaptive design to output designs in a single forward pass, and package density estimators with tools like Jax/PyTorch and BFlow using normalizing flows and consistency models. Empirically the policy matches or exceeds a tuned Bayesian baseline, yields tighter EIG bounds (≈3.5 nats on the toy), and runs in real time after offline training.
AI-written summary · about
chapters (12)
- 0:00Modern Bayesian design intro
- 1:29Why design matters
- 4:19Traditional Bayesian design challenges
- 9:22Variational computation speedup
- 13:20Optimizing designs with bounds
- 16:11Policy-based adaptive design
- 21:35Implicit models with neural nets
- 22:59Real-time adaptive design
- 27:14Advanced methods beyond learning
- 28:20Limitations and future directions
- 32:15Conclusion
- 32:57Q&A
related talks
32mDesigning Adaptive Experiments For Policy Learning And Inference
34mBayesian Optimization and it's application to Neural Networks
28mBayesian Optimization- Fundamentals, Implementation, and Practice
01h 30mBayBE: A Bayesian Back End for Experimental Planning in the Low-To-No-Data Regime
53mOff-policy Policy Optimization
31mCost-effective data annotation with Bayesian experimental design
25mScaling Bayesian Experimentation
30mTalk by Suhas Diggavi (UCLA)