Improvements on GFlowNets Applied to Molecular Discovery

Explore recent advancements in GFlowNets for molecular discovery in this 56-minute conference talk by Emmanuel Bengio from Valence Labs. Delve into multi-objective optimization using goal-based strategies and improved training techniques for both de-novo discovery and lead optimization. Gain insights into GFlowNets' potential for scientific discovery and active learning. The talk covers training and parameterizing GFlowNets, multi-objective approaches, limitations of scalarization, goal-conditioned GFlowNets, evaluation metrics, learned focus models, and understanding GFlowNet training through minimal graph problems. Conclude with key takeaways and a Q&A session to deepen your understanding of this cutting-edge approach to molecular design.

- Intro
- Training & Parameterizing a GFlowNet
- Multi-Objective GFlowNets
- Limitations of Scalarisation
- Goal Conditioned GFlowNets
- Evaluation Metrics
- A Learned Focus Model & Results
- Towards Understanding & Improving GFlowNet Training
- Understanding GFlowNets on a Minimal Graph Problem
- Conclusions & Takeaways
- Q+A