Improvements on GFlowNets Applied to Molecular Discovery | Emmanuel Bengio 

Try datamol.io - the open source toolkit that simplifies molecular processing and featurization workflows for machine learning scientists working in drug discovery: datamol.io/
Never miss another M2D2 talk, add the schedule to your calendar: m2d2.io/talks/m2d2/about/
Also consider joining the M2D2 Slack: m2d2group.slack.com/join/shar...
Abstract: In this talk, I’ll discuss recent progress on scaling GFlowNets to solve more complex molecular design problems. I’ll first share a recent refinement in multi-objective optimization using goal-based strategies, and I’ll then share some upcoming work on better training strategies for GFlowNets that have worked in both de-novo discovery and lead optimization. If time allows I’ll finish the talk on some experiments that I hope will convince you that GFlowNets have great potential for scientific discovery & active learning.
Speaker: Emmanuel Bengio - folinoid.com/
Twitter Prudencio: / tossouprudencio
Twitter Jonny: / hsu_jonny
Twitter datamol.io: / datamol_io
~
Chapters:
00:00 - Intro
05:22 - Training & Parameterizing a GFlowNet
10:02 - Multi-Objective GFlowNets
12:10 - Limitations of Scalarisation
14:08 - Goal Conditioned GFlowNets
18:46 - Evaluation Metrics
23:40 - A Learned Focus Model & Results
34:20 - Towards Understanding & Improving GFlowNet Training
41:11 - Understanding GFlowNets on a Minimal Graph Problem
45:57 - Conclusions & Takeaways
50:06 - Q+A

Наука