Emmanuel Bengio

Timeline

1. 2011–2016

   Education and conditional computation

   He holds a computer science BSc from Université de Montréal's DIRO department and studies there from 2011 to 2014. In 2015, he posts Conditional Computation in Neural Networks for faster models with Pierre-Luc Bacon, Joelle Pineau and Doina Precup; the paper uses reinforcement learning to optimize policies that drop blocks of units from neural networks. He studies for a CS master's at McGill from 2014 to 2016 under Pineau and Precup, with a thesis on conditional computation.

2. 2016–2022

   PhD and memorization study

   He pursues a PhD in computer science at McGill from 2016 to 2022 under Joelle Pineau and Doina Precup, with a thesis on generalization and optimization in deep reinforcement learning. In 2017, he is one of eleven authors of A Closer Look at Memorization in Deep Networks, which finds that deep networks can memorize noise data and tend to learn simple patterns first.

3. 2020

   Interference in TD learning

   In 2020, he publishes Interference and Generalization in Temporal Difference Learning with Joelle Pineau and Doina Precup at the 37th International Conference on Machine Learning. The paper finds that temporal-difference learning easily leads to low-interference, under-generalizing parameters.

4. 2021

   Introducing GFlowNets

   In June 2021, he posts Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation with Moksh Jain, Maksym Korablyov, Doina Precup and Yoshua Bengio; accepted at NeurIPS 2021, it models generation as a flow network and learns a stochastic policy that samples diverse objects such as molecular graphs in proportion to reward. In November 2021, he is one of six authors of GFlowNet Foundations, which develops further theoretical properties of GFlowNets.

5. 2022

   Joining Valence Labs

   In 2022, he joins Valence Labs at Recursion in Montréal as a staff ML scientist. He works on GFlowNets, protein models and drug discovery.

6. 2025

   Three papers at ICML

   In 2025, he coauthors three papers in the Proceedings of the 42nd International Conference on Machine Learning: Improved Off-policy Reinforcement Learning in Biological Sequence Design, Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation, and Random Policy Evaluation Uncovers Policies of Generative Flow Networks.

7. July 2026

   AquaGen molecular sampling

   In July 2026, he is first author of AquaGen, a generative model that produces all-atom, explicit-solvent molecular configurations from the Boltzmann distribution. It estimates absolute hydration free energy 4–10 times faster than standard GPU-based molecular dynamics with comparable accuracy. He and Sanjeev Raja describe the model in Valence Labs' Inside Valence series.

8. As of October 2026

   Research at LawZero

   He is a machine-learning scientist at LawZero and an associate industry member at Mila. His research topics include deep learning, generative models, GFlowNets, molecular modeling and reinforcement learning.