Sampling from unknown high-dimensional distributions is an interesting statistical problem that can benefit from machine learning methods. While traditional approaches, such as methods based on Markov chain Monte Carlo, can in principle sample exactly from the distribution, they can become computationally intractable in the presence of well separated modes. The recently introduced GFlowNets tackle this issue by learning policies to sample compositional objects proportionally to their rewards, leveraging neural networks to generalize using the structure in the distribution. This property also makes GFlowNets particularly interesting in realistic scenarios, such as designing biological sequences, for instance DNA and proteins, via active learning. In these applications, it becomes critical to sample candidates that are diverse and cover all the modes of the available reward (in-silico simulation), which potentially does not capture all the desired properties. A diverse sample improves the odds of discovering candidates that can satisfy further screening (in-vitro experiment). In this talk, we will introduce the core theoretical foundations of GFlowNets and illustrate its application on the design of biological sequences. We will provide examples and results of how GFlowNets can be successfully used to find high-reward and diverse biological sequences, and highlight the broader applicability of GFlowNets to various problems.

Alex Hernandez-Garcia is currently a postdoctoral researcher at Mila, the Quebec Artificial Intelligence Institute, in Montréal, working with Prof. Yoshua Bengio since December 2020. Before that, he obtained a PhD in Cognitive Science from the University of Osnabrück in Germany, and master's and bachelor's degrees in engineering at the University Carlos III of Madrid, Spain. His current focus is researching applications of machine learning to fight climate change, such as accelerating material discovery. More broadly, Alex is also interested in brain-inspired machine learning and computational neuroscience.

Moksh Jain is a graduate student at the University of Montréal and Mila, advised by Prof. Yoshua Bengio. His current research focuses on developing deep learning methods for efficient knowledge acquisition in active learning, and their applications to drug discovery. He obtained his bachelor's degree from NITK Surathkal and worked on developing efficient machine learning algorithms for resource-constrained devices at Microsoft.