Webinar Review: Using AI to Accelerate Scientific Discovery

Our trainees review webinars in their given fields and share abstracts to help colleagues outside their discipline make an informed choice about watching them. As our program bridges diverse disciplines, these abstracts are beneficial for our own group in helping one another gain key knowledge in each other’s fields. We are happy to share these here for anyone else who may find them helpful.

Using AI to Accelerate Scientific Discovery

Part of MIT’s Center for Brains, Minds, and Machines (CBMM) Seminar Series

Demis Hassabis, Founder and CEO of DeepMind

April 11, 2022

Watch on the MIT website >>

Ryan MillerAnalysis by Ryan Miller:

In this talk, Hassabis discusses recent advances in AlphaFold. In the beginning of the talk, the objective of DeepMind is presented as “Solving Intelligence to advance science and benefit humanity.” Hassabis describes that by solving intelligence, he means they would like to fundamentally understand the phenomena of intelligence, and then, of course, recreate that artificially to create artificial general intelligence.

With this framework, they developed what is called deep reinforcement learning, or that the stream of observations from the environment allows for the active reconstruction of a model environment. And with this model the AI system can learn how to achieve a specific goal or maximize a reward. As such, actions by the AI system may lead to new observations as it manipulates or probes the environment.

With this model, DeepMind developed AlphaGo to illustrate the potential of how an AI system can learn to play the game “Go” that has a search space of 10170 positions. To illustrate the power of the AI system a simulation of a game was presented. Moving forward in the talk Hassabis discusses AlphaFold, which marks the transition from leaning and building game intelligence to learning scientific discoveries. Specifically, how the AI system can search the state space and discover proper protein folding confirmations based on amino acid molecular and steric interactions. All in all, the talk was an incredibly unique perspective on predictive modelling and provided exciting future directions of the technology.