Stéphane d'Ascoli | Solving Symbolic Regression with Transformers

Sponsored by Evolution AI - www.evolution.ai/
Speaker: Stéphane d'Ascoli, Ph.D. candidate at Facebook AI
Papers: arxiv.org/abs/...
Abstract: Inferring mathematical laws from experimental data is a central problem in natural science, named symbolic regression. This task is particularly challenging as it cannot be framed as a differentiable minimisation problem, and has until now mainly been solved via genetic programming.
In this talk, Stéphane will show that Transformers pre-trained on vast synthetic datasets are able to solve this task with comparable performance to state-of-the-art genetic algorithms, with orders of magnitude smaller inference time. Stéphane will first demonstrate their capabilities on number sequence problems typical of IQ tests, then show how they can be adapted to real-world data.
Bio: Stéphane d'Ascoli is a Ph.D. candidate working on deep learning, sharing his time between the Center for Data Science of ENS Paris and Facebook AI Research. Prior to that, he studied Theoretical Physics at ENS Paris, and worked with NASA on black hole mergers.