Abstract:
Symbolic Regression searches over the space of analytical expressions (mix of variables, operators, and constants symbols) to fit experimental data. Symbolic regression is usually very slow compared to more basic regression methods, such as decision-trees or neural networks, preventing its use in settings where inference time is important, e.g. control. In this talk, I will present neural approaches to SR, especially pre-trained models that infer expressions from data in a zero-shot manner. I will also explain why they are not yet reliable, as well as directions we are currently exploring to improve their accuracy on "out-of-domain" datasets.
