Benjamin Bloem-Reddy

Invariance in machine learning methods: What, why, and then?

Statistical and machine learning methods that incorporate appropriate symmetries have proven hugely successful, from exchangeable probability models and de Finetti's theorem in Bayesian statistics, to convolutional neural networks and their extensions in deep learning. I will give an overview of the what and the why, with an emphasis on recent work and some open problems, and look towards what might be next.