Phenp-Coffee Talk

Speaker: Vishal Singh Ngairangbam (Durham University, U.K.)

Date and time: Today, 11th December, Thursday, at 2:30 p.m.

Venue: LH-3

Title: Theory-priors for deep learning in collider phenomenology

Abstract: We live in an era of disruptive AI technologies, making waves in every domain, from business and technological applications to their usage across virtually every scientific discipline, including fundamental fields like mathematics and physics. This is primarily fueled by a class of neural networks popularly known as deep-learning algorithms, which efficiently leverage various domain knowledge (or inductive biases). In this talk, I’ll discuss these algorithms specifically in the context of collider phenomenology. We will first touch upon Convolutional Neural Networks (CNNs), a quintessential class of deep learning algorithms, and explore their relation to the structure of perturbative QCD radiation patterns. We will then discuss some of their drawbacks, which highlight the naturalness of the point cloud representation for jets and events. This representation allows for a direct connection between architecture design-principles and underlying first-principle requirements like infrared and collinear safety. Finally, we will move on to discussing the crucial role of symmetries (or equivariance in machine learning parlance) and their connection to the matrix-element method.