ONLINE COLLOQUIUM: "Anomaly detection & event generation using (variational) autoencoder models" by Luc Hendriks (RU Nijmegen)
Friday, 20 November 2020 from to (Europe/Amsterdam)
Autoencoders and variational autoencoders are currently a popular choice for solving anomaly detection problems. I will discuss how these models work and how we successfully applied them to detect a possible beyond the Standard Model signal from the Standard Model background, by combining autoencoder techniques with more traditional outlier detection methods such as Gaussian Mixture Models and Isolation Forests. Additionally, these models can be used as generative models as well. In this talk, I will also discuss how we have applied a specific variational autoencoder technique that can generate collider events several orders of magnitude faster than the simulation.