Learning from the machine: unsupervised tagging

Febrero 28, 2019
De 3:00pm hasta 4:00pm

IFT Seminar Room/Red Room

Theoretical Physics, general interest
Speaker: 
Jernej F. Kamenik
Institution: 
Jozef Stefan Institute, Ljubljana, Slovenia
Location&Place: 

IFT Seminar Room/Red Room

Abstract: 
Drawing on the relationship between distributions of observables in events, and emergent themes in sets of documents, we can apply generative statistical modelling of jet substructure observables to discriminate between different a-priory unknown underlying short distance physical processes in multi-jet events. As a proof of principle I will discuss an unsupervised top-quark jet tagger, which can be trained on data alone. Additional theoretical input can be supplied in the form of Bayesian prior distributions of pure event samples. I will compare this proposal to existing traditional and machine learning approaches to jet tagging. Finally I will discuss applications to anomaly detection, especially in tails of distributions with relatively scarce training data sets.