This week the VII AMVA4NewPhysics workshop is under way in the premises of LIP in Lisbon. During these events the network gets together to discuss the status of the various projects, plan future events and activities, take action on arisen issues, and vote on budget and other topics. But this is a special event in the lifetime of the network, as we are getting toward the mature stage – we are in the fourth and last year of activity – and many of the projects are either complete or toward completion. So our PhD students have a lot to show, and we built the schedue around their talks, leaving ample space for their presentation and discussion in the agenda.
(Above: network members listen to Alessia Saggio’s presentation)
Among the results being shown, there are three which were described in detail in a document we submitted to the European Commission in real time, as October 31 was the deadline for production of “Deliverable 1.4”, titled “Final report on classification and regression tools for Higgs measurements”. You can download the document here.
The first of the three reported results was a reanalysis of the data made available by the ATLAS collaboration during the 2014 “Higgs kaggle challenge“, where 1750 teams competed to win a 13000 euro prize by classifying data and distinguishing the Higgs decay to tau leptons from backgrounds. Giles Strong applied new ideas on how to optimize neural networks (NN) taken from other domains, and checked what worked best. In the end he was able to obtain a solution which would have won the 2014 challenge, and in the process he learned what really can boost the performance of a neural network applied to a HEP classification problem.
The second result was produced by Pablo de Castro, who produced an innovative algorithm that allows neural networks to directly optimize the final result of a signal search performed through a NN-based classification, in the presence of systematic uncertainties that affect the knowledge of the density functions of the two classes. The algorithm, called “INFERNO” (from “inference aware neural optimization”) has been already described in an article now on the ArXiv. It can improve by large amounts the precision of measurements in the presence of nuisance parameters.
The third result is the cooking of Alexander Held, who applied the Matrix Element Method to the search for associated production of a top quark pair and a Higgs boson in ATLAS data. The rare and complicated process has been successfully extracted and its cross section measured, and results are now public.