1st Pan-European Advanced School on Statistics in High Energy Physics

28 Oct 2019, 09:00 → 1 Nov 2019, 14:00 Europe/Berlin

SR 4a/b (DESY Hamburg)

github-com/diana-hep/madminer

We are looking forward to welcome you to the first Pan-European Advanced School on Statistics in High Energy Physics.  The program consists of lectures on advanced or not so frequently covered topics, supplemented by tutorials. The school is open to PhD students and Post-docs. Participants should already have a good knowledge of basic statistical methods in data analysis.

Covered topics:

•  Bayesian inference
•  Machine learning
•  Unfolding
•  Gaussian Processes
•  Non-Parametric Inference

The registration fee is 50 Euro and has to be paid cash.

More information on the INSIGHTS Marie Sklodowska-Curie ITN  can be found here

Vidyo connection  Allen Caldwell whiteboard lecture monday and tuesday: showing camera image of whiteboard, no audio:

https://vidyoportal.cern.ch/join/KzCDUbNKYXFf9o5gj1GRmX37Id4

DESY map with venues https://www.desy.de/~obehnke/paneu/mapstat.pdf

Poster Poster_web.pdf

Monday 28 October

12:30 → 14:00 Registration

14:00 → 18:10 Session 1

14:00 Welcome to School

Speaker: Dr Olaf Behnke (DESY)

Slides programme_ref.key programme_ref.pdf

14:05 Bayesian Inference Lecture I

1) What is probability, leading to derivation of Bayes’ theorem 2) Comparison of Bayes & Frequentist for the Poisson problem with and w/o background, comparison of Bayes with Neyman, Feldman-Cousins 3) Detailed analysis of the on/off problem 4) The role of priors, ways to define them, deciding when they are important

Speaker: Allen Caldwell (MPI Munich)

Class Notes DAMC_script_1-4.pdf

Handout DESY-Handout.pdf

15:30 Coffee Break

16:00 Bayesian Analysis Toolkit Tutorial I

BAT session 1 will be devoted to brief intro to the Julia programming language, installing the package, and solving a simple Poisson problem

Speaker: Dr Oliver Schulz (MPI Munich)

Slides BATjl-Tutorial.pdf

18:30 → 20:00 Reception

Map Auditorium.png

Tuesday 29 October

09:00 → 12:30 Session 2

09:00 Bayesian Inference Lecture II

5) Model testing: the different schools 6) Detailed comparison and discussion of the Jeffreys’Lindley Paradox 7) Approaches to mapping the posterior pdf and our scheme for parallelizing MCMC

Speaker: Allen Caldwell (MPI Munich)

10:30 Coffee Break

11:00 Bayesian Analysis Toolkit Tutorial II

BAT session 2 will look at comparing fitting a spectrum and determining whether or not a signal is present.

Speaker: Dr Oliver Schulz (MPI Munich)

12:30 → 14:00 Lunch Break

14:00 → 18:00 Session 3

14:00 Likelihood Free Inference

General introduction + focus on likelihood-ratio based approaches

Speaker: Gilles Louppe (Liege University)

Slides talk-glouppe-neural-lfi-desy.pdf

15:30 Coffee Break

16:00 Likelihood-free Inference (Passive) Tutorial

Speaker: Gilles Louppe (University of Liège)

16:30 Probabilistic Programming

Lecture and (Active) Tutorial on a high energy physics problem

Speaker: Lukas Heinrich (CERN)

Slides HamburgStatsSchool.pdf

Wednesday 30 October

09:00 → 12:30 Session 4

09:00 Generative Models

Lecture: - Adversarial Training - Generative Models (GAN, WGAN, Autoencoder, Variational Autoencoder) - Bonus topics: Bayesian networks, Autoencoders for anomalies

Speaker: Gregor Kasieczka (Institut fuer Experimentalphysik / UHH)

Slides DeepLearning_PanEuroStat_2019_Generative.pdf

10:30 Coffee Break

11:00 Variational Autoencoder Tutorial

Training generative models on image data

Speaker: Gregor Kasieczka (Institut fuer Experimentalphysik / UHH)

Slides Exercise_Generative_Models_Problem_(1).ipynb links.pdf

12:30 → 14:00 Lunch Break

14:00 → 18:45 Session 5

14:00 Exemplary usages of GANs in the ATLAS experiment

Fast simulation will be a crucial tool for all LHC experiments from Run 3 and especially during HL-LHC. The increase in luminosity will not be matched by a similar increase in computing resources, therefore fast simulation will be the only way forward. Fast simulation is already used by the LHC experiments but it will need to be significantly improved to be used by all analyses. A possible way forward is to exploit the generative tool developed by the ML community, in particular GANs. In this lecture I will present the current state-of-the-art in ATLAS and use it to describe the challenges faced by this approach. Additional examples will be taken from another GAN developed to generate and simulate di-jet and top quark events. While the generation of event is not currently the bottle neck in the MC production chain, it will soon require significant resources as new, more precise but also more complex MC generators will be developed. Keywords: fast calorimeter simulation, fast event generation, GAN optimisation, voxalisation, conditioning

Speaker: Michele Faucci Giannelli (Edinburgh)

Slides Event_generation_and_simulation_using_GANs_-_Michele_Faucci_Giannelli.pptx

15:00 Traditional inference with ML tools I

Speaker: Lukas Heinrich (CERN)

Slides pyhf_desy_stats_school.pdf

15:30 Coffee Break

16:00 Traditional inference with ML tools II

Speaker: Lukas Heinrich (CERN)

16:30 Automated Differentiation

Speaker: Lukas Heinrich (CERN)

Slides grk.pdf

17:15 Bayesian Inference III

Speaker: Allen Caldwell (MPI)

19:30 → 22:00 School Dinner

Thursday 31 October

09:00 → 12:30 Session 6

09:00 Decision making under uncertainties"

In this seminar we present some aspects of the so-called "decision making under uncertainties"  that are typical in the context of business and strategical management, where the data represent  an important source of information that is not always exploited to the full extent of its potential.  In particular, we will describe a few cases taken from the experiences of Pangea Formazione,  where the Bayesian approach and more in general Bayesian probabilistic models, when paired with suitable data science solutions, have allowed to deal properly with uncertainties and to  provide valuable insights about corporate processes.

Speakers: Fabio Priuli (Pangea Formazione S.r.l.), Sara Borroni (Pangea Formazione S.r.l.)

Slides Hamburg_20191031.pdf

09:45 Gaussian Processes for Particle Physicists I

* Definition and basic properties * Mean functions, covariance functions and parameter estimation * Relation to other statistical learning techniques * Applications in physics

Speaker: Prof. Mikael Kuusela (Carnegie Mellon University)

Slides GPs_DESY_Oct_2019.pdf

10:30 Coffee Break

11:00 Gaussian Processes for Particle Physicists II

Speaker: Prof. Mikael Kuusela (Carnegie Mellon University)

11:45 Introduction to Unfolding Methods in HEP I

* Why is unfolding needed and why is it difficult? * Overview of common unfolding methods * Choice of the regularization strength * Uncertainty quantification

Speaker: Prof. Mikael Kuusela (Carnegie Mellon University)

Slides unfolding_DESY_Oct_2019.pdf

12:30 → 13:30 Lunch Break

13:15 → 19:00 Session 7

13:30 Introduction to Unfolding Methods in HEP II

* Why is unfolding needed and why is it difficult? * Overview of common unfolding methods * Choice of the regularization strength * Uncertainty quantification

Speaker: Prof. Mikael Kuusela (Carnegie Mellon University)

14:15 RooFitUnfold Tutorial

Introduction to Unfolding inside RooFit. Includes - preparing inputs - choosing unfolding method, - systematics handling and bias calculation. In this tutorial you learn to set up your analysis using unfolding from start to finish using realistic physics examples.

Speakers: Carsten Burgard (Nikhef), Dr Lydia Brenner (DESY), Mr Pim Verschuuren (Royal Holloway, University of London)

Slides ExercisesRooFitUnfold.pdf RooFitUnfoldIntroTutorialDESY.pdf

15:45 Coffee Break

16:15 RooFitUnfold Tutorial II

Speakers: Carsten Burgard (Nikhef), Dr Lydia Brenner (DESY), Pim Verschuren (Royal Holloway, University of London)

17:30 Approximate Bayesian Computing

This lecture will provide an overview of approximate Bayesian computation (ABC), starting with a focus on the motivation for the procedure, including when it makes sense to utilize it. The resulting approximation can be thought of as calculating the posterior under a contaminated data set; this interpretation provides a useful context for the procedure. The main challenge to using the approach is the computational difficulties, so the lecture will cover algorithms that ease this burden, including consideration of sequential Monte Carlo approaches, and the importance of careful choice of summary statistics.

Speaker: Chad Shafer (CMU)

Slides ABClectureSchafer.pdf

Friday 1 November

09:00 → 12:30 Session 8

09:00 Non-Parametric Inference I

These two lectures will provide an overview of the utility and challenges of nonparametric approaches to statistical inference. Nonparametric inference has great promise to "let data speak for themselves" and enable a researcher to avoid restrictive model assumptions. This can be especially important when massive sets of data need to be summarized in a way that preserved scientific information while not missing important features. Topics covered will include nonparametric density estimation and nonparametric regression. Focus will be placed on the curse of dimensionality, and the crucial role it plays in motivating particular nonparametric approaches. Underlying statistical theory will be presented.

Speaker: Chad Shafer (CMU)

Jupyternotebook NonparametricExamples.ipynb

Slides Material II Nonpar2DESY.pdf

Slides part I Nonpar1DESY.pdf

10:30 Coffee Break

11:00 Non-Parametric Inference II

Speaker: Chad Shafer (CMU)

12:25 Closing of the School

Speaker: Dr Olaf Behnke (DESY)

Slides Goodbye.pdf

12:30 → 14:00 Lunch Break