Gamma Learn Deep Learning applied to the CTA

Gamma. Learn: Deep Learning applied to the CTA data analysis Mikaël Jacquemont for the Gamma. Learn team Meeting Geneva University & LST 1 analysis hands on, 14/02/2020

Deep Learning applied to the CTA data analysis Gamma. Learn project Mikaël Jacquemont thesis • • Patrick Lambert Alexandre Benoit • Thomas Vuillaume • • Gilles Maurin Giovanni Lamanna *ASTERICS: european H 2020 project 2

Deep Learning applied to the CTA data analysis Gamma. Learn’s objectives • Gamma / hadron separation • Energy and direction regression From CTA data with Deep Learning Challenges • Hexagonal pixels • Running / Managing a lot of experiments • Comparing experiments • Computing center compatibility • … 3

Deep Learning applied to the CTA data analysis Challenges Hexagonal pixels • Convolution (& pooling): a key element of DL • Implemented for matrices • Standard methods – Oversampling Credits: T. Holch et al. – Rebinning – Image shifting + masked convolution Ø Preprocessing Ø Image distortion Axial addressing system 1 1 0 1 1 Convolution kernel mask 4

Deep Learning applied to the CTA data analysis Challenges Hexagonal pixels • Indexed. Conv – GEMM for arbitrary lattices • im 2 col operation based on the list of pixel neighbors – Hexagonal case Validated on CIFAR-10 and AID “Indexed operations for non-rectangular lattices applied to convolutional neural networks” VISAPP 2019 5

Deep Learning applied to the CTA data analysis Challenges Running / Managing a lot of experiments Gammalearn Data handlers Tensorboard. X Tensorboard Handlers Raw test data High level library Datasets Ignite Network checkpoints Experiment settings backup and logs Steps Criterions Summary Files ctaplot / Gamma. Board Visualization tools Experiment runner Experiment Optimizers Experiment settings Utils Automatic differentiation tool Network definition 6

Deep Learning applied to the CTA data analysis Challenges Running / Managing a lot of experiments 7

Deep Learning applied to the CTA data analysis Challenges Computing center compatibility • conda environment • Singularity container • Docker container coming soon 8

Deep Learning applied to the CTA data analysis Scientific Contribution • Multitask Architecture for full event reconstruction – Gamma / proton separation – Energy, Direction and Impact point (auxiliary task) • Focus on mono analysis for now – Interesting use-case – Preparation of the LST 1 analysis + Incoming of real data – First step to the stereo analysis • Data – LST 4 monotrigger (gamma and proton events) – Diffuse for training, point like for testing – Cuts • • • Minimum amplitude 300 Leakage 2 < 0, 2 Result of cleaning 9

Deep Learning applied to the CTA data analysis γ-Phys. Net • Physically guided architecture • Indexed. Conv • Uncertainty balancing • Masked loss Energy Global pool Impact flatten Direction Res. Net Particle type FC 10

Deep Learning applied to the CTA data analysis γ-Phys. Net performance Angular resolution --- γ-Phys. Net --- RF Energy resolution 11

Deep Learning applied to the CTA data analysis γ-Phys. Net performance --- γ-Phys. Net --- RF 12

Deep Learning applied to the CTA data analysis γ-Phys. Net performance Angular Resolution --- γ-Phys. Net (2, 8 k. Hz on V 100) --- H. E. S. S. II mono loose cuts --- H. E. S. S. II mono safe cuts Energy Resolution H. E. S. S. II Data Analysis with Im. PACT, Parsons R. et al. ICRC 2015 13

Deep Learning applied to the CTA data analysis Wrap up • Ecosystem for DL with CTA data • Mono analysis with Deep MT Learning – Very interesting results → consistent with template based analysis in H. E. S. S. but much (~1 e 3) faster 14

Deep Learning applied to the CTA data analysis Going further • Optimize γ-Phys. Net • Use these good results in mono for the stereo reconstruction • Test on real data → adaptation needed • Publication 15

Gamma. Learn: DL applied to the CTA data analysis Thank you for your attention Gamma. Learner framework: https: //gitlab. lapp. in 2 p 3. fr/Gamma. Learn Indexed. Conv package: https: //github. com/Indexed. Conv Ctaplot and Gammaboard: https: //github. com/vuillaut/ctaplot -> https: //github. com/cta-observatory/ctaplot 16

Gamma. Learn • Backup 17

Deep Learning for CTA data analysis Data • LST 4 monotrigger (gamma and proton events) • Diffuse for training, point like for testing • Calibrated and integrated images (625 k / 156 k) • Mono analysis • Cuts – Minimum amplitude 300 – Leakage 2 < 0, 2 – Result of cleaning 18

Gamma. Learn • Backup --- γ-Phys. Net --- H. E. S. S. II mono loose cuts --- H. E. S. S. II mono safe cuts 19