GammaLearn Results & Publications

Main Results

GammaPhysNet network architecture

The GammaPhysNet architecture

The GammaPhysNet architecture (Jaquemont et al 2021) has been developed specifically for the analysis of Cherenkov Telescope Array (CTA) data. It is a multi-task architecture with attention mechanism composed of a convolutional encoder and a multi-task decoder. Therefore, all the main event reconstruction tasks are performed simultaneously. This presents several advantages:

  • simplificity of use: only one network to train and to use for all the tasks
  • better generalization and performances: the network learns to extract the relevant information for all the tasks at the same time
  • less computational time and memory consumption: only one network to train and to use for all the tasks
  • less degenaracy between tasks as they are performed simultaneously and share the same encoder

Application to Monte-Carlo simulations

The efficiency of the GammaPhysNet architecture has been proven on Monte-Carlo simulations of the CTA LST-1 telescope (Jaquemont et al 2021, Vuillaume et al 2021).

Comparison of instrument response functions and performances on simulated data between the reference (Hillas+RF) and our network 𝛾-PhysNet DA. Top-left panels: angular resolution, energy resolution and energy bias (lower is better). Bottom-left: Sensitivity ratio (higher is better). Top-right: effective area (higher is better). Bottom-right: ROC curve (higher auc score is better).

Application to real data

The Crab Nebula

Application to real data has been done on CTA LST-1 data (Jacquemont et al 2021, Vuillaume et al 2021) showing improvements in sensitivity compared to the standard analysis. This was the first time that a full-event reconstruction was performed on real data with a deep learning architecture, and the first time that sensitivity improvement was achieved on real data. To achieve this result, a careful modulation of the background noise in the MC simulations was necessary to match the real data night sky background (NSB).

Reconstruction Excess Significance Background counts
Hillas+RF 379 γ 12.0 σ 308
Hillas+RF+Poisson noise 376 γ 11.9 σ 305
γ-PhysNet 395 γ 12.5 σ 302
γ-PhysNet+Poisson noise 476 γ 14.3 σ 317
Excess and significance results for Crab Nebula.

Mrk 501

An analysis of Mrk 501 data has also been performed (Vuillaume et al 2021) showing here again an increased sensitivity, but with a greater bias in the reconstructed directionß.

Smoothed (kernel=0.12°) count map of gamma-like events around Markarian 501.
Reconstruction Excess Significance Background counts
Hillas+RF 148.7 𝛾 7.6 𝜎 238.3
𝛾-PhysNet DA 192.7 𝛾 9.8 𝜎 226.3
Excess and significance results for Mrk 501.

Application to real data with domain adaptation

In order to be less sensitive to differences between Monte-Carlo simulations and real data (NSB, broken pixels, telescope PSF…), domain adaptation techniques have been applied. These techniques consist in including unlabeled real data in the training process in order to increase the network robustness to the differences between the training and the test data. Validation of different approaches and architecture has been shown on modified Monte-Carlo simulations (Dell’aiera et al 2023).

The IRFs obtained from the experiments using 𝛾-PhysNet (GPN), DANN, DeepJDOT and DeepCORAL. They are averaged on five seeds. The min and the max are displayed as a surface area. Top left is the energy resolution (lower is better). Top right is the energy bias (lower is better). Bottom left is the angular resolution (lower is better). Bottom right is the AUC per energy bin (high is better). All metrics are function of the gamma-ray true energy. For the energy and angular resolutions, the lower the better whereas for the AUC, the higher the better. An energy bias equals to zero is best.

Peer-reviewed publications:

  • Dell’aiera, M., Jacquemont, M., Vuillaume, T., and Benoit, A., Deep unsupervised domain adaptation applied to the Cherenkov Telescope Array Large-Sized Telescope, International Conference on Content-Based Multimedia Indexing (CBMI), 2023
    Open-source version: arXiv e-prints, 2023. doi:10.48550/arXiv.2308.12732.

  • M. Jacquemont, T. Vuillaume, A. Benoit, G. Maurin, P. Lambert and G. Lamanna, First Full-Event Reconstruction from Imaging Atmospheric Cherenkov Telescope Real Data with Deep Learning, 2021 International Conference on Content-Based Multimedia Indexing (CBMI), 2021, pp. 1-6, doi: 10.1109/CBMI50038.2021.9461918
    Open-source version:

  • Jacquemont, M.; Vuillaume, T.; Benoit, A.; Maurin, G. and Lambert, P. (2021). Multi-Task Architecture with Attention for Imaging Atmospheric Cherenkov Telescope Data Analysis. In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 5: VISAPP, ISBN 978-989-758-488-6 ISSN 2184-4321, pages 534-544. DOI: 10.5220/0010297405340544
    Open-source version:

  • Jacquemont M., Vuillaume T., Benoit A., Maurin G., Lambert P. (2021) Deep Learning for Astrophysics, Understanding the Impact of Attention on Variability Induced by Parameter Initialization. In: Del Bimbo A. et al. (eds) Pattern Recognition. ICPR International Workshops and Challenges. ICPR 2021. Lecture Notes in Computer Science, vol 12663. Springer, Cham.
    Open-source version:

  • Jacquemont, M.; Antiga, L.; Vuillaume, T.; Silvestri, G.; Benoit, A.; Lambert, P. and Maurin, G. (2019). Indexed Operations for Non-rectangular Lattices Applied to Convolutional Neural Networks. In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 5 VISAPP: VISAPP, ISBN 978-989-758-354-4, ISSN 2184-4321, pages 362-371. DOI: 10.5220/0007364303620371
    Open-source version:

  • Mikaël Jacquemont. GAMMALEARN: DEEP LEARNING APPLIED TO THE CHERENKOV TELESCOPE ARRAY (CTA). ICVSS 2018 - Computer Vision after Deep Learning, Jul 2018, Punta Sampieri, Italy. ⟨hal-01841581v2⟩

PhD thesis:

  • Mikaël Jacquemont. Cherenkov Image Analysis with Deep Multi-Task Learning from Single-Telescope Data. Machine Learning [cs.LG]. Université Savoie Mont Blanc, 2020. English. ⟨tel-03148673⟩


  • Pietro Grespan, Mikael Jacquemont, Rubèn López-Coto, Tjark Miener, Daniel Nieto-Castaño, Thomas Vuillaume et al Deep-learning-driven event reconstruction applied to simulated data from a single Large-Sized Telescope of CTA,
    37th International Cosmic Ray Conference. 12-23 July 2021. Berlin, Germany - [arxiv](

  • Vuillaume T., Jacquemont M., de Bony de Lavergne M., Sanchez D.~A., Poireau V., Maurin G., Benoit A., et al., Analysis of the Cherenkov Telescope Array first Large-Sized Telescope real data using convolutional neural networks, arXiv e-prints, 2021.

  • Mikaël Jacquemont, Thomas Vuillaume, Alexandre Benoît, Gilles Maurin, Patrick Lambert. Single Imaging Atmospheric Cherenkov Telescope Full-Event Reconstruction with a Deep Multi-Task Learning Architecture. Astronomical Data Analysis Software and Systems ADASS XXX, Nov 2020, Granada, Spain. ⟨hal-03043005⟩

  • D. Nieto, A. Brill, Q. Feng, M. Jacquemont, B. Kim, et al. Studying deep convolutional neural networks with hexagonal lattices for imaging atmospheric Cherenkov telescope event reconstruction. 36th International Cosmic Ray Conference, Jul 2019, Madison, United States. pp.753, ⟨10.22323/1.358.0753⟩. ⟨hal-02440010⟩

  • Mikaël Jacquemont, Thomas Vuillaume, A Benoit, Gilles Maurin, Patrick Lambert, et al. GammaLearn: a Deep Learning framework for IACT data. 36th International Cosmic Ray Conference, Jul 2019, Madison, United States. pp.705. ⟨hal-02197399⟩

  • Vuillaume, T., “GammaLearn - first steps to apply Deep Learning to the Cherenkov Telescope Array data”, in European Physical Journal Web of Conferences, 2019, vol. 214. doi:10.1051/epjconf/201921406020.