Deep Unsupervised Domain Adaptation for the Cherenkov Telescope Array

Deep learning thesis applied to astrophysics (GammaLearn)

Author	Under the supervision of
Michaël Dell'aiera	Thomas Vuillaume (LAPP) Alexandre Benoit (LISTIC)

michael.dellaiera@lapp.in2p3.fr

Presentation outline

Introduction to gamma-ray astronomy, principle of detection and workflow

Domain adaptation

Application to simulated data

Conclusion, perspectives

Introduction to gamma-ray astronomy, principle of detection and workflow

Gamma-ray astronomy

Study of the **high-energy gamma-ray sources** in the Universe.

The Cherenkov Telescope Array (CTA): * Next generation of gamma-ray observatory * First Large-Sized Telescope (LST-1) prototype is operational \ → No stereoscopy in this work

Study themes: * Mechanisms for particle accelerations * Accelerated particles as feedback on star formation / galaxy evolution * Physical processes at work close to neutron stars / black holes * Fundamental physics and cosmology

Principle of detection

γ-PhysNet

The challenging application to real observations

No ground truth → Training based on simulations (Monte Carlo, or MC)

Simulations are only close approximations of the real data.

A gamma event.

A gamma event with higher background noise and dead pixels (white).

The challenging application to real observations

Classification power with differing noise levels.
DOI 2203.05315

Count map of gamma-like events around Markarian 501.
The reconstructed position is not centered on the actual source.
DOI 2108.04130

→ Distribution discrepancies strongly affect performances.

Domain adaptation

Set of algorithms and techniques which aims to reduce domain discrepancies.

**Advantages:** - Very close to the data - Takes into account unknown differences → Robustly minimises discrepancies between MC and real data

**Unsupervised Domain Adaptation:** - Source are labelled training data\ → Simulated data - Target are unlabelled training data\ → Real data

State-of-the-art

[DOI 2009.00155](https://arxiv.org/abs/2009.00155)

Selection of the methods

* Domain Adversarial Neural Network (DANN)
[DOI 1505.07818](https://arxiv.org/abs/1505.07818)

* Deep Joint Distribution Optimal Transport (DeepJDOT)
[DOI 1803.10081](https://arxiv.org/abs/1803.10081) * Deep Correlation Alignment (DeepCORAL)
[DOI 1607.01719](https://arxiv.org/abs/1607.01719) * Deep Adaptation Networks (DAN)
[DOI 1502.02791](https://arxiv.org/abs/1502.02791)

* Well-known approaches * GAN-based methods are not suitable due to label shifts * Self-supervised architecture: to be tested soon

Validation pipeline of our approach

Validation of the methods

* Controlled perturbations on the simulated (labelled) datasets * Source = MC, Target = MC + perturbations
→ Validation with figures of merit (focus of this presentation)

Tests on real telescope acquisitions

* Source = MC, Target = Real data
→ Detection of known gamma-ray sources (future work)

Application to simulated data

Upper and lower bounds

Best scenario

Training distribution = Test distribution

→ Perfect test conditions \ → The best performance to be observed

Worst scenario

Training distribution ≠ Test distribution

→ Degraded test conditions \ → The worst performance to be observed

Goal: * To be as close as possible to the best scenario * With degraded test conditions * Using domain adaptation

Figures of merit for model evaluation

The resolution is computed per energy bin. Standard metric in gamma-ray astronomy. Lower is better.

Figures of merit for model evaluation

The bias is computed per energy bin. Standard metric in gamma-ray astronomy. Closer to '0' is lower bias, and therefore is better.

Figures of merit for model evaluation

The AUC is computed per energy bin. higher is better.

Setup

Train		Test
Source Labelled	Target Unlabelled	Unlabelled
MC	MC+Poisson(0.4) (MC*)	MC+Poisson(0.4) (MC*) Use labels to characterize domain adaptation

Application to MC: γ-PhysNet + DANN

Application to MC

AUC: Higher is better

Application to MC

AUC: Higher is better

Application to MC

Energy resolution: Lower is better

Application to MC

Energy resolution: Lower is better

Application to MC

Energy bias: Closer to '0' is lower bias, and therefore is better.

Application to MC

Energy bias: Closer to '0' is lower bias, and therefore is better.

Application to MC: Angular resolution

Augular resolution: Lower is better

Application to MC: Angular resolution

Augular resolution: Lower is better

Conclusion & Perspectives

Introduction of domain adaptation to tackle domain discreprencies on astronomical similations * Integration of domain adaptation in a multi-task framework * Successfully applied to LST-1 Monte-Carlo simulated data * Validation of domain adaptation in the context of simulated LST data * Application to real telescope acquisitions (ongoing), many challenges: * Label shifts * Moonlight condition

Acknowledgments

- This project is supported by the facilities offered by the Univ. Savoie Mont Blanc - CNRS/IN2P3 MUST computing center - This project was granted access to the HPC resources of IDRIS under the allocation 2020-AD011011577 made by GENCI - This project is supported by the computing and data processing ressources from the CNRS/IN2P3 Computing Center (Lyon - France) - We gratefully acknowledge the support of the NVIDIA Corporation with the donation of one NVIDIA P6000 GPU for this research. - We gratefully acknowledge financial support from the agencies and organizations listed [here](https://www.cta-observatory.org/consortium\_acknowledgment). - This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 653477 - This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 824064