Скачать или смотреть UoA ML Seminar: Alexandre Benoit - Analyzing gamma-ray astronomy data with deep learning

Machine Learning Seminar by Dr. Alexandre Benoit – Analyzing gamma-ray astronomy data with deep learning and first steps towards explainability


Short Bio

Alexandre received PhD degree in electronics and computer science from the University of Grenoble, INP in 2007 (France). Starting 2008, he has been an associate professor at Université Savoie Mont Blanc at LISTIC lab. In 2021, he reached a full professor position at LISTIC lab. His main research interest is related to image and time series understanding. He develops standard and deep learning approaches for a variety of application domains: remote sensing, multimedia and astrophysics. He develops specific approaches adapted to the sensor and data as well as visual perception models.

Abstract

In gamma-ray astronomy, the analysis of the images produced by Cherenkov telescopes consists in separating the gamma events from the background (cosmic rays), and reconstructing the parameters of the detected gammas. The first step is complex because cosmic rays can generate very similar images and the signal-to-noise ratio is typically lower than 1/1000. Besides, CTA, the next generation of observatories, will improve the sensitivity by an order of magnitude, with the counterpart of generating PB of data each year. As a result, standard analysis methods either are too slow or lack sensitivity at low energies.

In this talk, we present a deep multitask architecture, named γ-PhysNet, that outperforms a standard method relying on the Hillas parameter extraction and a multivariate method. Specifically, it achieves very interesting sensitivity below 200 GeV, and could enhance the study of transient phenomena. γ-PhysNet is also 800 times faster than the state-of-the-art method. While analysis models are trained on simulated data, we also show that γ-PhysNet obtains better results on the first exploitable data provided by the Large-Sized Telescope prototype of CTA. Finally, we present first steps towards the explainability of the model’s decisions.


https://ml.auckland.ac.nz/machine-lea...