Dimensionality reduction and sensitivity improvement for TACTIC Cherenkov data using t-SNE machine learning algorithm

Abstract

Very high energy (VHE) gamma and cosmic rays scatter with the atmospheric nuclei and create an extensive air shower (EAS) of electrons and positrons, which travel with relativistic speeds and induce the medium to emit a flash of Cherenkov light, which is detected by the camera. This indirect method to detect EAS is known as the imaging atmospheric Cherenkov technique (IACT). However, the main challenge in IACT is to segregate the EAS events induced by gamma primaries from huge cosmic ray background events. Since cosmic background events are ∼103 times more abundant than gamma rays, the gamma-cosmic ray discrimination plays a crucial role in evaluating the sensitivity of an IACT telescope, like the TACTIC. There are several methods to reduce the dimensionality of Cherenkov event images, such as principal component analysis (PCA), gamma domain dynamic supercuts and machine learning-based cuts which classify the gamma or hadronic events.In this study, we report applying a novel dimensionality reduction technique t-SNE (t-distributed stochastic neighbour embedding), on the image parameters of EASs. This technique enables us to visualize complex multi-dimensional features of EAS images into a compressed two-dimensional projection while preserving local neighbourhood features of the higher dimension. The efficacy of dimensionality reduction, which in this case is measured in terms of the statistical significance of detecting gamma-ray signal over cosmic ray background, is compared with the significance obtained using the conventional dynamic supercuts method. The application of the t-SNE technique to estimate signals from TACTIC observation data is also presented.