Search the repository

A+ | A- | | SLO | ENG

Keywords: marine biology, machine learning, diagnostics
Published in RUNG: 07.04.2025; Views: 355; Downloads: 0
This document has many files! More...

Abstract: Dark matter (DM), known to be a dominant matter component in the Universe, has been searched for extensively, yet remains undetected. One of the promising avenues of detecting a DM signal is to observe the so called ’DM sub-halos’ within our galaxy. These sub-halos, which are numerous within the Milky Way, are formed by the clustering of DM, as predicted by cosmological simulations, and most of them lack baryonic matter counterparts, making them challenging to detect. How- ever, the annihilation or decay of Weakly Interacting Massive Particles (WIMPs), a leading candidate for DM, within these sub-halos is expected to produce very high-energy (VHE) photons (called gamma-rays) at TeV energies, offering possible indirect DM detection. In this thesis, we focus on the Galactic Plane Survey (GPS) of the Cherenkov Tele- scope Array Observatory (CTAO), an upcoming ground-based gamma-ray obser- vatory, which promises unprecedented sensitivity and resolution in the detection of cosmic gamma-ray sources in the ∼ 30 GeV to ∼ 100 TeV energy range. As dark sub-halos are expected to appear as unidentified (point) sources in the CTAO GPS data, we employ a machine learning (ML)-based approach, the AutoSour- ceID framework, leveraging U-shaped networks (U-Nets) and Laplacian of Gaus- sian (LoG) filter, for automatic source detection and localization, and apply it to simulated GPS data. We establish detection thresholds for U-Nets trained on dif- ferently scaled counts (counts, square root or log of counts) and identify which approach offers best results (in terms of flux sensitivity and location accuracy). Our findings suggest that using log-scaled counts yields a factor of 1.7 lower flux threshold compared to counts alone. In addition, we compare our ML outcomes with traditional methods; however, this comparison is not straightforward, as ML and traditional approaches fundamentally differ in their methodologies and un- derlying assumptions. Nevertheless, The flux threshold obtained using log-scaled counts is comparable to that of the traditional likelihood-based detection method implemented in the Gammapy library, although further study is needed to estab- lish a more definitive comparison. These preliminary results also suggest that the flux threshold for detecting 90% of true sources with the ML approach is approx- imately two times lower than the sensitivity reported for the GPS in the CTAO publication. Although these results are not directly comparable due to differences in methodology, they hint that ML methods may offer superior performance in certain scenarios. Furthermore, we discuss the implications of our results on the sensitivity to DM sub-halos, improving it by a factor of 4, highlighting the possi- bility of detecting at least one sub-halo with a cross section approximately ⟨σv⟩ = 2.4 × 10−23 cm3 /s.
Keywords: Cherenkov Telescope Array Observatory, dark matter, sub-halos, machine learning, gamma-rays, master's thesis
Published in RUNG: 06.09.2024; Views: 1330; Downloads: 19
Full text (5,39 MB)

Abstract: This paper is about the first automated systems developed to recognize identity. While automated recognition in the twenty-first century is widely associated with images of the human face, its roots are to be found in attempts to visualize identity in other, non-figural types of trace left by human bodies, ranging as widely as shadows, astrological signs, handwriting, the prints left by palms and fingers and the acoustics of the human voice. This paper investigates one such system of recognition as it emerged from within the telecommunications industry context in the midcentury U.S. Ostensibly built to reduce human labor and cable bandwidth, Bell Labs developed three different phone devices in the 1950s to photograph, formalize and analyze the sounds of speech as they traveled through the telephony system. And while the device called “Audrey” indeed succeeded in recognizing spoken digits, it was its failure to recognize the speech contents without prior awareness of the identity of the speaker, that is to distinguish between the individuality of the speaking “medium” and their intended meaning, that arguably made the experiment a landmark in the history of machine-driven recognition. Accounting for the “noise” made by the body and the environment from which sound emanated into the device, which the lab’s technicians defined as ranging from “speech defects” to “inflection” and “background interference” proved more important than phonetic analysis in determining the intended message of given speech spectogram. Similarly to a range of experiments with noise by formalist filmmakers such as Tony Conrad, John Cage, Kurt Kren and others, it was on the principle of contingency and irreproducible uniqueness that Bell Lab technicians sought to train machine-driven intelligence.
Keywords: History of computer science, machine learning, Bell Labs, history of telecommunications, sound studies
Published in RUNG: 19.02.2024; Views: 2193; Downloads: 9
Full text (31,80 MB)

Abstract: We present a method based on the use of Recurrent Neural Networks to extract the muon component from the time traces registered with water-Cherenkov detector (WCD) stations of the Surface Detector of the Pierre Auger Observatory. The design of the WCDs does not allow to separate the contribution of muons to the time traces obtained from the WCDs from those of photons, electrons and positrons for all events. Separating the muon and electromagnetic components is crucial for the determination of the nature of the primary cosmic rays and properties of the hadronic interactions at ultra-high energies. We trained a neural network to extract the muon and the electromagnetic components from the WCD traces using a large set of simulated air showers, with around 450 000 simulated events. For training and evaluating the performance of the neural network, simulated events with energies between 10^18.5 eV and 10^20 eV and zenith angles below 60 degrees were used. We also study the performance of this method on experimental data of the Pierre Auger Observatory and show that our predicted muon lateral distributions agree with the parameterizations obtained by the AGASA collaboration.
Keywords: Pierre Auger Observatory, indirect detection, surface detection, ground array, ultra-high energy, cosmic rays, muons, machine learning, recurrent neural network
Published in RUNG: 04.10.2023; Views: 2769; Downloads: 9
Full text (1,08 MB)
This document has many files! More...

Abstract: Mass composition anisotropy is predicted by a number of theories describing sources of ultra-high-energy cosmic rays. Event-by-event determination of a type of a primary cosmic-ray particle is impossible due to large shower-to-shower fluctuations, and the mass composition usually is obtained by averaging over some composition-sensitive observable determined independently for each extensive air shower (EAS) over a large number of events. In the present study we propose to employ the observable ξ used in the TA mass composition analysis for the mass composition anisotropy analysis. The ξ variable is determined with the use of Boosted Decision Trees (BDT) technique trained with the Monte-Carlo sets, and the ξ value is assigned for each event, where ξ=1 corresponds to an event initiated by the primary iron nuclei and ξ=−1 corresponds to a proton event. Use of ξ distributions obtained for the Monte-Carlo sets allows us to separate proton and iron candidate events from a data set with some given accuracy and study its distributions over the observed part of the sky. Results for the TA SD 11-year data set mass composition anisotropy will be presented.
Keywords: Telescope Array, indirect detection, surface detection, ground array, ultra-high energy, cosmic rays, composition, anisotropy, machine learning, boosted decision tree
Published in RUNG: 04.10.2023; Views: 2114; Downloads: 8
Full text (1,14 MB)
This document has many files! More...