arXiv subjects

Stefano Vergani

Publications and source records attributed to Stefano Vergani.

6 recordsLinked to original sources

arXiv2026

Physics at the Edge: Benchmarking Quantisation Techniques and the Edge TPU for Neutrino Interaction Recognition

This work presents a comprehensive benchmark of different quantisation techniques for convolutional neural networks applied to neutrino interaction recognition. Utilising simulation for a generic liquid argon time-projection chamber, models are quantised and then deployed on the Google Coral Edge TPU. Models are tasked with recognising which neutrino interaction is simulated in the image between neutral current, muon-neutrino charged current, and electron-neutrino charged current. Four Keras models are tested, and accuracy is measured across two different pipelines: using post-training integer quantisation and quantisation-aware training. Inference speed is benchmarked against an AMD EPYC 7763 CPU and NVIDIA A100 GPU. A study of the energy consumption is also presented, with attention to potential costs and environmental issues. Results show that, among the four models tested, accuracy degradation is limited and, in particular, Inception V3 presents almost no accuracy degradation across the two quantisation and deployment pipelines. The speed of the edge TPU is comparable to that of the CPU, and one order of magnitude slower than the GPU. Moreover, the energy consumption of all models deployed on the edge TPU is several orders of magnitude lower than that of the CPU and GPU. In the energy consumption-latency parameter space, CPU, GPU, and edge TPU performances can be clearly separated. This paper explores possible future integrations of edge AI technologies with neutrino physics.

physics.ins-det↗

arXiv2026

STAMP/STPA Informed Characterization of Factors Leading to Loss of Control in AI Systems

A major concern amongst AI safety practitioners is the possibility of loss of control, whereby humans lose the ability to exert control over increasingly advanced AI systems. The range of concerns is wide, spanning current day risks to future existential risks, and a range of loss of control pathways from rapid AI self-exfiltration scenarios to more gradual disempowerment scenarios. In this work we set out to firstly, provide a more structured framework for discussing and characterizing loss of control and secondly, to use this framework to assist those responsible for the safe operation of AI-containing socio-technical systems to identify causal factors leading to loss of control. We explore how these two needs can be better met by making use of a methodology developed within the safety-critical systems community known as STAMP and its associated hazard analysis technique of STPA. We select the STAMP methodology primarily because it is based around a world-view that socio-technical systems can be functionally modeled as control structures, and that safety issues arise when there is a loss of control in these structures.

cs.CY↗

arXiv2025

LArTPC hit-based topology classification with quantum machine learning and symmetry

We present a new approach to separate track-like and shower-like topologies in liquid argon time projection chamber (LArTPC) experiments for neutrino physics using quantum machine learning. Effective reconstruction of neutrino events in LArTPCs requires accurate and granular information about the energy deposited in the detector. These energy deposits can be viewed as 2-D images. Simulated data from the MicroBooNE experiment and a simple custom dataset are used to perform pixel-level classification of the underlying particle topology. Images of the events have been studied by creating small patches around each pixel to characterise its topology based on its immediate neighbourhood. This classification is achieved using convolution-based learning models, including quantum-enhanced architectures known as quanvolutional neural networks. The quanvolutional networks are extended to symmetries beyond translation. Rotational symmetry has been incorporated into a subset of the models. Quantum-enhanced models perform better than their classical counterparts with a comparable number of parameters but are outperformed by classical models, which contain an order of magnitude more parameters. The inclusion of rotation symmetry appears to benefit only large models and remains to be explored further.

physics.ins-det↗

arXiv2023

Dipole-Coupled Neutrissimo Explanations of the MiniBooNE Excess Including Constraints from MINERvA Data

We revisit models of heavy neutral leptons (neutrissimos) with transition magnetic moments as explanations of the $4.8σ$ excess of electron-like events at MiniBooNE. We perform a detailed Monte Carlo-based analysis to re-examine the preferred regions in the model parameter space to explain MiniBooNE, considering also potential contributions from oscillations due to an eV-scale sterile neutrino. We then derive robust constraints on the model using neutrino-electron elastic scattering data from MINERvA. We find that MINERvA rules out a large region of parameter space, but allowed solutions exist at the $2σ$ confidence level. A dedicated MINERvA analysis would likely be able to probe the entire region of preference of MiniBooNE in this model.

hep-ph↗

arXiv2022

A First Application of Collaborative Learning In Particle Physics

Over the last ten years, the popularity of Machine Learning (ML) has grown exponentially in all scientific fields, including particle physics. The industry has also developed new powerful tools that, imported into academia, could revolutionise research. One recent industry development that has not yet come to the attention of the particle physics community is Collaborative Learning (CL), a framework that allows training the same ML model with different datasets. This work explores the potential of CL, testing the library Colearn with neutrino physics simulation. Colearn, developed by the British Cambridge-based firm Fetch.AI, enables decentralised machine learning tasks. Being a blockchain-mediated CL system, it allows multiple stakeholders to build a shared ML model without needing to rely on a central authority. A generic Liquid Argon Time-Projection Chamber (LArTPC) has been simulated and images produced by fictitious neutrino interactions have been used to produce several datasets. These datasets, called learners, participated successfully in training a Deep Learning (DL) Keras model using blockchain technologies in a decentralised way. This test explores the feasibility of training a single ML model using different simulation datasets coming from different research groups. In this work, we also discuss a framework that instead makes different ML models compete against each other on the same dataset. The final goal is then to train the most performant ML model across the entire scientific community for a given experiment, either using all of the datasets available or selecting the model which performs best among every model developed in the community.

hep-ex↗

arXiv2021

Explaining the MiniBooNE Excess Through a Mixed Model of Oscillation and Decay

The electron-like excess observed by the MiniBooNE experiment is explained with a model comprising a new low mass state ($\mathcal{O}(1)$ eV) participating in neutrino oscillations and a new high mass state ($\mathcal{O}(100)$ MeV) that decays to $ν+γ$. Short-baseline oscillation data sets are used to predict the oscillation parameters. Fitting the MiniBooNE energy and scattering angle data, there is a narrow joint allowed region for the decay contribution at 95% CL. The result is a substantial improvement over the single sterile neutrino oscillation model, with $Δχ^2/dof$ = 19.3/2 for a decay coupling of $2.8 \times 10^{-7}$ GeV$^{-1}$, high mass state of 376 MeV, oscillation mixing angle of $7\times 10^{-4}$ and mass splitting of $1.3$ eV$^2$. This model predicts that no clear oscillation signature will be observed in the FNAL short baseline program due to the low signal-level.

hep-ph↗