Searcharxiv⌕ Search

arXiv subjects

Riccardo Di Sipio

Publications and source records attributed to Riccardo Di Sipio.

12 recordsLinked to original sources
arXiv2025

Toward Ethical AI Through Bayesian Uncertainty in Neural Question Answering

We explore Bayesian reasoning as a means to quantify uncertainty in neural networks for question answering. Starting with a multilayer perceptron on the Iris dataset, we show how posterior inference conveys confidence in predictions. We then extend this to language models, applying Bayesian inference first to a frozen head and finally to LoRA-adapted transformers, evaluated on the CommonsenseQA benchmark. Rather than aiming for state-of-the-art accuracy, we compare Laplace approximations against maximum a posteriori (MAP) estimates to highlight uncertainty calibration and selective prediction. This allows models to abstain when confidence is low. An ``I don't know'' response not only improves interpretability but also illustrates how Bayesian methods can contribute to more responsible and ethical deployment of neural question-answering systems.

cs.CL↗
arXiv2025

Rethinking LLM Training through Information Geometry and Quantum Metrics

Optimization in large language models (LLMs) unfolds over high-dimensional parameter spaces with non-Euclidean structure. Information geometry frames this landscape using the Fisher information metric, enabling more principled learning via natural gradient descent. Though often impractical, this geometric lens clarifies phenomena such as sharp minima, generalization, and observed scaling laws. We argue that curvature-based approaches deepen our understanding of LLM training. Finally, we speculate on quantum analogies based on the Fubini-Study metric and Quantum Fisher Information, hinting at efficient optimization in quantum-enhanced systems.

cs.CL↗
arXiv2025

The Curved Spacetime of Transformer Architectures

We present a geometric framework for understanding Transformer-based language models, drawing an explicit analogy to General Relativity. Queries and keys induce an effective metric on representation space, and attention acts as a discrete connection that implements parallel transport of value vectors across tokens. Stacked layers provide discrete time-slices through which token representations evolve on this curved manifold, while backpropagation plays the role of a least-action principle that shapes loss-minimizing trajectories in parameter space. If this analogy is correct, token embeddings should not traverse straight paths in feature space; instead, their layer-wise steps should bend and reorient as interactions mediated by embedding space curvature. To test this prediction, we design experiments that expose both the presence and the consequences of curvature: (i) we visualize a curvature landscape for a full paragraph, revealing how local turning angles vary across tokens and layers; (ii) we show through simulations that excess counts of sharp/flat angles and longer length-to-chord ratios are not explainable by dimensionality or chance; and (iii) inspired by Einstein's eclipse experiment, we probe deflection under controlled context edits, demonstrating measurable, meaning-consistent bends in embedding trajectories that confirm attention-induced curvature.

cs.LG↗
arXiv2021

The Dawn of Quantum Natural Language Processing

In this paper, we discuss the initial attempts at boosting understanding human language based on deep-learning models with quantum computing. We successfully train a quantum-enhanced Long Short-Term Memory network to perform the parts-of-speech tagging task via numerical simulations. Moreover, a quantum-enhanced Transformer is proposed to perform the sentiment analysis based on the existing dataset.

cs.CL↗
arXiv2019

Unfolding as Quantum Annealing

High-energy physics is replete with hard computational problems and it is one of the areas where quantum computing could be used to speed up calculations. We present an implementation of likelihood-based regularized unfolding on a quantum computer. The inverse problem is recast in terms of quadratic unconstrained binary optimization (QUBO), which has the same form of the Ising hamiltonian and hence it is solvable on a programmable quantum annealer. We tested the method using a model that captures the essence of the problem, and compared the results with a baseline method commonly used in precision measurements at the Large Hadron Collider (LHC) at CERN. The unfolded distribution is in very good agreement with the original one. We also show how the method can be extended to include the effect of nuisance parameters representing sources of systematic uncertainties affecting the measurement.

physics.data-an↗
arXiv2019

Bidirectional Long Short-Term Memory (BLSTM) neural networks for reconstruction of top-quark pair decay kinematics

A probabilistic reconstruction using machine-learning of the decay kinematics of top-quark pairs produced in high-energy proton-proton collisions is presented. A deep neural network whose core consists of a Bidirectional Long Short-Term Memory (BLSTM) is trained to infer the four-momenta of the two top quarks produced in the hard scattering process. The MadGraph5+Pythia8 Monte Carlo event generator is used to create a sample of top-quark pairs decaying in the $μ$+jets channel, whose final-state objects are used to create the input to the deep neural network. Distortions due to limited resolution of the experimental apparatus are simulated with the Delphes3 fast detector simulator. The level of agreement between the Monte Carlo predictions and the BLSTM for kinematic distributions at parton level is comparable to that obtained using a benchmark method that finds the jet permutation that minimizes an objective function.

hep-ex↗
arXiv2019

Shower it again, Pythia

The Parton-Shower algorithm implement in the Pythia generator is applied multiple times to the same parton-level configuration to estimate the systematic uncertainty affecting large-radius jet substructure variables associated with the stochastic nature of the algorithm. Results are presented in the case of boosted $h\rightarrow b\bar{b}$ and $t\rightarrow bqq$. The code is publicly available on the repository https://github.com/rdisipio/ReShower.git

hep-ex↗
arXiv2019

DijetGAN: A Generative-Adversarial Network Approach for the Simulation of QCD Dijet Events at the LHC

A Generative-Adversarial Network (GAN) based on convolutional neural networks is used to simulate the production of pairs of jets at the LHC. The GAN is trained on events generated using MadGraph5 + Pythia8, and Delphes3 fast detector simulation. We demonstrate that a number of kinematic distributions both at Monte Carlo truth level and after the detector simulation can be reproduced by the generator network with a very good level of agreement. The code can be checked out or forked from the publicly accessible online repository https://gitlab.cern.ch/disipio/DiJetGAN .

hep-ex↗
arXiv2015

GPGPU for track finding in High Energy Physics

The LHC experiments are designed to detect large amount of physics events produced with a very high rate. Considering the future upgrades, the data acquisition rate will become even higher and new computing paradigms must be adopted for fast data-processing: General Purpose Graphics Processing Units (GPGPU) is a novel approach based on massive parallel computing. The intense computation power provided by Graphics Processing Units (GPU) is expected to reduce the computation time and to speed-up the low-latency applications used for fast decision taking. In particular, this approach could be hence used for high-level triggering in very complex environments, like the typical inner tracking systems of the multi-purpose experiments at LHC, where a large number of charged particle tracks will be produced with the luminosity upgrade. In this article we discuss a track pattern recognition algorithm based on the Hough Transform, where a parallel approach is expected to reduce dramatically the execution time.

physics.ins-det↗