Pierrick Pochelu

Pierrick Pochelu

Auto-generated profile, based on publicly available information

Pierrick Pochelu

HPC Platform, University of Luxembourg, Luxembourg

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 November 2025 in Engineering Applications of Artificial Intelligence
Verified
Journal Article
Closed access
PPPierrick Pochelu
Pierrick Pochelu
HCHyacinthe Cartiaux
Hyacinthe Cartiaux
JSJulien Schleich
Julien Schleich

7 May 2024
Verified
Conference Paper
Open access
PPPierrick Pochelu
Pierrick Pochelu
OLOscar Castro Lopez
Oscar Castro Lopez
In this paper, we present a comprehensive empirical study to evaluate four prominent Computer Vision inference frameworks. Our goal is to shed light on their strengths and weaknesses and provide valuable insights into the challenges of selecting the right inference framework for diverse situations. Additionally, we discuss the potential room for improvement to accelerate inference computing efficiency.

18 February 2026 in Journal of Open Source Software
Verified
Journal Article
Open access
PDF file
PPPierrick Pochelu
Pierrick Pochelu
SerializableSimpy is a Python framework for building discrete-event simulations (DES). DES modeling and analysis are useful in time-sensitive applications, such as manufacturing, logistics, and distributed systems. The framework provides core DES building blocks, including logical processes (LPs) and event causality. It also includes components for state and synchronization, such as stores, resources, and priority queues. Users focus on business logic, not the eventprocessing engine. After modeling the system, users can start, pause, and resume.

20 October 2026 in Open Research Europe
Verified
Journal Article
Open access
PPPierrick Pochelu
Pierrick Pochelu
Software development pipelines are increasingly targeted by attacks. These attacks are diverse, ranging from vulnerable code and exposed secrets to software supply chain compromise and runtime exploitation. While numerous security tools exist, they are often implemented as isolated controls, creating gaps in visibility and protection across the software development lifecycle. This paper presents an open-source GitLab-based DevSecOps demonstrator that integrates 16 automated security controls spanning development, software supply chain security, and runtime protection. The architecture combines static and dynamic analysis, dependency and container scanning, artifact signing and verification, compiler hardening, fuzz testing, Web Application Firewall (WAF), and Runtime Application Self-Protection (RASP). The approach is demonstrated on two intentionally vulnerable applications implemented in Python and C and mapped to OWASP Secure Product Design Principles, NIST SSDF, and OWASP SAMM. Experimental results show that all intentionally introduced vulnerabilities are detected, prevented, or mitigated by at least one security control. The demonstrator highlights the value of a defense-in-depth strategy in which complementary security mechanisms provide lifecycle-wide coverage while remaining fully open source, reproducible, and easily deployable within GitLab CI/CD environments.

7 April 2026
Verified
Thesis
Closed access
PPPierrick Pochelu
Pierrick Pochelu
Ensembles de réseaux de neurones répartis et parallèles appliqués au domaine de l'énergie: L'apprentissage ensembliste de réseaux de neurones profonds (Deep Neural Network - DNN) combine les prédictions de plusieurs DNN pour améliorer les performances de résultats obtenus à l’aide d’un réseau unique (DNN) notamment en réduisant l'erreur de généralisation.Ces techniques d’ensemble ont le potentiel d'améliorer des applications complexe et stratégique ML en particulier celles du domaine de l’énergie. Cependant, implémenter efficacement ces techniques d’ensembles soulèvent plusieurs questions: Comment orchestrer efficacement les différentes étapes du cycle de vie des ensembles de DNN (apprentissage, inférence) en vue d’une exécution globale performante tirant partie des infrastructures HPC. Comment entrainer plusieurs DNN indépendants, sélectionner un ensemble performant, et exposer les prédictions de l'ensemble construit pour des applications clientes.Dans ce travail, nous proposons des procédures pour construire un ensemble précis de DNN avec de multiples accélérations des étapes son cycle de vie. Nos procédures garantissent de trouver automatiquement de bons ensembles en se basant sur la diversité des DNN, leur précision individuelle et optimisation du coût de calcul de l'ensemble. En ce qui concerne les deux objectifs qualité des prédictions et la vitesse d'inférence, nous avons découvert que notre procédure construit des ensembles de DNN dont le résultat de prédiction apporte un gain significatif (rapidité, précision des prédictions, reproductibilité, économie en consommation de ressources) supérieur par rapport à l’utilisation d’un DNN unique.Ces technologies ont été implémentées au sein de l’entreprise, sur des cas d’usages concrets et opérationnels (classification d’images, contrôle optimal d’installation par renforcement learning). L’application sur ces cas d’usage, entre dans la stratégie de l’entreprise son orientation vers la production d’énergie décarbonée.

14 March 2024 in arXiv
Verified
Preprint
Open access
PDF file
PPPierrick Pochelu
Pierrick Pochelu
Among numerical libraries capable of computing gradient descent optimization, JAX stands out by offering more features, accelerated by an intermediate representation known as Jaxpr language. However, editing the Jaxpr code is not directly possible. This article introduces JaxDecompiler, a tool that transforms any JAX function into an editable Python code, especially useful for editing the JAX function generated by the gradient function. JaxDecompiler simplifies the processes of reverse engineering, understanding, customizing, and interoperability of software developed by JAX. We highlight its capabilities, emphasize its practical applications especially in deep learning and more generally gradient-informed software, and demonstrate that the decompiled code speed performance is similar to the original.