On the Feasibility of EASA Learning Assurance Objectives for Machine Learning Components - IRT Saint Exupéry - Institut de Recherche Technologique Access content directly
Preprints, Working Papers, ... (Preprint) Year : 2024

On the Feasibility of EASA Learning Assurance Objectives for Machine Learning Components

Abstract

Despite the significant success of using Machine Learning (ML) in numerous industrial applications, how to integrate these technologies in safety-critical contexts poses many challenging questions. Several industrial and academic research groups, as well as various standardization committees are actively working to provide (partial) answers to these questions. In this document, we focus on one such initiative led by the EASA, which proposes a series of guidelines and requirements to develop ML-based systems for critical applications in the aviation domain. In this paper we investigate whether these requirements can be satisfied when using ML to solve a relatively simple regression task, that of building a neural network surrogate of the International Geomagnetic Reference Field (IGRF) model. Though we acknowledge all the structuring efforts towards the ambitious certification goal, our analysis pinpoints several important issues with some of these guidelines, such as ambiguous definitions, prohibitive computational costs, or currently very limited theoretical guarantees. Our analysis compels us to remain cautious about the various general recommendations proposed for designing trustworthy ML components for safety-critical systems. These conclusions call for the academic and industrial communities concerned by "Trustworthy AI" to strengthen their collaboration and pursue the research efforts necessary to address the existing challenges and establish sound methodologies for building safe ML-based applications.
Fichier principal
Vignette du fichier
ERTS-paper49-final.pdf (1.18 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04575318 , version 1 (14-05-2024)

Identifiers

  • HAL Id : hal-04575318 , version 1

Cite

Florence de Grancey, Sébastien Gerchinovitz, Lucian Alecu, Hugues Bonnin, Joseba Dalmau, et al.. On the Feasibility of EASA Learning Assurance Objectives for Machine Learning Components. 2024. ⟨hal-04575318⟩
0 View
0 Download

Share

Gmail Mastodon Facebook X LinkedIn More