Internship: Explainable AI for the qualification of aerospace algorithms

<p><strong>Background</strong></p><p class="MsoNormal" style="text-align:justify;">Machine learning is increasingly used in aerospace for tasks that have traditionally relied on human experts or on fixed rule-based procedures. For such tasks, data-driven algorithms promise gains in speed, consistency and sensitivity. Before they can be used in practice however, they must be qualified, which means it has to be demonstrated with sufficient substantiation that the algorithm performs its intended function reliably within a defined set of conditions. The qualification approaches used in the sector today were developed for human operators and for conventional systems whose behaviour can be traced back to explicit requirements. They do not translate directly to algorithms whose behaviour is learned from data, and emerging guidance from aviation authorities names explainability as a key ingredient of trustworthy AI, while leaving open how it should be demonstrated in practice.</p><p style="min-height: 1.7em;"></p><p class="MsoNormal" style="text-align:justify;">Performance figures on a test set do not show why an algorithm reaches a decision, whether it relies on physically meaningful information, or how it behaves outside the data it was trained on. Explainable machine learning offers a wide range of methods to address this, from feature attribution and saliency methods, to example-based and concept-based explanations, but it is still an open question which of these can provide evidence that is strong enough to support qualification. This project therefore investigates how explainability methods can contribute to the qualification of machine learning algorithms for aerospace, starting from an use case, and provides a comparative assessment of several methods.</p><p style="min-height: 1.7em;"></p><p><strong>Assignment</strong></p><p class="MsoNormal" style="text-align:justify;">The assignment will include the following tasks:</p><ul><li><p>A literature study on the qualification of machine learning algorithms in safety-critical sectors and on explainability methods for machine learning (e.g., attribution and saliency methods, example-based and concept-based explanations, interpretable models, uncertainty quantification), including the ways in which explanations themselves are evaluated.</p></li><li><p>A use case provided by NLR, for instance automated defect detection, and identification of the substantiation needed to qualify the algorithm, translated into criteria that explanations should meet.</p></li><li><p>Implementing several explainability methods for the use case and analysing them against these criteria, for example their faithfulness to the model, their robustness and their usefulness to domain experts.</p></li><li><p>Assessing how the resulting explanations can serve as evidence in the qualification of the algorithm, and formulating recommendations for the use of explainability in qualification.</p></li></ul><p style="min-height: 1.7em;"></p><p><strong>Results</strong></p><p class="MsoNormal">The final outcomes of this assignment will be:</p><p class="MsoListParagraphCxSpFirst">§&nbsp; A structured assessment, as part of a literature study, of explainability methods for machine learning and their suitability for qualification in aerospace, together with a justified selection of the use case.</p><p class="MsoListParagraphCxSpMiddle">§&nbsp; An implementation and comparative evaluation of several explainability methods on the selected use case, with recommendations on how explainability can support the qualification of the