< Retour au sommaire

Soutenance de thèse de Romain Xu-Darme

Titre: Algorithms and evaluation metrics for improving trust in machine learning: application to visual object recognition

le

Lieu: 700 avenue Centrale, 38400 Saint-Martin-d'Hères, room 306 (Grenoble)

Suivre en visio

Meeting ID: 593 962 6342

The integration of decision-making algorithms based on deep neural networks in critical applications - such as healthcare or automotive - depends on their ability to elicit trust in their decision, a major issue that is at the heart of this thesis. More precisely, our work focuses on three prerequisites to trust in a machine learning model : the performance of the model, i.e. its ability to perform the requested task ; its ability to estimate the uncertainty of its decision ; its explainability, which characterises the possible degree of interactiveness between the user and the model in order to obtain information regarding its inner-workings. Explainability is either inherent to the model (self-explaining model) or the result of post-hoc explanation methods - that attempt to explain the behaviour of a pre-existing model. In the context of computer vision applications, self-explaining models first extract an intermediate representation of the image, composed of semantic variables called attributes. Then, they build their result upon these attributes using a decision-making process that is transparent for the user. However, this thesis points out limitations in the intermediate representation of current state-of-the-art self-explaining models, and proposes to improve attribute extraction through a new unsupervised part learning algorithm - called PARTICUL - that is embedded with an uncertainty measure that acts as a proxy for the visibility of each individual part. Finally, the explainability of a model also depends on the quality of the chosen explanation method. In practice, explanation methods are evaluated using dedicated metrics, with respect to a set of properties covering both the content and the form of the explanation. Our thesis first shows that some existing metrics imperfectly evaluate their corresponding property, then presents new metrics to mitigate the issues found. Additionally, when explaining the behaviour of a complex model for visual classification, we show that these properties of explanations actually correspond to conflicting expectations from the user. Hence, no explanation can simultaneously satisfy all properties and is, by default, the result of a compromise. This observation leads us to reaffirm the importance of the correctness of the information above all other properties, so that the user can make an informed decision when choosing whether to trust the model.