Explanations and meaningful information: at the interface between technical capabilities and legal frameworks
PLSC 2022
D. Bourgeois, S. Vergnolle
For many individuals, decisions taken by a computer are preferable to ones made by humans because they are often considered to be more objective. At the same time, individuals feel uncomfortable with autonomous vehicles roaming the roads. This clash can be explained by the fact that decisions made by computers, and more specifically by AI-enabled systems, can be virtually impossible to understand from the outside.
Learning Representations of Source Code from Structure and Context
MSc. Thesis
D. Bourgeois
Early transformer models would view source code as plain text. In this work, we show the value of heterogenous representations, in particular tree and graph-based forms such as ASTs. We also show how transformers can be considered equivalent to Graph Neural Networks (GNN) in some forms, and leverage this fact to propose novel code understanding models.
GNNExplainer: Generating explanations for Graph Neural Networks
NeurIPS 2019
R. Ying, D. Bourgeois, J. You, M. Zitnik, J. Leskovec
Graph Neural Networks (GNNs) represent information very compactly, yet when machine learning models operate on them we lose the ability to understand predictions' rationale. In this work we propose a model-agnostic framework to extract meaningful explanations of any GNN's predictions.
A Dynamic Embedding Model of the Media Landscape
WWW 2019
J. Rappaz*, D. Bourgeois*, K. Aberer
Leverage the unsupervised news coverage similarity maps from prior work, we show that market dynamics greatly affect coverage patterns. In particular, acquisitions tend to concentrate and unify the set of covered events over time.
Selection Bias in News Coverage: Learning It, Fighting It
WWW 2018
D. Bourgeois, J. Rappaz, K. Aberer
Leveraging advanced recommendation systems, we analyze large amounts of news coverage from many sources. We extract similarity maps from the embedding, in order to understand similarities and biases in coverage, and propose a method for diversifying news consumption.
Using holistic information in the Trigger
LHCb Public Note
D. Bourgeois, C. Fitzpatrick, S. Stahl
The trigger sifts through large amounts of particle collision data in near real-time. This must be done in software, at a whopping 30MHz rate. In doing so, it must remove large amounts of data while retaining the most important information (i.e. the rarest decays). For this purpose, we devise high recall, configurable machine learning methods to efficiently retain the most useful information for later processing.
New approaches for track reconstruction in LHCb's Vertex Locator
JHEP 2018
C. Hasse, J. Albrecht, B. Couturier, D. Bourgeois, V. Coco, N. Nolte, S. Ponce
The VErtex LOcator (VELO) is the closest subdetector to the collision point. We reconstruct collision tracks and the residuals of each hit with respect to the initial estimated straight line. Using recurrent networks, we estimate the position and uncertainty of a VELO track's closest to beam state. The resolution of this prediction as well as its ability to estimate the uncertainty is shown to be superior to existing methods.
