Claire Théobald (Opailleur) will defend his thesis, entitled “Bayesian Deep Learning for mining and analyzing astronomical data”, on Tuesday June 6th at 1:30pm in room A008.
In this thesis, we address the issue of trust in deep learning predictive systems in two complementary research directions. The first line of research focuses on the ability of AI to estimate its level of uncertainty in its decision-making as accurately as possible. The second line, on the other hand, focuses on the explainability of these systems, that is, their ability to convince human users of the soundness of their predictions.
The problem of estimating the uncertainties is addressed from the perspective of Bayesian Deep Learning. Bayesian Neural Networks assume a probability distribution over their parameters, which allows them to estimate different types of uncertainties. First, aleatoric uncertainty which is related to the data, but also epistemic uncertainty which quantifies the lack of knowledge the model has on the data distribution. More specifically, this thesis proposes a Bayesian neural network can estimate these uncertainties in the context of a multivariate regression task. This model is applied to the regression of complex ellipticities on galaxy images as part of the ANR project “AstroDeep”. These images can be corrupted by different sources of perturbation and noise which can be reliably estimated by the different uncertainties. The exploitation of these uncertainties is then extended to galaxy mapping and then to “coaching” the Bayesian neural network. This last technique consists of generating increasingly complex data during the model’s training process to improve its performance.
On the other hand, the problem of explainability is approached from the perspective of counterfactual explanations. These explanations consist of identifying what changes to the input parameters would have led to a different prediction. Our contribution in this field is based on the generation of counterfactual explanations relying on a variational autoencoder (VAE) and an ensemble of predictors trained on the latent space generated by the VAE. This method is particularly adapted to high-dimensional data, such as images. In this case, they are referred as counterfactual visual explanations. By exploiting both the latent space and the ensemble of classifiers, we can efficiently produce visual counterfactual explanations that reach a higher degree of realism than several state-of-the-art methods.
Thesis Supervisors :
Invited member :