Robust Deep Learning for Autonomous Driving

TL;DR

This paper develops new methods for reliable uncertainty estimation in deep neural networks to enhance safety and robustness in autonomous driving systems, addressing failure prediction, domain adaptation, and out-of-distribution detection.

Contribution

It introduces the true class probability (TCP) criterion for better failure prediction, a learning scheme for TCP, and extends uncertainty measures to out-of-distribution detection in autonomous driving.

Findings

TCP outperforms existing uncertainty measures in failure prediction

The learned confidence improves domain adaptation via better pseudo-label selection

Evidential models effectively detect misclassification and out-of-distribution samples

Abstract

The last decade's research in artificial intelligence had a significant impact on the advance of autonomous driving. Yet, safety remains a major concern when it comes to deploying such systems in high-risk environments. The objective of this thesis is to develop methodological tools which provide reliable uncertainty estimates for deep neural networks. First, we introduce a new criterion to reliably estimate model confidence: the true class probability (TCP). We show that TCP offers better properties for failure prediction than current uncertainty measures. Since the true class is by essence unknown at test time, we propose to learn TCP criterion from data with an auxiliary model, introducing a specific learning scheme adapted to this context. The relevance of the proposed approach is validated on image classification and semantic segmentation datasets. Then, we extend our learned confidence approach to the task of domain adaptation where it improves the selection of pseudo-labels in self-training methods. Finally, we tackle the challenge of jointly detecting misclassification and out-of-distributions samples by introducing a new uncertainty measure based on evidential models and defined on the simplex.