Skip to main content

Safety Assurance for Deep Neural Networks

Supervisor

Suitable for

MSc in Computer Science
Computer Science, Part B
Mathematics and Computer Science, Part C
Computer Science and Philosophy, Part C
Computer Science, Part C

Abstract

Safety Assurance for Deep Neural Networks

Professor Marta Kwiatkowska is happy to supervise projects in the area of safety assurance and automated verification for deep learning, including Bayesian neural networks. For recent papers on this topic see  http://qav.comlab.ox.ac.uk/bibitem.php?key=WWRHK+19, http://qav.comlab.ox.ac.uk/bibitem.php?key=RHK18 and http://qav.comlab.ox.ac.uk/bibitem.php?key=CKLPPW+19, and also https://www.youtube.com/watch?v=XHdVnGxQBfQ.

Below are some concrete project proposals, but students’ own suggestions will also be considered:

  • Robustness of attention-based sentiment analysis models to substitutions. Neural network models for NLP tasks such as sentiment analysis are susceptible to adversarial examples. In a recent paper (https://www.aclweb.org/anthology/D19-1419/) a method was proposed for verifying robustness of NLP tasks to symbol and word substitutions. The method was evaluated on CNN models. This project aims to develop similar techniques for attention-based NLP models (www-nlp.stanford.edu/pubs/emnlp15_attn.pdf).
  • Attribution-based safety testing of deep neural networks. Despite the improved accuracy of deep neural networks, the discovery of adversarial examples has raised serious safety concerns. In a recent paper (http://qav.comlab.ox.ac.uk/bibitem.php?key=WWRHK+19) a game-based method was proposed for robustness evaluation, which can be used to provide saliency analysis. This project aims to extend these techniques with the attribution method (http://arxiv.org/abs/1902.02302) to produce a methodology for computing the causal effect of each feature and evaluate it on image data.
  • Uncertainty quantification for end-to-end neural network controllers. NVIDIA has created a deep learning system for end-to-end driving called PilotNet (http://devblogs.nvidia.com/parallelforall/explaining-deep-learning-self-driving-car/). It inputs camera images and produces a steering angle. The network is trained on data from cars being driven by real drivers, but it is also possible to use the Carla simulator. In a recent paper (http://arxiv.org/abs/1909.09884) a robustness analysis with statistical guarantees for different driving conditions was carried out for a Bayesian variant of the network. This project aims to develop a methodology based on these techniques and semantic transformation of weather conditions (see http://proceedings.mlr.press/v87/wenzel18a/wenzel18a.pdf) to evaluate the robustness of PilotNet or similar end-to-end controllers in a variety of scenarios.