PhD in Studentship in Reliable Deep Learning

Overview of the Project
Deep Learning methods for perception and control are highly effective for many tasks. However, in many applications, it is unavoidable that practical deployments will involve data of different statistics to the training data, or even adversarial attack; and performance significant degrades as a result. This project will investigate guaranteeing the deep network’s performance in these kind of situations. Potential approaches could include meta-learning for robustness to domain-shift and adaptation to domain shift, theoretical analysis of neural network robustness to adversarial attack through learning theory, or enabling manual verification of the decision though explainable AI techniques. Potential application domains include both deep learning for computer vision, as well as deep reinforcement learning for robot control.

Candidate’s profile
Project specific skills: Deep learning.
Programming skills: python, pytorch, tensorflow
A good Bachelors degree (2.1 or above or international equivalent) or Masters degree in a relevant subject (physics, mathematics, engineering, computer science, or related subject)
Proficiency in English (both oral and written)
Knowledge of computer vision, reinforcement learning and gradient-based optimisation techniques are highly desirable.