This project breaks the current 'weak' assumption in the literature that the performance of deep learning models reported on a holdout dataset is an indicator of the performance on all future and yet to be encountered conditions during deployment. In reality, performance fluctuates and can drop below critical thresholds when the robot travels through particular places, times and conditions.
This PhD project will investigate new methods and approaches to performance monitoring of deep learning-based perception without the need for ground truth data including uncertainty estimation, dataset shift and domain shift detection, and change detection.
We expect the outcomes of this project to include the development of new algorithms that can monitor the performance of deep learning-based perception modules in real-time, by extracting cues from extrospective and introspective data during robotic deployment.
Skills and experience
To successfully complete this project, you will need to have:
- strong programming skills in a language such as Python
- good general knowledge of deep learning and computer vision
- a strong interest in robotics application and enabling robots to work in the real world.
You may be able to apply for a research scholarship in our annual scholarship round.
Contact the supervisor for more information.