• Heriot-Watt University, Edinburgh
  • The latest start date is January 2024

Deep Neural Networks (DNNs) achieve the state-of-the-art performance in several tasks in computer vision, natural language processing, image reconstruction, and many other areas. However, their black-box nature makes them inscrutable. In particular, they often learn spurious correlations between data, especially when the latter is not sufficiently varied. This problem, which is aggravated under distribution shifts (i.e., when the training and testing data differ significantly), can be traced back to the fact that all its knowledge has to be distilled from the training data. While prior knowledge, e.g., physical models, can be encoded in the training data via augmentation, embedding such information into the network's architecture has the potential to guarantee better generalization, robustness, and smaller data requirements.

In this project, which will be done in close collaboration with SeeByte Ltd, Edinburgh, we will look into generative models applied to underwater imagery. The goal is to improve the performance of current DNN-based algorithms for detection by exploiting the knowledge of the sonar acquisition process and, possibly, other physical models like motion and geometry. By designing physics-informed DNNs, we hope to obtain algorithms that not only are more reliable than conventional DNN ones, but also require smaller training datasets.

The student will be integrated in a dynamic team, supervised by Dr. Joao Mota, at Heriot-Watt University, and will interact regularly with SeeByte Ltd., a company that produces state-of-the-art software for managing unmanned and remote robotics assets in the maritime domain. The student will also be integrated into the National Robotarium, the UK's centre for Robotics and Artificial Intelligence.

For full details and to apply please visit: https://www.findaphd.com/phds/project/embedding-physical-models-into-deep-neural-networks-for-sonar-detection-co-funded-by-seebyte-ltd/?p154120