Generalisability and robustness of deep learning for automated image analysis

While deep learning has achieved highly accurate performance for many different applications, such as image analysis and natural language processing, there is still significant limitation in its generalisability and robustness. In particular, deep learning models often perform well on the dataset that the models are developed for, but then perform poorly when applied to a different dataset, even when the datasets are for the same tasks.

Various approaches have been developed recently to address these issues. In this project, we will further investigate this problem and the potential solutions will be based on domain adaptation, domain generalisation, out-of-distribution learning, meta-learning, etc.

This PhD study will involve a comprehensive literature review of the state-of-the-art, selection of a certain type of methodology and application domain (i.e., general computer vision, biomedical imaging, autonomous driving, robotics, etc.), method development and extensive experimental studies.

How to Apply

Express your interest in this project by emailing Associate Professor Yang Song. Include a copy of your CV and your academic transcript(s).