Cangxiong’s interests lie in the mathematics of machine learning. His recent focus is deep learning with differential privacy. He has been exploring how this learning can be achieved and privacy guarantees being proved by using compressive learning and dynamical systems methods.
Before joining IMI, Cangxiong was a research associate at the Department of Computer Science at the University of Bath. Funded by Innovate UK and in partnership with the Foundry and DNEG, he investigated how privacy of the training data in federated learning can be breached. In his work, he demonstrated how training images can be reconstructed using gradients from training and our knowledge of the model architecture in deep networks for image classification. Before joining the University of Bath, Cangxiong worked for startups in Cambridge UK and in Beijing China, where he used machine learning and statistics to model consumer behaviour and used that knowledge to improve pricing models and recommendation systems.
Cangxiong obtained his PhD in number theory from the University of Cambridge, where he generalised Kronecker limit formula to arbitrary number fields. He obtained his MPhil and Bachelor’s degrees from the University of Hong Kong. In his MPhil thesis, he proposed a differential geometric approach to study rational points on elliptic curves over function fields.
In his spare time, Cangxiong enjoys playing chamber music on the violin with other musicians and hiking and cycling in nature.