Researcher: Ferdia Sherry

We have been very interested in structure-preserving deep learning over the past few years, with one of the main applications being the design of robust neural networks. State-of-the-art neural networks have repeatedly been shown to suffer from a lack of robustness, for example in the form of adversarial examples in image classification: images that have been imperceptibly perturbed so as to fool image classifiers. This is the cause of significant concern in safety-critical applications such as medical imaging. One of the main threads of research within structure-preserving deep learning is concerned with the connections between neural networks and dynamical systems, specifically the design of neural network architectures based on the appropriate discretisation of parametrised continuous dynamical systems that have desirable stability properties. In this project, we are further pursuing this direction: applying this idea to the problem of robust image classification, generalising it to robust classification on other forms of data such as graph data, and applying it to Plug-and-Play image reconstruction with provable guarantees.

https://doi.org/10.1017/S0956792521000139 -- https://doi.org/10.1137/22M1527337 -- https://arXiv.org/abs/2306.17332 -- https://arXiv.org/abs/2311.06942