Adrian Goldwaser

PhD Thesis: Neural network generalisation in heavily overparameterised environments (working title)

Neural networks and deep learning methods have been seen to perform incredibly well on a wide variety of tasks. Despite this, there is very little underlying theory which explains why they work despite predictions of classical learning theory. My PhD focuses on understanding why they generalise so well to new data despite being able to simply memorise the training data.

Who or what inspired you to pursue your research interests?

My research interests were inspired by how a lot of neural network explanations and application advice does not explain why things work to a satisfactory degree. I think that as we begin to use neural networks for progressively more areas and more vital areas, it also becomes increasingly important to work on the theory behind them to understand when they will work and under what situations they will fail. It also provides insight to make sure they do what we want instead of what we tell them to do.