Research
My research explores how neural networks represent information, generalise across domains, and operate under practical compute constraints. My work spans robust feature learning, memory-efficient training, transformer compression, computer vision, and computational pathology.
Featured
Nuclear masks for robust histopathology
Models trained on data from one hospital often struggle when used on data from other hospitals. We guide the training of a standard image classifier using nuclear masks, helping it focus more on nuclear shape and arrangement. This approach improves accuracy and robustness across hospitals, and the model does not need masks during testing.
Curriculum Learning & Training Dynamics
A 2017 study of individual-example and mini-batch dynamics during neural-network training: how quickly examples are learned, how their losses move across epochs, how Highway Networks allocate gated computation, and why apparently sensible curricula can still lose to random shuffling.