Dr Edward Milsom

Research interests:

• Deep Learning
• Optimisation
• Bayesian methods

Ed’s research investigates the foundations of deep learning, including optimisation, theory, and Bayesian deep learning. Modern AI uses neural networks to learn relationships and meaning in complex, unstructured data like images, text ,and audio. However, neural network training currently utilises many heuristics and architectural choices that are poorly understood, leaving much room for improvement,  with the end goal being to reduce training time, save energy, increase performance on tasks, and move towards more reliable / trustworthy models. Recent work includes investigating methods for automatically determining optimal learning rates for different layers of a neural network (using tricks inspired by Bayesian methods), and utilising spectral properties of gradient descent updates to improve training.

LINKS:

Edward Milsom on the University of Bath research portal