Stochastic Optimisation and ML models

Written by Josh Stevens
December 2, 2020

Continuing our discussion series with CCIMI students, we sit down with Derek to chat about his area of research and PhD journey so far.

Can you introduce yourself and describe your PhD project?
I am Derek Driggs, and I am developing new stochastic optimisation algorithms that can be used to speed up the training of machine learning models. Research shows that large models trained on enormous datasets exhibit the best performance, but training these models can take days, even when using high-performance computing hardware. Fast stochastic optimisation algorithms make it possible to train these models in a reasonable amount of time. Before coming to Cambridge, I studied applied mathematics at the University of Colorado Boulder. My Master’s thesis also focused on developing fast algorithms for fitting statistical models.

What excites you the most about your project?
One great quality of my field of research is that it has many applications. During my PhD, I have used new optimisation algorithms to de-noise CT scans, reconstruct old audio recordings, and fit statistical or machine learning models for a variety of prediction tasks. Because optimisation is important in so many areas, I have been able to meet and work with experts from several fields.

What was the biggest challenge and what did you find most enjoyable in the first year of your PhD?
I spent my first year learning as much as I could from lectures and talks throughout the university. Cambridge is one of the only places on Earth where students can regularly attend lectures and talks by world-leading academics. By walking a few feet from my office to the Isaac Newton Institute, I put faces to the names appearing on important papers that I had read, which was not something I could do as easily before beginning my programme at Cambridge. Of course, these unique opportunities present their own challenges. It is tempting to spend too much time gaining surface-level knowledge of many fields, but to perform high-quality research for a PhD, it is necessary to have a deep understanding of a small field. I think this shift in focus occurs naturally over the first year.

What advice would you have for students interested in the Mathematics of Information PhD?
I would advise students to use and contribute to the Mathematics of Information community. Everyone’s research is unusually collaborative (for maths research), so it is common to have new ideas after discussing someone else’s project. Even if a conversation is ultimately not useful for research, there are social benefits as well!