MINDS CDT, 2019
Researcher, The Alan Turing Institute
His doctoral research focused on resource-aware deep learning, developing an approach he and his collaborators called progressive intelligence. The central idea is that not every situation requires the same level of computational effort. A system that always runs a full, slow, expensive model wastes resources on cases where a fast, cheaper output would have been perfectly adequate. Progressive intelligence addresses this by producing low-cost, lower-confidence outputs first, then investing more processing only where the initial result is not sufficient.
To make this work in practice, Jack researched branched neural networks: architectures that include early exiting branches allowing inference to stop as soon as a satisfactory result is produced. His thesis produced two papers exploring how those early exit policies can be designed and applied. He also wrote a survey of the related field of Graceful Degradation, which concerns how machine learning systems can decline in performance gradually rather than failing suddenly.
During his PhD, Jack worked with Dstl and The Alan Turing Institute, including a three-month secondment at Porton Down. His PhD was funded by Dstl, and the practical experience of working in non-academic settings, alongside the relationships he built through those collaborations, led directly to his current role.
At The Alan Turing Institute his work spans a wider range of applications and modalities than his PhD, but the thread running through both is the same: making large deep learning models work efficiently and reliably in practice.