Jack Dymond

MINDS CDT, 2019

MINDS alumnus | Data Scientist, The Alan Turing Institute

Building AI that knows when to stop

Research focus: Resource-aware deep learning and progressive intelligence

Jack Dymond is a MINDS alumnus whose doctoral research developed deep-learning systems that vary their computational effort according to what each situation requires. He is now a Data Scientist at The Alan Turing Institute, working in the Applied Research Centre for Defence and Security.

View Jack’s professional profile

Connect with Jack on LinkedIn

View Jack’s research outputs

Contact the MINDS team

Career at a glance

MINDS status	Alumnus
MINDS metadata	MINDS CDT, 2019
Doctoral research	Resource-aware deep learning
Research approach	Progressive intelligence and branched neural networks
Current role	Data Scientist
Current organisation	The Alan Turing Institute
Current centre	Applied Research Centre for Defence and Security (ARC)
MINDS-linked experience	Dstl-sponsored PhD and three-month secondment described in approved MINDS material

Matching computational effort to the problem

Deep-learning systems do not always need to use the same level of computational effort for every decision. Some situations may need a full, slower and more expensive model, while others can be handled effectively using a faster and less computationally demanding output.

During his MINDS PhD, Jack Dymond researched an approach to resource-aware deep learning known as progressive intelligence. The system produces a lower-cost output first, then invests additional processing only when the initial result is not sufficient.

His work focused particularly on branched neural networks. These models include early-exit points that allow processing to stop as soon as the system has produced an output that meets the required threshold.

The aim is not simply to make a model smaller or faster in every circumstance. It is to help the same system use its resources more intelligently, adjusting the level of computation to the demands of each case.

Why the research matters

Using computing resources more efficiently

A system that always runs its fullest model may spend unnecessary processing time and energy on cases that could have been handled by a quicker output.

Supporting different levels of confidence

Progressive systems can provide an initial result quickly, then invest more computational effort when greater confidence is needed.

Helping large models work in practice

Resource-aware approaches can contribute to making large deep-learning systems more efficient and reliable in operational settings.

Progressive intelligence and early-exit networks

Progressive intelligence treats computational effort as something that can be adjusted rather than fixed. A model produces an initial output using a lower level of computation and continues processing only where the result needs to be improved.

Jack explored this through branched neural networks, which contain intermediate outputs known as early exits. These allow inference to stop once a satisfactory result has been produced, rather than requiring every input to pass through the complete model.

Research and outputs

Jack’s doctoral outputs included a survey of Graceful Degradation and two papers exploring how early-exit policies can be designed and applied in branched neural networks. Graceful Degradation concerns how machine-learning systems can decline in performance gradually rather than failing suddenly.

Jack’s wider publication record is available on his Google Scholar profile.

MINDS doctoral project: Achieving Progressive Intelligence for Low Power Systems

From MINDS to The Alan Turing Institute

During his PhD, Jack worked with Dstl and The Alan Turing Institute, gaining experience of research outside a solely academic setting. His PhD was sponsored by Dstl and approved MINDS material describes a three-month secondment at Porton Down. His PhD collaborations gave Jack practical experience in non-academic research and helped him build the relationships that ultimately led to his current role at The Alan Turing Institute.

Jack’s current Turing profile states that he worked on multiple projects with the Applied Research Centre before joining The Alan Turing Institute in January 2024. He is now a Data Scientist in the Applied Research Centre for Defence and Security, where he works on deep-learning-based solutions to challenges in the defence and security sector.

His work at the Institute spans a wider range of applications and modalities than his PhD, with the same underlying aim: making large deep-learning models work efficiently and reliably in practice.

Advice for current researchers

I had no idea when I picked up that literature review, on the recommendation of my supervisors, that it would shape the rest of my career. That’s really the point I’d want to make to current MINDS researchers: every opportunity matters, even when it’s not obvious why. If you have a specific direction in mind post-PhD, reach out directly and start building relationships now - don’t wait until submission. The earlier you lay that groundwork, the stronger a foundation you’ll have when the time comes.