Skip to main content Skip to navigation

Dr Suhaib Fahmy, Royal Academy of Engineering/Leverhulme Trust Research Fellowship

Edge Intelligence for Real-World Distributed Machine Learning

The potential for machine learning to revolutionise industries and sectors - from healthcare to transportation - is drawing significant research interest.

Yet, the computational complexity of training and deploying methods like deep neural networks remains a significant barrier to wider application in the real world.

There is growing research activity around more efficient inference using neural networks, but training remains a challenge and is limited to powerful datacentre machines.

Dr Suhaib Fahmy's Fellowship will explore novel computing architectures to enable efficient inference and training of neural network models near the edge of the network.

This will allow prototyping and deployment of real machine learning applications nearer to the data sources.

It will involve moving away from applications with centralised historical stored data to those with real-time streaming data, for use in a variety of real world applications.

Machine learning can enable huge datasets to be processed beyond the scope of human capability - which will prove invaluable in multiple and sectors, including healthcare.

 

Dr Fahmy will aim to narrow the gap between machine learning theory research and real world applications.

Virtualised architectures

Existing research on efficient neural network architectures, focused primarily on inference, requires low level optimisation for a specific application. Dr Suhaib Fahmy's Fellowship explores virtualised architectures that could support multiple machine learning applications and distributed streams of data on shared hardware, allowing these accelerators to be deployed as a reusable, shared network service.

The key aim of the Fellowship is to narrow the growing gap between machine learning theory research and real world distributed applications with constraints on predictability, latency, privacy, and efficiency by offering an open hardware accelerated framework that can be deployed for prototyping, experimentation, and deployment of a wide range of applications.

Dr Fahmy will also examine how some limited support for training can be added to these efficient edge accelerators, to enable distributed training using data at the edge, rather than sending it all to a centralised cloud datacentre. This will offer improvements in scalability and privacy, and improve latency and bandwidth use.