ITU researcher receives grant to enable powerful machine learning in smaller hardware environments

Associate Professor of Computer Science at the IT University of Copenhagen, Pınar Tözün, has secured 2.7 million kroner from the Novo Nordisk Foundation to develop novel mechanisms to get more value out of data using the computing power of small devices.

Pinar TozunComputer Science DepartmentResearchClimate ITgrantssoftware

Written December 16, 2022 7:01 AM by Theis Duelund Jensen

Today real-world applications such as speech recognition in virtual assistants (e.g., mobile phones) or image recognition in satellites are powered by machine learning models. However, training models in the cloud is costly, and that is not even factoring in the electricity usage and carbon footprint. The costly models are then deployed on smaller devices backing up one or more applications.

There is therefore a need for optimisation of processes in the field, and that is where Associate Professor of Computer Science at the IT University, Pınar Tözün’s new research project Machine Learning on Tiny Hardware (MOTH) comes into play.

The goal of the MOTH project is to develop novel mechanisms to increase the value of data using the computing power of smaller devices. In contrast to the computing resources in the cloud, smaller devices are more resource-constrained (one or two orders of magnitude) thus posing a challenge for optimisation.

However, enabling more operations at these devices would reduce latency, cost, and power required to deploy machine learning models on them and would additionally secure a safer processing of data. Ideally, Pınar Tözün’s findings will benefit all institutions that collect and process data on resource-constrained devices.

“In this project, the goal is to tackle this problem in the context of machine learning models. What level of complexity in models can we actually work with when we deploy them on modern edge devices, and can we do that dynamically? Can we update models from time to time, and what is the cost of that? The current practice involves optimising the models for the particular setting. It is quite static and not very flexible when you want to implement updates or move to different hardware,” says Pınar Tözün.

If successful, the research could pave the way for a reduction of data traffic to data centers or bigger hardware resources. If smaller devices closer to data sources can produce more value, the need for computation at high-performance data centers will be reduced. The effect is a reduction in latency of operations, the monetary cost, as well as the carbon footprint.

“Ultimately, the project is about getting more value out of resource constrained devices,” says the researcher. “We are not going to build new hardware. What we are working on is optimising the software side of the equation which may impact the way we deploy machine learning models in the future.”

The project is slated to run for three years and will start in August 2023. The 2.7 million kroner has been awarded by the Novo Nordisk Foundation.

Further information