Power-Aware FPGA Mapping of Convolutional Neural Networks

Theme: Digital Economy
Area: Artificial intelligence technologies

In recent years, Machine Learning and Artificial Intelligence techniques have demonstrated significant impact in a number of applications and have adopted by a variety of industries. Nevertheless, the current machine learning models require a large number of computations, making their deployment on embedded devices that are compute and energy constrained a challenging problem.

The project will focus on the research and development of methodologies and tools that would enable the reduction of energy and power consumption of the deployment of machine learning algorithms on an embedded device, maintaining at the same time their performance and delivering the required processing capability. The research will focus on the derivation of models for power consumption and their integration in existing frameworks that allow to optimise the mapping of a Machine Learning algorithm into an embedded platform meeting performance and energy/power targets. As machine learning algorithms are in general resilient to approximations, novel lossy techniques will be investigated tailored for machine learning workloads that would deliver the necessary energy/power gains, with minimum impact on the resources and performance of the system.

The outcome of the project will be a set of methodologies and tools that would allow the deployment of machine learning based systems on embedded devices, enabling as such the development of IoT nodes with machine learning capabilities in a reduced energy/power footprint, leading to the extended life of the system.