Skip to Content

Adaptive Acceleration of Neural-Networks in Resource Constrained Edge and Cloud Environments

PhD project title

Adaptive Acceleration of Neural-Networks in Resource Constrained Edge and Cloud Environments  

Outline description, including interdisciplinary, intersectoral and international dimensions (300 words max)

In the past decade, neural-networks have gained a lot of popularity and are being widely used in object recognition, data analysis and classification tasks of smart-car,-city,-health applications. Traditionally, neural-networks are designed for optimal prediction capabilities through large-scale and dense connectivity without considerations for practical implementations, and thus they end-up being very computationally intensive and power hungry.  However, as Neural-Networks are increasingly being used in various devices in Edge and Cloud environments, where power consumption is a primary concern, the attention of industry and users turns into energy-efficient Neural-Networks. To this end, the primary aim of this project is the acceleration of Neural-Networks at the right energy and accuracy based on dynamic task-scheduling and accuracy-adaptation on heterogeneous devices at the Edge and Cloud.  This will be achieved by designing new low-cost hardware mechanisms and combining them with intelligent runtime support that ensure each task is dynamically precision tailored based on the informational value of the processed data as well as the operating conditions and available resources on shared multi-tenant environments at the Edge and Cloud. The developed cross-layer mechanisms will allow to exploit the inherent resilient properties of NNs and dynamically tune the data-dependent precision controls to opportunistically scale resource provisioning for step-changing energy-efficiency and resilience.

As the main case study, we will use Object Recognition in Video Streams, which is extremely important in emerging secure surveillance and smart-city applications including self-driving cars. The Neural-Network and mechanisms will be developed on reconfigurable hardware, i.e. Field Programmable Gate Arrays (FPGAs), which offers the required flexibility for architectural optimizations that is not available in existing processors used by most of the current works.  Such devices have already started being used in Cloud datacentres like Amazon’s and are excellent candidates for emerging Edge deployments, which we will enable by the developed mechanisms.     

The project is closely aligned with one of the main interdisciplinary themes of scalable-secure-intelligence of the ECIT GRI, which also forms a central part of the announced Belfast City-Deal. The project brings together experts on hardware design, and data science enhancing the rapidly growing area of Artificial-Intelligence in Queen’s and in Northern-Ireland as detailed in the recent report from the Matrix-panel thus supporting the targeted creation of a regional Centre of Excellence on AI. The project promotes the mobility of the Fellow to a world leading industrial research lab, where he will be hosted for 4 months and will establish a international academia-industry collaboration. The research will lead to impactful articles on top tier international venues both in hardware design and data science on which the supervisory team has annual presence.     


Key words/descriptors



Energy-Efficiency, Artificial-Intelligence, Neural-Networks, Accelerators, Heterogeneous Systems, Runtime, Smart City, Smart Cars, Surveillance, Object Recognition, Video Streaming   

Fit to CITI-GENS theme(s)

The project is aligned with most CITI-GENS themes, since it targets the acceleration of Neural-Networks that are central in emerging smart city, smart health and smart manufacturing applications. All such domains will combine data analysis at resource constrained devices at the Edge with further processing at the Cloud as we target. The targeted application is in particular aligned with Information Technology and Creative Industries themes of CITI-GENS.  

The project is intersectoral and interdisciplinary combining computer engineering and data science on a novel research area that is a gaining a lot of commercial interest as outlined in the recent report from the Matrix-panel , and recent products and services announced by international companies like Amazon, Xilinx, Google, IBM.

The project will lead to the implementation of novel hardware for neural networks with the potential to be published on top-tier venues in the area of design automation, hardware design, multimedia and machine learning like IEEE international design automation conferences (DAC, DATE), International Symposium on Microarchitecture (MICRO), Field-Programmable Custom Computing Machines  (FCCM), Computer Vision and Pattern Recognition (CVPR), International Conference on Computer Vision (ICCV), International Conference on Machine Learning (ICML), Neural Information Processing Systems (NeurIPS) as well as prestigious journals.

The team will aim at participating on international hardware and neural-networks design competitions in some of which both academic supervisors have won in the past. The student will become member of the HIPEAC network of excellence which has over 2500 researchers from across the globe on embedded and high performance systems where he will have the chance to demonstrate his results on annual international summer schools and workshops.

Supervisor Information



First Supervisor: Georgios Karakonstantis                                                         School: EEECS, ECIT (computer science)

Second Supervisor: Yang Hua                                                                               School: EEECS, ECIT  (data science)

Third Supervisor:  Michaela Blott                                                                        Company: Xilinx, Dublin ,Ireland


Name of non-HEI partner(s)


XILINX is the inventor of the FPGA and programmable SoCs with approximately 3000 employees worldwide. Their highly-flexible programmable silicon, enabled by a suite of advanced software and tools, drives rapid innovation across a wide span of industries and technologies - from aerospace and defense, automotive, industrial, scientific, and medical devices, and more recently in datacenters for energy-efficient acceleration of cloud computing. Xilinx delivers the most dynamic processing technology in the industry, enabling rapid innovation with its adaptable, intelligent computing. Xilinx owns almost half of the market that is set to touch $7.9 Billion by 2020.

XILINX has developed advanced IP modules and frameworks for Neural-Networks, which are already used by many users worldwide and are integrated in smart-city applications and self-driving cars. Michaela Blott, the 3rd supervisor is world expert on scalable Neural-Networks accelerators which are promoted through the research lab in Ireland where the Fellow will be hosted during their internship.  

Contribution of non-HEI partner(s) to the project:






The project is a unique opportunity for a PhD student to develop energy-efficient Neural-Networks on heterogeneous devices at the Edge and Cloud on which XILINX has extensive commercial interest.

XILINX will provide one of their latest ACAP boards to the project along with access to their commercially available Neural-Network framework and relevant development tools, while offering student supervision by an experienced researcher.  At the beginning of the project, the Fellow will be trained through the XILINX-XUP program on prototyping Neural-Networks on a heterogeneous board. During the project the student will be trained on negotiation skills, project-management and teamwork from XILINX, in addition to the training offered by Queen’s-EEECS on scientific-writing, presentations, and programming/scripting and the specially designed Queen’s training for COFUND students. Furthermore, the Fellow will be hosted in XILINX for a research internship for at least 4 months allowing him/her to get an industrial experience, while networking with engineers that could boost their career.

Overall, the Fellow will gain skills on an industry related research project on applications that have high commercial value and interest.  The long-term evaluation of the power consumption of Neural-Networks running on an advanced board will provide XILINX useful insights. The project will help identify mechanisms that can improve the energy-efficiency of Neural-Network accelerators, with the potential to give XILINX a market advantage especially in the foreseen deployments at the Edge. It is expected that joint publications will be produced which in combination with public presentations will help the Fellow develop his presentation skills, while enhancing the research image of XILINX. The project will establish a long-term relationship between ECIT and XILINX and will enable the identification of research challenges and the subsequent preparation of follow up funding proposals. Finally, the developed material are going to enhance current courses on Artificial-Intelligence and Cloud in EEECS, which is extremely timely given the increased utilization of FPGAs in datacenters and the use of heterogeneous devices as accelerators of Artificial-Intelligence workloads.

Note that in the past 2 years, we have been collaborating with XILINX Research in a major EU project, the Vineyard, in which I participated as CoI and Xilinx served as member of the industrial advisory board. The main aim of the project was the utilization of Field Programmable Gate Arrays (FPGAs) in data-centres. XILINX as the most important stakeholder in the market provided suggestions and directions on our work that focused on maximizing the sharing of FPGA resources among datacentre users. Our work resulted in 6 scientific papers in renowned conferences and invited presentations.

Research centre / School


Subject area

Electrical and Computer Engineering, Computer Science, Data Science