Rethinking Deep Learning on Remote Smart Sensors

Develop new neural network compression mechanisms to accelerate neural networks on low powered FPGA and embedded GPU smart sensors
Description of the Project: 

Neural networks for deep learning have been proven successful for many different domains, such as autonomous driving, conversational agents, autonomous robotics and computer vision. Neural network models are typically trained and executed on GPUs, but these have significant energy costs and lack portability needed for remote smart devices. FPGAs and embedded GPUs solve this problem, but cannot host large trained models. Thus, mechanisms to compress neural networks are needed to fit within hardware resource constraints without losing accuracy of AI inferences the model can make.

This PhD topic will combine hardware programming (embedded GPUs and programmable FPGAs) and neural network transformation approaches.

It is a very timely topic, with neural network approximation projects being led by industry: Distiller (Intel), Tensorflow Lite (Google) and FINN (Xilinx), which signifies the move away from focusing solely on high performance AI, towards so-called Edge Computing I.e. ubiquitous low powered AI sensors deployed from homes and vehicles, to robots and factories.

Resources required: 
Donation FPGA boards from Xilinx: Zedboards and Pynq boards.
Project number: 
140027
First Supervisor: 
University: 
Heriot-Watt University
Second Supervisor(s): 
First supervisor university: 
Heriot-Watt University
Essential skills and knowledge: 
Neural networks, deep learning, GPU/FPGA programming,
Desirable skills and knowledge: 
Compiler construction, approximate computing.
Industry placement details: 
Xilinx. If the student wants to focus on FPGA based neural network acceleration, the Xilinx Research group in Dublin may be keen to collaborate with PhD research. They develop the FINN Python framework for compiling neural networks to FPGAs.