Trained machine learning hardware
Concurrent trends exist towards both solving problems using artificial intelligence, and minimising power consumption for use in smart-phones or IoT (Internet of Things) hardware.
The goals for the project are to reduce power consumption and cost (compared to implementation on GPUs) for hardware execution of pre-trained neural networks, which are a popular form of machine learning suited to many problems: especially audio/video classification/recognition where deep learning networks are often used.
We propose frameworks consisting of software-based training with matched hardware-based execution, where multiple methods for reducing computational complexity in hardware may be exploited in the interest of reducing power consumption, with minor compromises in network accuracy. In this project, we explore the parameter choices that affect the software-based training and the hardware-based execution.
Electrical and Electronic Engineering