Energy Efficient High-Throughput Neuromorphic Processor for Deep Learning

The technical objective of this work is to develop a low power, high throughput neuromorphic system for online learning. The system will be based on an FPGA to enable low cost development and easy deployment. We propose to implement a transfer learning system as this will allow the FPGA to train a convolutional neural network with new data quickly and without significant hardware resource requirements. The system will be integrated with a software defined radio to be able to collect communications data continuously to be utilized for transfer learning. We have developed a prototype of the FPGA based transfer learning system in Phase I and aim to make the system more efficient, scaled up, more accurate, and general purpose so that it would be more adept at learning cognitive communications modulations in a cubesat environment. In addition to the FPGA based online learning system, we will investigate the design of accurate and efficient deep learning algorithms for the FPGA system to process cubesat communications data. This dual pronged approach of developing hardware and software will lead to the most efficient overall system, having low power, high accuracy, and high speed. We will integrate the system with a software defined radio to mimic satellite communications and carry out field tests of the system to ensure proper reliability and functionality. Our key deliverables will be the FPGA system design, deep learning algorithm evaluation and best algorithm for cognitive communications learning on the FPGA, and software defined radio integration of the FPGA system.