RESERVOIR COMPUTING AND BENCHMARKING OF NEUROMORPHIC SYSTEMS FOR SWaPCONSTRAINED AUTONOMOUS PROCESSING
Abstract
Autonomous processing is critical in several military and intelligence applications (e.g. target identification/ classification). Furthermore, the growing amount and availability of visual data in the past two decades shows the pressing need to develop autonomous agents in data analysis tasks. Research has shown that the Air Force could require up to 117,000 personnel dedicated to motion imagery exploitation alone, which is one-third of the active-duty Air Force. Though these estimates use simplistic analysis, the actual numbers are still untenable. Maintaining nonhuman eyes-on imagery with autonomous agents and cueing human analysts to observe only those images deemed of interest will significantly reduce the personnel overhead. Moreover, such real-time processing is also necessary for applications in autonomous control in unmanned aircraft vehicles (UAV), where the vehicle needs to react quickly to changing environments. In all most all these applications autonomous processing units have to be deployed onto size, weight, and power (SWaP)-constrained platforms. Moreover state-of-the-art vonNeumann computing systems are not designed for such autonomous processing with spatio-temporal patterns in SWaP constrained systems. In this work, our overarching goal is to develop efficient neuromorphic hardware architectures suitable for SWaP constrained systems to address real-time visual information processing tasks. To achieve this goal, we take a two-pronged approach. First, we propose to design a SWaP constrained fully-scalable architecture for an emerging neuromorphic computing paradigm called reservoir computing (RC). RC employs a random recurrent neural network (the reservoir) to project sensor data onto a high-dimensional space. The dynamics of the reservoir facilitate integration of sensory data in spatial and temporal dimensions, effectively enabling short-term memory. RC is especially suited for real-time autonomous processing because their performance is robust to noise and failure in the structure of the network. Second objective is to perform SWaP benchmarking of different cognitive tasks in a neuromorphic system, to identify the best neuromorphic platforms for autonomous processing. In both the objectives, the proposed architectures and conventional algorithms will be designed in software, in Register Transfer Language (RTL), and ported onto reconfigurable platforms like Field Programmable Gate Arrays, and custom ASICs for a direct comparison of their SWaP efficiencies. Our target application is complex autonomous information processing over continuous input streams, such as anomaly detection in images.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2016
- Source ID
- FA87501610108
Entities
People
- Dhireesha Kudithipudi
Organizations
- Rochester Institute of Technology
- Rome Laboratory
- United States Air Force