CNS Core: Small: Online Performance Monitoring of Neuromorphic Services

Drexel University, Philadelphia PA

Investigators

Abstract

Machine learning applications that are implemented via spiking neural networks (SNNs) can be executed using very little energy on a new type of computer architecture called neuromorphic processors. These processors mimic the structure and operation of biological neurons and synapses within the brain and are especially suitable for executing SNNs whose computations are guided by the location and frequency of spikes occurring within the network. Checking the correctness of results generated by a program is well understood for traditional processors. This project develops techniques to check the correctness of results generated by SNNs executing on neuromorphic processors. The intellectual merits of the project lie in the development of an online performance monitoring framework for neuromorphic processors. This framework includes both duplicate and compare, as well as low-cost model-based approaches to detect faults affecting the internal circuitry of these processors. Performance monitoring units collect real-time behavioral data from these circuits, in terms of spike times, to feed the analytical models for subsequent analysis. Efficacy of the monitoring framework is judged in terms of hardware and energy costs, intrusion overhead on executing SNNs, and fault coverage, using cycle-accurate simulations as well as fault-injection experiments on neuromorphic hardware. Tool and techniques developed by this project will promote use of neuromorphic computing within the broader science and engineering community in the United States, sustaining the leadership role in machine learning and artificial intelligence. The project involves undergraduate students in research via the Vertically Integrated Projects program at Drexel University. Collaborators from academia and industry deliver guest lectures on research and development in neuromorphic computing, with these lectures being integrated within relevant courses. Partnering with Project Eureka! the project also engages high-school girls in computer programming with the aim of broadening participation in computing. The code base containing the SNN simulators, performance monitors, along with the analytical models is open source and will be maintained as such for five years. Software deliverables, including stable releases of the simulators, will be made available to the broader research community via a public repository. This repository will also contain the various SNN-benchmark applications and their execution traces, as well as tutorials and documentation. The URL for the repository is https://github.com/drexel-DISCO. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.