Highlight

Automated Design of Neuromorphic Networks for Scientific Applications at the Edge

Left - Memristive neuromorphic system	Right - Spiking neural network designed by EONS
Left - Memristive neuromorphic system | Right - Spiking neural network designed by EONS

Achievement

A team of researchers at Oak Ridge National Laboratory (ORNL), Purdue University, and the University of Tennessee utilized a previously presented approach, EONS, to design spiking neural networks for a memristive neuromorphic implementation for scientific data applications.  The researchers used a multi-objective approach in EONS to maximize network accuracy on the scientific data application task and minimize network size and energy.  They illustrated that EONS determines both the network structure and the parameters, removing the burden from the user on determining the appropriate spiking neural network structure, and they show that the resulting networks are very different from the layered structure of typical neural networks. They also showed that the multi-objective approach produces smaller, more energy efficient networks than the original EONS approach and produces comparable accuracy to a back-propagation style training approach.

Significance and Impact

This work demonstrates an approach that can automate the design of spiking neural networks for deployment into low size, weight, and power neuromorphic hardware implementations for scientific applications. This approach will lower the barrier of entry for scientists to utilize neuromorphic systems for their own applications.

Research Details

  • Researchers extended a multi-objective EONS approach by demonstrating the approach specifically for the automated design of small, energy efficient SNNs for a memristive neuromorphic system for scientific data classification tasks, with an edge deployment scenario specifically in mind. 
  • Researchers showed that the new approach produces networks that perform equivalently on accuracy but are significantly smaller and significantly more energy efficient than the networks trained by the original EONS algorithm and networks trained using a back propagation-based algorithm.
  • Researchers showed that it is sufficient to minimize either size of network or energy usage in order to achieve networks that are both smaller and more energy efficient.

Citation and DOI  

Catherine D. Schuman, J. Parker Mitchell, Maryam Parsa, James S. Plank, Samuel D. Brown, Garrett S. Rose, Robert M. Patton, and Thomas E. Potok. "Automated Design of Neuromorphic Networks for Scientific Applications at the Edge." International Joint Conference on Neural Networks (IJCNN) 2020. 

Overview

Designing spiking neural networks for neuromorphic deployment is a non-trivial task.  It is further complicated when there are resource constraints for the neuromorphic implementation, such as size or power constraints, that may be present in edge applications.  In this work, researchers utilized a previously presented approach, EONS, to design spiking neural networks for a memristive neuromorphic implementation for scientific data applications.  They specifically used a multi-objective approach in EONS to maximize network accuracy on the scientific data application task, but also to minimize network size and energy.  They illustrated that EONS determines both the network structure and the parameters, removing the burden from the user on determining the appropriate spiking neural network structure, and showed that the resulting networks are very different from the layered structure of typical neural networks. Finally, they showed that the multi-objective approach produces smaller, more energy efficient networks than the original EONS approach and produces comparable accuracy to a back-propagation style training approach.

Last Updated: May 28, 2020 - 4:01 pm