A team of researchers at Oak Ridge National Laboratory (ORNL) have developed a neuromorphic platform to enable neuromorphic algorithms and applications research. The platform includes a Python and C++ based API, a high-performance spiking neural network simulator, and FPGA accelerator integration. To demonstrate the capabilities of Caspian, the researchers implemented an evolutionary algorithm (EONS) and a back-propagation-based algorithm (SLAYER) for training Caspian networks. The researchers also analyzed the performance characteristics of the Caspian simulator and showed improved execution performance compared to a previous event-based simulator.
Significance and Impact
This work enables advanced algorithms and application research for neuromorphic computing. The provided simulation engine allows for neuromorphic training and testing on existing computing systems, and the resulting models can then be efficiently deployed in hardware using of the shelf FPGAs. Researchers can use the Python API to quickly prototype concepts in simulation to determine feasibility before specific neuromorphic hardware is needed.
- Researchers created a high level neuromorphic API and spiking neural network simulator for advanced algorithm development and rapid prototyping of applications.
- Researchers demonstrated both an evolutionary algorithm approach (EONS) and a back-propagation-based approach (SLAYER) for training Caspian networks.
- Researchers analyzed the runtime performance of the Caspian simulator and showed improved scalability and overall execution time as compared to a previously developed simulation engine.
- Researchers discussed how this platform can enable efficient hardware deployments by using a custom neuromorphic architecture on an FPGA.
Citation and DOI
J. Parker Mitchell, Catherine D. Schuman, Robert M. Patton, and Thomas E. Potok. "Caspian: A Neuromorphic Development Platform" Neuro-Inspired Computational Elements (NICE) Workshop 2020.
Caspian provides a development platform for neuromorphic algorithms and applications research. The C++ and Python APIs allow for quick and flexible development of new algorithms and applications. Initial training can be done in simulation and then deployed to an FPGA platform for fast, efficient inference in hardware. In this work, researchers demonstrate using both an evolutionary algorithm (EONS) and a back-propagation-based algorithm (SLAYER) to generate well performing spiking neural networks within the Caspian platform. Researchers also discuss future work to develop a series of FPGA designs compatible with the developed programming environment.
Last Updated: June 18, 2020 - 9:30 am