Neuromorphic intermediate representation: A unified instruction set for interoperable brain-inspired computing

Area of Science:

• Neuromorphic Computing
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) and neuromorphic hardware are gaining traction for machine learning applications.
• Variability in current software and hardware solutions hinders reproducibility of SNN research.
• A unified framework is needed to bridge mathematical models and diverse implementations.

Purpose of the Study:

• To establish a common reference frame for computations in digital neuromorphic systems.
• To create a standard that abstracts discretization and hardware constraints for SNNs.
• To enhance interoperability and accessibility across neuromorphic platforms.

Main Methods:

• Introduction of Neuromorphic Intermediate Representation (NIR) as a computational standard.
• NIR defines composable model primitives for hybrid systems (continuous-time dynamics and discrete events).
• Demonstration of NIR by reproducing three SNN models across 7 simulators and 4 hardware platforms.

Main Results:

• NIR successfully captures computational models, abstracting away implementation-specific details.
• NIR demonstrated broad compatibility, supporting numerous neuromorphic systems.
• Reproducibility of complex SNN models was achieved across diverse hardware and software.

Conclusions:

• NIR decouples neuromorphic hardware and software development, fostering interoperability.
• NIR improves accessibility to various neuromorphic technologies for researchers.
• NIR represents a significant advancement for brain-inspired computing, promoting energy-efficient systems.