Hierarchical, Compositional Corelet Programming for Neuromorphic Chips
Arnon Amir, IBM
Recent results in deep convolutional neural networks on a new, spiking, core-based neuromorphic architecture approach state-of-the-art classification accuracy on a number of datasets while processing more than 1200 frames per second, (yet consuming only 25 to 275 mW). The corelet programming language (CPL) was developed to program this highly efficient, non von-Neumann architecture. Using MATLAB® object-oriented programming, CPL provides network abstraction, encapsulation, and efficient hierarchical composition of reusable network components (corelets). The talk introduces CPL, its development environment (CPE), the corelet library, and a short live demo as time permits. CPE has been released and taught to a community of more than 100 developers from dozens of institutes around the world, who use it to develop new neuromorphic algorithms, networks, and systems.
Recorded: 3 Nov 2016