Nengo: The Neuromorphic Compiler and the Rise of Neuromorphic Computing Chips
Another article (see below) came out on the increased activity in neuromorphic computing, including mention of our company Applied Brain Research Inc. Whether this chip, IBM's, Intel's announcement or the even more advanced research projects such as NeuroGrid, these are exciting times for neuromorphs everywhere!
Where do we Applied Brain Research (https://appliedbrainresearch.com/) fit in? Our role is as the operating system and software compiler for neuromorphic chips. We also are the developers of several leading neuromorphic applications such as REACH; our robotic control system which merges robotic control theory, deep learning, online learning and adaptive control together. This yields a robotic control system that is ideal for human-robot interaction as it remains compliant during motion so is less dangerous to work with. The adaptive control adapts the control signals in areas of the operational space where the deep learning training cases did not cover it well. The result is a control system that stays under control where pure deep learned robotic control has unpredictable failure points and requires many times the number of training cases. Studying the brain, our only know example of intelligence, has definite engineering benefits. It is still early but the advances being made by all the companies mentioned and ours are changing the face of computing forever, opening up a whole new world of possibilities.
This discussion isn't about deep learning versus neuromorphics. It is about how do we together use all the best tools we have to create solutions that solve pressing problems enabling us all to do so much more. Nengo (https://www.nengo.ai/) was pioneered by our team of scientists who reverse engineered brain circuits to build it, long in advance of the kind of computing we need to quickly run the size of systems they have built. Since its inception, they have been running their models on super computers, GPUs and PCs. None of these are structured the way the brain is; massively parallel, using a send- information-only-when-needed spiking architecture the combination of which enables truly amazing neural networks that run on less power. This architecture is another way we can extend Moore's Law. The coming age of neuromorphic hardware which uses all the above brain design principles is a super thrilling event for us. Finally, hardware at the performance scale of our algorithms.
An era is now arriving where we can see that massively parallel neuromorphic algorithms could be run in real-time. This will accelerate both the hardware and software to enable more dynamic, real-time, low-power, adaptive AI for robotic control, vision, cognitive computing and other applications.
For more on our company see our website above and our intro video at https://www.youtube.com/watch?v=zGpwSeMtCQc .
To start learning Nengo please see https://www.nengo.ai/.
See the article below on another form of neuromorphic chip coming;