How would you program a board with 8 million spiking neurons?

Intel Labs Neuromorphic computing team headed by Mike Davies released a series of exciting posts showing the scale and efficiencies that the Intel neuromorphic group has achieved with their neuromorphic chip Loihi and the 8 million neuron board built with Loihi chips, named Poholki Beach. ( https://www.dhirubhai.net/feed/update/urn:li:activity:6556717856871055361/ ).

But how do you build faster, lower-power, real-time, adaptive AI systems with these emerging neuromorphic chips like Loihi and the Poholki Beach boards? Use the same system that first programmed spiking deep learning, reinforcement learning and adaptive control on Loihi, ABR's Nengo visual neuromorphic development platform (download free for academic and personal use at https://www.nengo.ai).

The team at our company, Applied Brain Research (ABR - https://www.AppliedBrainResearch.com ), has been doing leading research into neuromorphics for over 20 years and have been developing Nengo, our neuromorphic compiler for over 15 years. It has taken a long time and lots of work from many leading labs around the world, including the one we launched our company out of at the University of Waterloo, to progress our understanding of spiking biological networks well enough that the last decade has seen spiking chips now yielding commercially relevant performance and use cases. Intel's announcement and our collective performance benchmarks are leading the way in making neuromorphics a commercially valuable architecture. We have been working for and with Mike's team at Intel Labs since 2015, developing Nengo's Loihi backend and spiking applications on Loihi.

So back to how do you get started programming all those independent spiking neurons on these new chips? Our Nengo neuromorphic compiler can be used to build, run and debug small and large neuromorphic systems visually in real-time. Nengo models run on many platforms including Loihi, CPU, GPU and FPGA, so you can build a model once and then run it where you need to. For the best performance, run it on Loihi but you can develop it today for Loihi even before it is released using the Nengo Loihi emulator available now at https://www.nengo.ai/nengo-loihi/. Note as well the chips do not provide all the secrets to how to get the performance from a spiking chips, it matters how you use the neurons. There are many required parameters for each neuron, including connections in and out, weights and time constants. Nengo takes care of that for you. Allowing you to program at a high level with the help of Nengo's visualization graphical interface while still allowing you to set the low-level parameters when you need to.

The visual debugging of models in Nengo is one feature our users love. The only way to visually and interactively run and debug your spiking models is with Nengo. Orchestrating millions of neurons is a mental challenge, doing it visually and interactively is a lot easier.

The largest Nengo model built to date, the Spaun brain model, has 6.5 million spiking neurons structured into a many integrated dynamic neural ensembles. For a visual of the Spaun system inside the Nengo visual development tool, see the fly through of Spaun included in the introduction video at the bottom of the https://www.nengo.ai webpage.

For commercial applications, neuromorphics is showing its value for low-power, adaptive, always-on edge-AI applications such as low-power edge deep learning for vision and audio processing, drone control and edge sensor integration. Using these capabilities we are developing leading real-time, edge-AI applications for cars, drones, phones, robots and cameras. Intel's results and ours ( https://arxiv.org/abs/1812.01739 ) show the progress neuromorphics has made in the past few years to enabling these development activities. Neuromorphics performance is now commercially relevant and the architecture has much to offer in the coming years.

Companies looking to make their sensor, vision, audio and control processing lower-power, more scalable, higher speed and adaptive, should consider neuromorphic solutions. We can help your team understand what neuromorphics can do today and in the coming iterations and help you get started well ahead of your competition, before the coming rush when neuromorphic chips become widely available. If you release a product first whose battery lasts longer, sensors detect faster & adapt better, or whose drone flies longer,... that can change your market share. We can help understand where and when to use neuromorphics for your situation and if it makes sense, get you started immediately on bringing the best neuromorphic computing has to offer to your products.

Visit us today at www.appliedbrainresearch.com to learn more, or email us at [email protected] for more information.