Chairman & Co-CEO, Applied Brain Research Inc.
We make AI smaller, greener and smarter, inspired by decades of brain research.
Smart devices require AI to be much smaller, faster and more capable. Voice interaction with cars, phones, appliances and all devices is convenient but requires structured conversation, an internet connection, isn’t private and costs device makers a lot for in cloud compute for years after sale. What if a full AI speech recognition system could be combined with natural language understanding all on one small computer chip that coat $2 to make? Then devices could talk all the time, privately, fluidly, quickly and completely to their owners and no cost the device maker. Gone would be those frustrating structured command word sequences that offline devices have to use now. Want something from your device? Just talk to it. Say it however you want, knowing that the sound is not going to a server in the cloud and it will respond without that awkward pause from online speech systems. It the device needs to get something from the internet it can send just the request call, not your voice or sounds from your home. The maker doesn’t pay for cloud compute costs for ongoing device interactions and needs much less computing assets to support the deployed devices.
We have designed this full speech and language chip and costed it out, spoken to major car companies, cell phone and wearables makers and we are making it. We are raising a Series-A to fab the chip and expand our sales, marketing and software teams. We already have customers licensing our algorithms in wearables and speech for medical devices.
So how do we fit all that AI into such a small chip?
One of things we discovered in our brain research is a mathematical model of how brains store representations of streams of events. This mathematical model it turns out is the optimal way to compress data over time. We named it the Legendre Memory Unit and we have implemented it in software and in chip form. We have built keyword spotting (KWS), automated speech recognition (ASR), natural language processing (NLP) and time-series data processing software systems with it, and they are all smaller, faster, as or more accurate and consume less power and memory space than the current state-of-the-art (SotA) networks. For NLP models we achieved 10x reduction in training compute resources and our NLP scales with the sequence length of the text. O(N) compute and O(1) memory scaling, whereas in production SotA NLP scales as O(N^2).
This means those large language models that have become a politically-charged, climate change topic, can now be processed for 1/10th the power. This is needed as the models are now 500x larger than a year ago and the move is now to continual training to customize and personalize the model constantly. This is s desperately needed innovation. We are in evaluation at two of the top 5 tech firms globally with this technology as of this month. We have patented our LMU globally and our patents have been granted in Canada and the USA already and the global applications are being processed.
We have a complete deployment stack for our technology coming out with the chips as well, called NengoEdge, a SaaS platform extension of our popular real-time AI compiler www.Nengo.ai. Nengo is the world’s leading neuromorphic research compiler. The commercial market for neuromorphics will open in the coming years as well but now we are focused on our LMU chip. We have one of the world’s leading real-time AI teams. The LMU is our latest discovery in a long line of dynamic AI firsts for our team. We built the world’s largest working brain model, Spaun using Nengo. Now we have found a massive set of commercial problems in AI size and efficiency our LMU algorithm and other methods can solve today.
Let’s make AI work better for everyone by making it more interactive, greener and smarter!