A new company launched Monday by former NASA chief Dan Goldin aims to deliver a major boost to the field of neural computing.
KnuEdge's debut comes after 10 years in stealth; formerly it was called Intellisis. Now, along with its launch, it's introducing two products focused on neural computing: KnuVerse, software that focuses on military-grade voice recognition and authentication, and KnuPath, a processor designed to offer a new architecture for neural computing.
"While at NASA I became fascinated with biology," said Goldin in an interview last week. "When the time came to leave NASA, I decided the future of technology would be in machine intelligence, and I felt a major thrust had to come from inspiration from the mammalian brain."
Goldin's startup began by focusing on speech recognition in the presence of noise.
KnuEdge's technology, which has already been tested under military conditions, makes it possible to authenticate a person's voice for computers, Web and mobile apps, and IoT devices with only a few words spoken into a microphone in any language and in real-world noisy conditions, the company said. Potential applications span industries such as banking, entertainment and hospitality.
"We don't care if you have a cheap or expensive microphone," he said. "It can go on anyone's computational platform, and we're beginning to approach error rates measured in parts per billion."
Whereas most work in neural networks focuses on information structures known as "dense" matrices, KnuEdge tapped "sparse" matrices instead, Goldin said.
"As far as we know, the sparse matrix in the human brain is the most efficient," he said. "We used that as a model. We felt it was important to build a model that replicated how the brain works."
KnuEdge's KnuPath LambdaFabric processor technology, meanwhile, was inspired by a roadblock the company encountered while working on its voice recognition offering. Essentially, the company realized it could not achieve the performance it needed with traditional CPUs and GPUs, it said, so it created a new team to build an application-specific integrated circuit (ASIC).
KnuPath processors are low-wattage, 256-core chips with 16 bidirectional I/O paths, the company said. With the ability to scale up to 512,000 devices, they offer rack-to-rack latency of just 400 nanoseconds.
"Each core can run a different calculation simultaneously," Goldin said.
KnuPath can operate alone or be integrated with other devices. Potential applications include signal processing and machine learning for the Internet of Things (IoT).
Production platforms and systems will be available in the second half of this year. KnuEdge plans to make its application programming interfaces (APIs) available worldwide on a freemium model.
So far, KnuEdge claims private funding totaling US$100 million and $20 million in revenue.
KnuVerse differs from Siri or Google Now in that it focuses on improving audio quality to the point where it can be used for secure recognition of an individual in any environment, said Jim McGregor, principal analyst with Tirias Research.
"Think of it as a version of Google Now on steroids that has been proven by the U.S. government and military," McGregor said.
KnuPath, on the other hand, is "essentially a new processing architecture designed around creating intelligence from massive amounts of data, or what is typically called deep learning," he said. "This is a move away from the CPU paradigm to a entirely new form of computing."
The most obvious competitor in that space today is IBM, but there are other companies developing neural network technology as well, McGregor noted. "Look for a wave of innovation in this area over the next few years."
Together, KnuEdge's technologies could help create smart cities, he added. While many details remain to be revealed, "seeing new technology like this makes it an exciting time to be in the industry."