The company has told eeNews Europe that it has developed its first core, the SNAP-64 processor, and plans to follow an "ARM-like" business model. In other words it will license its technology and circuit designs for use by semiconductor companies and will receive a royalty – typically a percentage of the chip price – on each chip sold that includes its intellectual property.
The company’s technology, based on spiking neural adaptive networks, has been in development for many years although the company was founded in December 2013. It is largely the result of several years work by Peter Van Der Made, chief technology officer and interim CEO. Spiking neural networks are modelled much more closely on human brain activity that conventional weighted neural networks in that they transmit information by a series of pulses from neuron to neuron.
Unusually for a startup, BrainChip is already a public company having been supported and funded through a reverse takeover by a former Australian minerals exploration company, Aziana Ltd. Earlier this year Aziana acquired BrainChip Inc and changed the name to Brain Chip Holdings Ltd. but brings with it cash resources and a listing on the Australian Securities Exchange (ASX).
BrainChip is not the first company to seek to implement neural networks in hardware. IBM has announced its TrueNorth chip (see IBM True North puts brain on a chip), Qualcomm’s Zeroth processor core is included on its Snapdragon 820 system chip, and Intel has included a neural network core for pattern classification inside its Quark SE chip courtesy of a licensing deal. NeuroMem Inc. (Petaluma, Calif.) offers the technology for license and in chip form (see Startup’s tech is Intel’s Quark neural network).
However, BrainChip claims its model of brain activity is superior and its performance faster because of the fact that it uses a spiking model of the neural network information transfer. That said, Qualcomm claims its Zeroth processor core has also evolved from focusing on biologically realistic spiking neural networks to also include artificial neural networks for on-device deep learning.
BrainChip also claims that it is the first company to hold a granted patent on digital spiking neural network technology (see Patent number US 8,250,011 “Autonomous learning dynamic artificial neural computing device and brain inspired system” Neil Rinaldi, former CEO of Aziana and now non-executive director of BrainChip Holdings Ltd, told eeNews Europe that BrainChip Inc. has been drawing up protection of its intellectual property. The company has five additional patents pending and a further 76 patents in the process of being filed.
The SNAP-64 has been implemented in FPGA and the company is looking to implement the core in a leading-edge manufacturing process at a foundry. "The expectation is that many licensees will want to implement the core as co-processor in a system chip. On-chip it can be configured to meet a wide variety of applications." BrainChip is also prepared to engage with more advanced customers who wish to adapt the overall architecture for specific applications.
"We have engaged discussions with a number of companies in the mobile phone and IoT sectors," said Rinaldi. Other applications include robotics and autonomous learning machines used in exploration and unmanned vehicles, speech recognition, extraction of speech and sound from a noisy background and visual image recognition.
The SNAP-64 has been designed and run in software simulation as well as implemented on an FPGA but technical questions remain as to the complexity it represents in terms of neuron and synapse connections, the resolution of the stimulus and spiking signal resolution, and the interface provided between the SNAP-64 and conventional Von Neumann processors.
The company is looking to appoint a permanent CEO.
Startup plans neural network circuit for low-power wireless sensors
Startup’s pattern matching tech is inside Intel’s Quark neural net
IBM’s Watson to help CVS manage patient health data