Artificial Intelligence (AI) is everywhere. That is to say, the media appear to be writing about AI on a daily basis and many companies are working to understand what AI truly is. However, today's AI is essentially a set of software-based rules, programmed by humans. It may be artificial, but we wouldn't call it intelligent. True AI implies an entirely autonomously learning system that can recognise patterns without any pre-programmed set of rules. The scope of the latter is limited by the programmer's coding skills and imagination. ASX-listed BrainChip (ASX:BRN) has developed a hardware-only AI solution that can truly learn autonomously by being fed data sets, such as video feeds and sounds. Working through sufficient amounts of data, this chip starts to recognise and associate patterns by itself. The key advantage is that the chip will look for any sort of pattern, as opposed to rules-based AI solutions that are programmed to only look for specific patterns. The number of application areas for this type of AI is endless, which is why we believe BRN has a bright future ahead.
Replicating nature in silicon
The key aspect of BRN's technology is that it mimics a biological neuron, or actually several thousands of biological neurons. Neurons in our brain pass on input signals, or spikes, to neighboring neurons depending on the responses they got to earlier spikes they forwarded. The type of spikes that get passed on by neighboring neurons are apparently considered more relevant by these neighboring neurons and get attributed a higher weight by the particular neuron that forwarded these spikes. Conversely, spikes that are not forwarded by neighboring neurons are apparently less relevant and receive a lower value. This process of attributing, so-called, synaptic weights to incoming spikes allows biological neurons to learn autonomously.
BRN has replicated this process in a computer chip design that can contain several thousands of artificial neurons. The company's Spiking Neuron Adaptive Processor, or SNAP, is a hardware-only computer chip solution that doesn't require any software code to operate.
Faster, cheaper and smaller
One of the main advantages of BRN's chip compared to software-based AI is speed. Software-based AI programs, such as Deep Learning, run on computers that work sequentially meaning a program is executed one step at a time. So while the processor is calculating, the entire program needs to wait for the outcome of that calculation before it can use that outcome in the next step of the program. However, like the human brain, BRN's chip works in parallel. All artificial neurons on the chip can fire and receive spikes simultaneously, which leads to a tremendous speed advantage. A pattern that may take BRN's chip minutes to recognise and classify, may take a software-based AI program hours or even days.
Another advantage is cost. Deep Learning programs, like IBM's True North, require very substantial amounts of computer power that can fill entire floors with computer equipment. Not only is the Capex for these computers extremely high, the Opex is very substantial as well because of the high energy consumption of these systems. Given BRN's speed advantage, the required Capex to achieve the same result will be substantially lower.
Lastly, there is a clear size advantage to BRN's solution. Obviously, a room full of computer equipment doesn't fit well in mobile devices and on equipment at the edge of the Internet of Things, such as sensors. BRN's chip solution is small enough to be embedded in mobile phones, for instance for voice recognition, either stand alone or as part of a System-on-a-Chip (SOC).
Endless application areas
The number of applications areas for BRN's technology is sheer endless and includes face and voice recognition in mobile devices and security applications, fault detection in engines through sound patterning, autonomous vehicles, robotics, computer gaming, body implants, weather forecasting, biotech and genetics etc. Before being deployed or embedded, the chips can be trained in a specific application area, e.g. face recognition, by feeding the master chip relevant data sets from which to learn. The resulting synaptic weights can subsequently be copied to all chips manufactured for that specific purpose.
There is nothing artificial about BRN's addressable market
The market for highly specialised neuromorphic chips is expected to reach US$ 5BN by 2022, from just US$ 500M in 2014. However, we believe the broader market for embedded AI is likely to be many times larger. Consequently, we believe the future for BRN looks very bright indeed.
TMT Analytics provides institutional grade equities research for companies in the Technology, Media and Telecom sectors as well as bespoke sector research.
Find out more about TMT Analytics at (VIEW LINK)