Artificial neural networks, a computational framework that mimmics biological learning processes using statistics and large data sets, are behind many of the technological marvels of today. Google is famous for employing some of the largest neural networks in the world, powering everything from their search recommendations to their machine translation engine. They’re also behind numerous other innovations like predictive text inputs, voice recognition software and recommendation engines that use your previous preferences to suggest new things. However these networks aren’t exactly portable, often requiring vast data centers to produce the kinds of outputs we expect. IBM is set to change that however with their TrueNorth architecture, a truly revolutionary idea in computing.
The chip, 16 of which are shown above welded to a DARPA SyNAPSE board, is most easily thought of as a massively parallel chip comprising of some 4096 processes cores. Each of these cores contains 256 programmable synapses, totalling around 1 million per chip. Interestingly whilst the chip’s transistor count is on the order of 5.4 billion, which for comparison is just over double of Intel’s current offering, it uses a fraction of the power you’d expect it to: a mere 70 milliwatts. That kind of power consumption means that chips like these could make their way into portable devices, something that no one would really expect with transistor counts that high.
But why, I hear you asking, would you want a computerized brain in your pocket?
IBM’s TrueNorth chip is essentially the second half of the two part system that is a neural network. The first step to creating a functioning neural network is training it on a large dataset. The larger the set the better the network’s capabilities are. This is why large companies like Google and Apple can create useable products out of them, they have huge troves of data with which to train them on. Then, once the network is trained, you can set it loose upon new data and have it give you insights and predictions on it and that’s where a chip like TrueNorth can come in. Essentially you’d use a big network to form the model and then imprint on a TrueNorth chip, making it portable.
The implications of this probably wouldn’t be immediately apparent for most, the services would likely retain their same functionality, but it would eliminate the requirement for an always on Internet connection to support them. This could open up a new class of smart devices with capabilities that far surpass anything we currently have like a pocket translator that works in real time. The biggest issue I see to its adoption though is cost as a transistor count that high doesn’t come cheap as you’re either relying on cutting edge lithography or significantly reduced wafer yields. Both of these lead to high priced chips, likely even more than current consumer CPUs.
Like all good technology however this one is a little way off from finding its way into our hands as whilst the chip exists the software stack required to use it is still under active development. It might sound like a small thing however this chip behaves in a way that’s completely different to anything that’s come before it. However once that’s been settled then the floodgates can be opened to the wider world and then, I’m sure, we’ll see a rapid pace of innovation that could spur on some wonderful technological marvels.