Neural networks can do some impressive things these days. They can describe images, translate speech, play videogames, and even have dreams. However, these networks are huge, and can be extremely slow to run. The DeepDream software, running on my desktop PC takes about ten minutes to process an image. When Google Now or Siri want to understand your speech, they have to stream the data off to a supercomputer elsewhere, producing a frustrating delay.
It’ll be a long time before the processor in your phone is able to run these million-neuron networks at a reasonable speed. Given the impending slowdown in Moore’s Law, it might take a very long time indeed. So, how can we bring the power of these algorithms to bear in the mobile world? IBM thinks it has an answer.
A Brain in a Box
The machine is the size of a slimmed-down mini-fridge, or a medicine cabinet. It consists of 48 little boxes, each about the size of a computer hard drive. The boxes contain computer chips that are, well, a little weird. IBM calls it “TrueNorth.” It’s a specialized chip designed, from the ground up, to run neural networks. Implemented in the silicon are physical analogs of neurons and synapses (the connections between neurons).
Each chip simulates about one million neurons, with 256 million synapses between them. Together, the 48 million silicon neurons exceed the number of brains in the cerebral cortex of a rat (which max out at about 21 million). That’s a heck of a lot of learning power in such a small box.
Rather than using software to simulate the behavior of a neural network, these chips cut out the middle man and build the “neurons” directly out of silicon. That has a lot of advantages. Normally, a neural network of this size would be run on racks and racks of servers. These installations can use as much power as a city block. In contrast, each TrueNorth chip draws on 70 milliwatts of power — so little that it can run flat out for a week on a standard cell phone battery. It draws more than one hundred times less power than a high-end Intel CPU. Together, the 48-processor machine likely consumes less energy than your PC.
Making Intelligence Mobile
While the current machine is pretty bulky by mobile standards, it’s also a very early prototype that doesn’t push the limits in terms of chip fabrication. In the near future, it’s very likely that such a machine could be made even smaller and lower powered, letting companies embed it in mobile devices, or intelligence-intensive mobile machines like self driving cars and autonomous drones. This would radically change the way artificial intelligence is used, taking the technology back from the cloud.
Unfortunately, the chips don’t allow for the network to be trained using hardware acceleration — yet. You’ll still need a traditional supercomputer to allow the network to learn to perform a task. However, once a network has been trained, its behavior can be executed easily on a TrueNorth chip. In the future, these chips may be able to implement the learning algorithms themselves in hardware, allowing for the creation of true, learning “electronic brains”. It remains to be seen how long it’ll take to make this technological leap, though I don’t know of any specific obstacles.
Needless to say, these sorts of chips require new ways of thinking about software. That’s why IBM built the brain-scale machine. It’s a demonstration for a boot camp they’re hosting for researchers from 30 different research organizations on five continents. At the end of the symposium, the machine will be disassembled, and each processor will go home with a different research institution. Their job is to figure out how to build good software for this strange, organic chip. At a hackathon held in Colorado, researchers were able to get the chip to recognize images and speech in only a few days, suggesting that current “Deep Learning” research will largely carry over. That said, there’s still a lot of work to be done on how to get the most out of these chips.
A Chip For Every Season
If rumors of the death of Moore’s Law bear out, this sort of technology is going to become increasingly important to push the frontiers of machine intelligence. A human brain has about 100 billion neurons. If a future version of TrueNorth could learn independently, and we fully understood the algorithms that power the brain, you’d need about 2000 of the 48-chip machines in order to be able to rival a human brain for power.
But that’s not out of the question. According to some napkin math, that machine would fit in a room. It would draw less power than an electric car. If you want to live in a future of intelligent machines, TrueNorth shows a path to achieving that goal, even if Moore’s Law falters. For now, it’s still poised to revolutionize the way we use AI in our cars and mobile devices. That’s pretty exciting all by itself.
What do you think? Is dedicated hardware the future of AI? Let us know your thoughts in the comments below.
Image Credits: mouse by kontur-vid via Shutterstock