In Melbourne, Australia, Stefan Harrer is running an artificial software brain atop an artificial hardware brain in an effort to analyze a brain that isn’t artificial at all. Ultimately, he and his colleagues envision merging these three brains together so that the artificial can augment the real.
Harrer is an IBM researcher stationed at the company’s Australian research lab. Together with neurologists at the University of Melbourne, he’s developing a computing system that can analyze your brain waves in an effort to predict epileptic seizures.
‘Our aim is to replace broken neural systems with machines.’
The trick is that Harrer and his colleagues are building the system using a neural network, computer software that mimics the web of neurons in the human brain. This is the same breed of neural networkthat identifies photos you post to Facebook, recognizes the commands you speak into your Android phone, and more. If you feed a photo of your mother into a neural network, it can learn to recognize your mother. And now, Harrer is feeding scans of brain waves into a neural network, in the hopes that it can learn to recognize epilepsy.
“We’re trying to extract all the meaningful information from all the background noise,” Harrer says. “We want to be able to detect a specific seizure for a specific patient.”
But there’s also another trick. Harrer and team are running this neural network on an experimental IBM chip called TrueNorth, which, like the neural network they’re using, is built in the image of the human brain. Because it uses a similar architecture, TrueNorth is rather adept at running neural networks. And because it consumes very little power, Harrer and team hope to one day use the chip to build a wearable device that, working in tandem with a brain implant, monitors for seizures around the clock and notifies patients before they happen.
“We want to do this on a wearable system that you put on a subject—on a patient—and have it do analysis in real-time, 24/7,” Harrer says. “That’s the only way this technology will have an impact beyond cool research papers.”
That may sound like science fiction. But at the University of Melbourne, neurologists have already run a study in which a less complex implant gathered EEG readings from epilepsy patients over the course of about three years. This data, in fact, is what Harrer and team are using to train their neural network.
We’re still a long way from a time when we can attach an (artificial) neural network to a human body—Harrer’s work is still in the preliminary stages—but that’s the ultimate aim here. And that’s certainly doable, according to Kimford Meador, a neurologist at Stanford University Medical Center who has no connection to the work in Australia. “If you have an implant near the seizure’s origin,” Meador says, “you can detect them quite reliably.”
Harrer’s project, which is described in a peer-reviewed paper being presented at theACM Computing Frontiers conference in May, is part of sweeping movement among companies and researchers to develop so-called deep neural networks. Along with the widespread use of deep neural nets at Facebook and Google, Twitter is using them to identify pornography on its social network. Microsoft is using them to translate Skype calls from one language to another. And academics at the University of California at Berkeley are using them to teach robots to screw on bottle caps. The difference is that Harrer is using fake brains to analyze real brain waves—and that he’s experimenting with TrueNorth, a chip that’s not yet available on the open market.
Herrar’s work is particularly intriguing because it’s difficult for today’s machines to automatically predict seizures on the fly, in part because they aren’t using the latest in AI. “No one has really applied machine learning to this kind of task,” Meador says. But there is a clear path using such machine learning. As Meador explains, manually detecting a seizure just before it happens isn’t that difficult—at least where some patients are concerned. So, if you have data from seizures, you should be able to teach a neural network to detect on its own. Thanks to that earlier study from the University of Melbourne, Harrer and team have the data.
The harder part, Meador says, is trying to predict the probability of seizure well ahead of time. That takes some much deeper analysis. But it’s also more useful, and indeed, this is part of what Harrer and his colleagues at the University are striving for. Because of new insights that have arisen since that original implant study, says Mark Cook, the neurologist at the University of Melbourne who led the study, they “should be able to see deeper into the structure underlying the seizure activity.”
Fixing Broken Systems
But the tech that makes all this practical is the TrueNorth chip. Nowadays, companies like Google, Facebook, and Microsoft typically run their neural networks across myriad machines inside massive computer data centers. They train the neural network inside the data center. When you allow Facebook to identify people in your photos, for instance, your laptop, tablet, or phone is communicating with those neural nets via the Internet. With TrueNorth, IBM aims to make it easier to run neural networks on the laptop, tablet, or phone itself—and maybe on a wearable that talks to an implant in your head.
The idea is that, after detecting brain wave patterns that indicate a seizure, the device would notify you by sending a wireless signal to your smartphone. And even just a little bit of notice before a seizure happens can be beneficial to patients, says Dean Robert Freestone, a senior research fellow at the University of Melbourne who is working closely with Harrer. “It can give them a new sense of freedom and eliminate a lot of risk in their lives,” he says. The device could provide enough notice to, say, pull your car to the side of the road before a seizure happens.
Harrer acknowledges that this full-blown warning system is still years off. But modern AI technology keeps getting smarter. He and his collaborators even think that their implants could one day be used to prevent seizures entirely: the system would detect a seizure coming on and send out electrical impulses in order to stop it. In essence, an artificial brain could end up augmenting and improving a real brain. “Our aim is to replace broken neural systems with machines—machines that can interact with the brain in a very natural way,” says Freestone.
This too is entirely doable, says Stanford’s Meador. At least one Silicon Valley company is already offer a device that will stimulate the brain in an effort to curb seizures, he says. But it’s not nearly as sophisticated as the device that Harrer envisions. “This could let us to something that we can’t do at the moment,” Cook says. “It provides us with a better way.”