This site is reader-supported. When you click through links on our site, we may be compensated.
When Elon Musk first started talking about launching a brain-computer interface company, he made a number of comments that set expectations for what that idea might entail. The company, he said, was motivated by his concerns about AI ending up hostile to humans: providing humans with an interface directly into the AI's home turf might prevent hostilities from developing. Musk also suggested that he hoped to avoid any electrodes implanted in the brain, since that might pose a barrier to adoption.
At his recent public launch of the company (since named Neuralink), worries about hostile AIs did get a mention—but only in passing. Instead, we got a detailed technical description of the hardware behind Neuralink's brain-computer interface, which would rely on surgery and implanted hardware. In the process, Neuralink went from something in the realm of science fiction to a company that would be pushing for an aggressive evolution of existing neural-implant hardware.
Those changes in tone and topic are a sign that Musk has been listening to the people he hired to build Neuralink. So, how precisely is Neuralink pushing the envelope on what we can already do in this space? And does it still veer a bit closer to science fiction in some aspects?
The big picture
Before taking a look at the individual components that Neuralink announced recently, let's start with an overview of what the company hopes to accomplish technology-wise. The plan is to access the brain via a hole less than eight millimeters across. This small hole would allow Neuralink to implant an even smaller (4mm x 4mm) chip and its associated wiring into the brain. The chip will get power from, and communicates with, some wireless hardware located behind the ear, much like current cochlear implants.
Inside the brain, the chip will be connected to a series of small threads that carry electrodes to the relevant area, where they can listen in on the electrical activity of neurons. These threads will be put in place using a surgical robot, which allows the surgeon to insert them in a manner that avoids damaging blood vessels.
The chip will take the raw readings of neural activity and process them to a very compact form that preserves key information, which will be easier for their wireless hardware to transmit back across the skull. Electrical impulses can also be sent to the neurons via the same electrodes, stimulating brain activity. Musk thinks that it would be safe to insert as many as 10 of these chips into a single brain, though Neuralink will obviously start testing with far fewer.
All of that is an evolution of some of the existing work on brain-computer interfaces. But the details behind some of these features provides a better sense of how Neuralink is pushing the field forward.
The Neuralink introduction included a video of the brain during surgery, revealing how the wrinkly organ constantly shifts with breathing and blood flow. This makes implanting electrodes a challenge, especially since much of the brain is laced with blood vessels that the electrodes could easily puncture. Plus, due to their incredibly small size, the electrodes themselves are susceptible to damage.
The robot keeps a surgeon in charge, but it turns the process of electrode implantation into something closer to a video game. Using a microscope integrated into the robot, a surgeon is given a static view of the underlying brain, thanks to software that compensates for the pulsing and shifting. With the static view, implanting the electrodes becomes something like a point-and-click activity: the surgeon selects a location, and the robot inserts the electrode there while after compensating for any ensuing movement of the underlying tissue. Although video showed its insertion method as looking like a violent stab, the hardware protects the electrodes from damage at this point.
This method certainly has the potential to make electrode implantation safer, in part by minimizing the risk of blood-vessel damage. But let me be clear: while the electrodes are small enough that they're not dramatically larger than the neurons they interact with, there's still the potential for damage to those neurons or their support cells during the electrode insertion, as well as some disruption of the connections among neurons. That potential may be lowered by the robot, but it's not going away.
One other issue that the robot doesn't obviously solve is that several of the images displayed during the Neuralink introduction showed the chips being located somewhere other than where the electrodes were targeted. There's certainly enough play in the wiring of the electrodes to allow a bit of distance between the two, but it's hard to understand how this can be managed with a single, small surgical incision.
In existing systems, the electrodes are their own distinct hardware component, but Neuralink is looking to change this. The company hopes to do so by producing the metal portion of the electrodes as it's building layers of metal into the chips used for processing the electrode data. This provides some real advantages, as the process technology used there is already operating at the sort of fine scales that make structure of the electrodes easy.
This setup would also do away with any bulky connector hardware currently needed to link electrodes with the rest of the system—they're already part of it. Presumably, Neuralink will manufacture chips with electrodes of different lengths to allow for flexibility in the implantation process.
In use, multiple electrodes will be combined into a single “thread,” with polymer layers providing insulation to avoid cross-talk. Additional polymer layers will protect the thread from the environment of the brain, which Vanessa Tolosa of Neuralink described as “harsh.” The electrode and polymer materials were both chosen to limit inflammatory and other immune responses.
Overall, this part of Neuralink's approach seemed solid, although a full evaluation will have to wait for longer-term studies of a thread's safety and useful lifetime inside an actual brain. Scar development was a real problem with early electrodes made by others, but further development has limited this problems. Presumably, Neuralink has already learned from others here.
<em>Listing image by <a href="https://www.youtube.com/watch?v=r-vbh3t7WVI&feature=youtu.be">Neuralink</a></em> <div id="action_button_container"></div> <div itemprop="articleBody" class="article-content post-page"> <div class="gallery shortcode-gallery gallery-wide"> <ul> <li data-thumb="https://cdn.arstechnica.net/wp-content/uploads/2019/08/IMG_3385-150x150.jpg" data-src="https://cdn.arstechnica.net/wp-content/uploads/2019/08/IMG_3385.jpg" data-responsive="https://cdn.arstechnica.net/wp-content/uploads/2019/08/IMG_3385-980x698.jpg 1080, https://cdn.arstechnica.net/wp-content/uploads/2019/08/IMG_3385-1440x1025.jpg 2560" data-sub-html="#caption-1550389"> <figure style="height:698px;"> <div class="image" style="background-image:url('https://cdn.arstechnica.net/wp-content/uploads/2019/08/IMG_3385-980x698.jpg'); background-color:#000"></div> <figcaption id="caption-1550389"> <span class="icon caption-arrow icon-drop-indicator"></span> <div class="caption"> Multiple generations of implant chips were on display at Neuralink's launch. </div> <div class="credit"> <span class="icon icon-camera"></span> John Timmer </div> </figcaption> </figure> </li> <li data-thumb="https://cdn.arstechnica.net/wp-content/uploads/2019/08/Screen-Shot-2019-08-12-at-7.35.13-PM-150x150.png" data-src="https://cdn.arstechnica.net/wp-content/uploads/2019/08/Screen-Shot-2019-08-12-at-7.35.13-PM.png" data-responsive="https://cdn.arstechnica.net/wp-content/uploads/2019/08/Screen-Shot-2019-08-12-at-7.35.13-PM-980x551.png 1080, https://cdn.arstechnica.net/wp-content/uploads/2019/08/Screen-Shot-2019-08-12-at-7.35.13-PM-1440x810.png 2560" data-sub-html="#caption-1550583"> <figure style="height:698px;"> <div class="image" style="background-image:url('https://cdn.arstechnica.net/wp-content/uploads/2019/08/Screen-Shot-2019-08-12-at-7.35.13-PM-980x551.png'); background-color:#000"></div> <figcaption id="caption-1550583"> <span class="icon caption-arrow icon-drop-indicator"></span> <div class="caption"> Cleaner look at said chips direct from Neuralink's presentation. </div> <div class="credit"> <span class="icon icon-camera"></span> <a class="credit-link" href="https://www.youtube.com/watch?v=r-vbh3t7WVI&feature=youtu.be">Neuralink</a> </div> </figcaption> </figure> </li> </ul> </div>
This is perhaps the most intriguing part of the Neuralink effort. According to Neuralink's DJ Seo, the company is iterating chip designs at a three-month pace and has brought the size down by a factor of seven over multiple generations. The design Neuralink is currently testing has a number of components, but the most interesting of those are the components that take input from the electrodes.
Neural signals are both analog and extremely noisy. But in the systems we've looked at so far, information is carried by what are called “spikes.” These involve a dramatic change in voltage that quickly returns to background levels.
Neuralink's chip is designed to take the raw input from its electrodes, filter out some of the noise, and then identify the spikes. It then registers the spikes in an extremely compact data format (Seo said compression is about 200-fold) and sends it along to the controller sitting outside the skull. Right now, the chip uses a USB-C connection for data transfer, but it's intended to handle communication and receive power wirelessly. It can apparently get enough power that way to stimulate 64 electrodes at once, and it can switch targets to stimulate additional ones.
A lot of Neuralink's vision may sound difficult to believe, but the company's roadmap, in many ways, starts as an extension of existing work. We have surgically implanted electrodes in both humans and animal subjects, and we have successfully read neural activity. In some cases, we've even used those readings to perform tasks, like controlling a mouse cursor or even a robotic arm, as in the video below. But there appear to be a number of key advances in what Neuralink is working on, one obvious, the others less so.
Note the word choice: “appear” is needed because it wasn't always clear what's ready for use and what, in the words of Neuralink scientist Phil Sabes, is “aspirational.” (Some of the clearly aspirational ideas were well into the realm of science fiction; we'll come back to those in a bit.)
The obvious Neuralink advance is the size of the implant. As you could see in that video, existing electrodes are hooked up to a substantial box that protrudes from the skull, needed in part because the electrodes need adaptors before it can talk to any other hardware. Neuralink wants to get rid of all of that. It wants the surgery to be elective and outpatient, and it wants the recipients of its hardware to be able to go home with it and go about their lives as normal. In service of that goal, the company has gotten rid of most of this complexity, reducing the implant to electrodes directly integrated into a small chip on a circuit board with a USB-C connection.
The chip, and the fact that it is integrated with its electrodes, is the key to this advance. While there has been some hardware demonstrated in academic labs that allows mice and rats to move freely with implanted electrodes, that hardware's still quite a bit bulkier, and the animals are undoubtedly aware of its presence. If Neuralink gets its wireless connection working, the results would be somewhat better than the cutting edge present in research labs, and made in volume.
This is where the research community was heading, but Neuralink should get the community there with the consistency of hardware that's produced at scale. And the Neuralink scientists made it clear they want to cooperate with researchers when the hardware's ready. (When talking about using this in animal research, Musk joked, “We even care about rats, even though they carry the Black Death.”)
But Neuralink wants its hardware to be approved for use in humans, where it would represent a much more dramatic advance. Use in humans, however, will require extensive testing of everything. The safety of the electrodes and their lifespan within the brain environment, the surgical robot and implantation process, and the chip itself (which is meant to be implanted under the skull) will all need extensive validation. Assuming everything passes, however, Neuralink could dramatically lower the complexity of placing implants in humans, as well as the impact of having an implant on the patient's day-to-day existence.
What's less clearly an advance is the data coming out of this system. From the images shown during Neuralink's presentations, the raw data is relatively noisy, and it seems to have a lot of drift where the entire pattern of neural activity shifts up and down somewhat. I don't think that the noise is enough to keep the processor from identifying clear spikes in the activity of the neurons it's monitoring, but I'm less certain any borderline activity will stand out enough to register. This is something that may have to be sorted out through trial and error.
Revolution or science fiction?
Overall, if things work as they're expected to, the system currently in developed by Neuralink represents a significant advance over existing brain-computer interfaces. At over 1,000 electrodes in a compact format, there are plenty of existing use cases, both in humans and in animal research, that could benefit from something like a near-future iteration of this hardware.
But Musk and the Neuralink execs aren't interested in existing use cases. Their goal is to make this so simple to implant that it can be done on an outpatient basis. Implantees would be sent home with an app on their phone that monitors the device and lets them control it.
That's the sort of game-changing technology that Musk has been drawn to in the past, but the brain is a radically different engineering problem than rockets or cars. This is not something where rapid iteration and learning through failure are necessarily going to work. Musk seemed to have accepted this, acknowledging how hard the work will be and that it will take longer than he'd like. That's a good attitude to have, as some of the ideas floated during Neuralink's announcement are going to be extremely challenging.
To begin with, we have little experience with having electrodes like these reside deep in the brain for decades; we simply haven't been doing this sort of work for long enough for that to be possible. And, while the scarring caused by early attempts seems to have been sorted out, we really don't know if a set of electrodes will be able to consistently listen to and stimulate the same neurons 30 years after they're implanted. This issue is compounded by the fact that Neuralink is talking about implanting multiple devices—Musk said they think that as many as 10 would be safe, but Neuralink will probably limit things to “only” five.
Why would you need more than one implant? Because some of the things Musk is talking about will require talking with different parts of the brain. For example, he mentioned giving an artificial limb haptic feedback, an excellent goal. But controlling the limb would require a connection to the motor cortex, while the feedback would have to be fed in to the sensory processing system. These are in distinct areas of the brain and would therefore require separate sets of electrodes to communicate with.
There's no reason in principle this couldn't be done, but it hasn't yet been tried. In part, that's because any one implant creates the risk of damage to neurons, the connections among them, and the tissues that support them. Each additional implant would increase that risk. Obviously, the risk is typically low enough that implants are often considered a reasonable option for people, but nonetheless the risk definitely exists. Adding additional implants and targeting multiple brain regions will undoubtedly increase it.
As we mentioned in our earlier analysis of Neuralink, there's also the coding problem. For any of this to work, we have to understand how the information transmitted by neurons is encoded. This will vary among different regions of the brain; a set of signals traveling along the optical nerve won't mean the same thing as an identical set of signals flowing around the hippocampus as it tries to dredge up a memory.
Even when implants target the same brain region, there's going to be variation from implant to implant and patient to patient. That's because a given brain region contains a large population of neurons that often do distinct or partially overlapping things. For example, in the visual cortex, some groups of cells will register the presence of vertical features; others horizontal ones; still others pick up motion; and so on. Each set of electrodes will pick up signals from different subsets of these populations, meaning each individual's system will have to learn to understand the particular intricacies of the brain activity it's listening in on. Sending signals back will probably require learning by both the implant controller and the person implanted.
Aiming for science fiction can find (science) fact
None of these challenges are insurmountable. But they're all real, and they're what stands between an obvious extension of existing technology and Musk's mid-term vision. That vision is currently in the science fiction realm: multiple implants, put in place through elective, outpatient surgery, communicating wirelessly with a small bit of hardware behind the ear, and it's all controlled by a cellular phone. No individual part of that vision is too far beyond existing technology, but assuming that all of these individual challenges can be overcome (or at least overcome in a specific, convenient manner) is extraordinarily optimistic.
Obviously anything beyond that is clearly in the realm of science fiction, much like the whole idea motivating this: Musk's hope that getting humanity integrated into computers' home turf might make AI less hostile.
But that's not to say that this is going to be a wasted effort, even if none of these visions come to past. Based on the development of most technologies, the people at Neuralink will struggle to get their preferred solutions to all these challenges to work. Yet there's a reasonable chance that they'll get something to work and end up solving somewhat different problems in the process. And it's exactly because they've chosen to tackle an incredibly interesting set of challenges that whatever they do end up solving could be pretty consequential.
<div id="action_button_container"></div> </div>