Neuralink White Paper [2019]

[Elon Musk & Neuralink's "An Integrated Brain-Machine Interface Platform With Thousands Of Channels" inspires this post.]

What is Neuralink's primary goal, and what have they done to achieve it?

Neuralink's primary goal is to make a high-bandwidth, high-resolution brain computer interface that is also highly scalable and suitable for chronic implantation. While each of the four components of this goal -- bandwidth, resolution, scaling, longevity -- is difficult on its own, Neuralink has made significant progress on each one over the last few years. This post will summarize what Neuralink has achieved in each category, according to their published paper (note: apparently even more has happened than just what they've publicly announced).

Bandwidth

The very high electrode density in Neuralink's device requires a very high bandwidth transmission interface. Even the USB-C interface (which can carry up to 40Gb/s) discussed in the paper, isn't fast enough to transfer all of the raw neural data.

To solve this problem, Neuralink implemented an ASIC-based "recording stack." The ASIC does initial data processing, which consists of "amplifying small neural signals, rejecting out-of-band noise, sampling and digitizing the amplified signals, and streaming out the results for real-time processing." (pg. 5) The USB-C cable then carries the processed data out of the brain and into a "base station," which uses a 10Gb Ethernet connection and to send data via UDP to downstream consumers.

Up to 3 implants can sync with one base station (useful for when one person has multiple implants for increased density), and examples of a downstream consumer of the data include long-term storage and data visualization (pg. 6).

Neuralink mentioned they planned to use wireless transmission in their first product (not USB-C, as described in the paper), but bandwidth details about this system were not provided.

Resolution

Resolution over a large number of neurons is a key goal of Neuralink's project. While current prosthetic control BCIs and speech synthesizer BCIs use relatively few electrodes (<256), there are close to 100 billion neurons in the brain and transferring "high fidelity information" (like thoughts or much finer motor control) will be nearly impossible without much higher resolution (pg. 1).

Neuralink's solution to the resolution problem is "threads," long, flexible strips embedded with electrodes. The paper lists two types of threads Neuralink is currently using, called "Linear Edge" and "Tree", named after their shape. Linear Edge is one long thread with 32 electrodes all on one side, whereas Tree places each of its 32 electrodes on its own "mini-thread", 16 of which branch to one side of the main trunk and 16 to the other side. Electrodes are placed close together in each probe (50um in the former, 75um in the latter), which enables high density recording.

Because the individual electrodes are very small, a surface modification is used to "lower impedance...and increase the effective charge-carrying capacity [of the electrodes]" (pg. 3). The two coatings trialed thus far are polystyrene sulfonate (PEDOT:PSS) and iridium oxide (IrOx). While the former coating does have lower impedance, it's less clear whether PEDOT:PSS is as suitable for chronic implantation in humans as IrOx.

The threads' flexibility also increases resolution. Because threads aren't confined to a specific area, as in a traditional electrode array like a Utah Array, they can be spread across a wide area and reach more and deeper neurons.

Scaling

Neuralink's vision of wide-spread adoption of BMIs requires new methods and considerations beyond what is typically necessary for small-batch, highly specific medical BCI implants. BCI implantation surgery cannot be long and involved like it currently is, and devices must be mass-manufacturable.

Neuralink's solution to the insertion scaling problem is an advanced robotic system that can insert electrode threads quickly and automatically, with minimal human oversight (long-term, the goal is with no oversight). This robot has three key features: an imaging stack, an automatic insert mode, and mechanisms to ensure minimal down-time.

The imaging stack has software to locate thread loops and the target brain area, and then build a depth map of the area for insertion. This software is assisted by the imaging stack's ability to emit several wavelengths of light, each of which hooks into one of the software components (see supplementary video #1 for a demo of these different light sources during insertion). The imaging stack provides the technology for the automatic insertion mode, which can reach insert speeds of up to "6 threads (192 electrodes) per minute" (pg. 4). Further supporting this speed is the sub-minute time necessary to swap the needle pincher cartridge and the ultrasonic cleaning method used for the needle, as they significantly decrease any maintenance or cleaning downtime.

Neuralink has also laid the groundwork for large-scale manufacturing. For electronics, they use a modular approach, packaging amplifiers, analog-to-digital converters, and digital output serializers together on one ASIC, which is then integrated with the rest of the required technology on a PCB. Separating manufacturing processes enables individual optimizations and avoids the difficulties of trying to optimize one monolithic manufacturing process. Theoretically, this should increase yields and lower costs. For large-scale electrode manufacturing, Neuralink developed a "wafer-level micro fabrication process" of the arrays that contain the electrode thread and built a separate "sensor" area that interfaces with the ASICs (pg. 2).

Longevity

The final piece of the puzzle is ensuring that the mass-manufactured, high-resolution, high-bandwidth device can last in the hostile environment of the brain. To accomplish this, parts and processes must be biocompatible.

Biocompatible materials are materials that won't harm the tissue they're embedded in. The electrode threads use a lightweight, flexible, biocompatible polymer called polyimide, which is also heat and chemical resistant. The electronics, on the other hand, move away from polymer usage and towards metal usage, employing a titanium case coated with parylene-c. The coating provides extra protection against potentially catastrophic fluid ingress.

Because brain tissue is extremely important to a person's functioning, minimizing damage from surgical tools during the thread insertion procedure is critical. In addition to health concerns, inflammation and scarring of tissue torn or displaced during surgery significantly lessens signal quality and thus implant effectiveness, which is the whole point of the surgery. As a result, the robot is programmed to plan thread insertion paths that avoid vasculature in the brain, and threads are designed for maximal thinness and flexibility to reduce tissue displacement during insertion.

Conclusion

There is still more progress to be made towards Neuralink's final design goal of a wireless interface that connects to an iPhone app, about which details were not given in the paper. While wireless systems typically use more power and have lower transmission rates than wired ones of similar complexity, the achievements Neuralink has already made in thread technology, scalable insertion, and materials are all immediately applicable to the wireless design and should greatly accelerate development.

Sources

1. An Integrated Brain-Machine Interface Platform With Thousands of Channels: https://www.biorxiv.org/content/10.1101/703801v4