Speaker unverified: belief

Transcript context

…So maybe can you tell me at high-level what the app is, what the software is outside of the brain? So maybe working backwards from the goal. The goal is to help someone with paralysis. In this case, Noland. Be able to navigate his computer independently. And we think the best way to do that, is to offer them the same tools that we have to navigate our software. Because we don’t want to have to rebuild an entire software ecosystem for the brain, at least not yet. Maybe someday you can imagine there’s UXs that are built natively for BCI, but in terms of what’s useful for people today, I think most people would prefer to be able to just control mouse and keyboard inputs, to all the applications that they want to use for their daily jobs, for communicating with their friends, et cetera. And so the job of the application is really to translate this wireless stream of brain data, coming off the implant, into control of the computer. And we do that by essentially building a mapping from brain activity to sort of the HID inputs, to the actual hardware. So HID is just the protocol for communicating like input device events, so for example, move mouse to this position or press this key down. And so that mapping is fundamentally what the app is responsible for. But there’s a lot of nuance of how that mapping works, and we spent a lot of time to try to get it right, and we’re still in the early stages of a long journey to figure out how to do that optimally. So one part of that process is decoding. So decoding is this process of taking the statistical patterns of brain data, that’s being channeled across this Bluetooth connection to the application. And turning it into, for example, a mouse movement. And that decoding step, you can think of it in a couple of different parts. So similar to any machine learning problem, there’s a training step, and there’s an [inaudible 05:32:39] step. The training step in our case is a very intricate behavioral process where the user has to imagine doing different actions. So for example, they’ll be presented a screen with a cursor on it, and they’ll be asked to push that cursor to the right. Then imagine pushing that cursor to the left, push it up, push it down. And we can basically build up a pattern or using any sort of modern ML method of mapping of given this brain data, and then imagine behavior, map one to the other. And then at test time you take that same pattern matching system. In our case it’s a deep neural network, and you run it and you take the live stream of brain data coming off their implant, you decode it by pattern matching to what you saw at calibration time, and you use that for a control of the computer. Now a couple sort of rabbit holes that I think are quite interesting. One of them has to do with how you build that best template matching system. Because there’s a variety of behavioral challenges and also debugging challenges when you’re working with someone who’s paralyzed. them has to do with how you build that best template matching system. Because there’s a variety of behavioral challenges and also debugging challenges when you’re working with someone who’s paralyzed. Because again, fundamentally you don’t observe what they’re trying to do, you can’t see them attempt to move their hand. And so you have to figure out a way to instruct the user to do something, and validate that they’re doing it correctly, such that then you can downstream, build with confidence, the mapping between the neural spikes and the intended action. And by doing the action correctly, what I really mean is, at this level of resolution of what neurons are doing. So if, in ideal world, you could get a signal of behavioral intent that is ground truth accurate at the scale of one millisecond resolution, then with high confidence, I could build a mapping from my neural spikes, to that behavioral intention. But the challenge is again, that you don’t observe what they’re actually doing. And so there’s a lot of nuance to how you build user experiences, that give you more than just a course on average correct representation of what the user’s intending to do. If you want to build the world’s best mouse, you really want it to be as responsive as possible. You want it to be able to do exactly what the user’s intending, at every step along the way, not just on average be correct, when you’re trying to move it from left to right. And building a behavioral calibration game, or our software experience, that gives you that level of resolution, is what we spend a lot of time working on.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.