Like most neurotech folks, I’m thrilled that people with clout are finally taking deep brain-machine interfaces (BMI) seriously. Not only does it mitigate a plausible threat to humanity — the chance that artificial intelligence will outpace and dominate us — it offers enormous value in the short and long term, the chance to expand what humanity really is.
But…BMI is tough.
As in, really tough. Not just electric-cars tough, or spaceflight tough. Accomplishing deep BMI — where deep is a term I’m using to mean a true, high-bandwidth link capable of expanding human cognition — will be a decades-long project. Decades of engineering, decades of biotech, rigorous systems work, marketing, testing, commercialization, and hundreds of overlapping science projects to fill in the many gaps.
In other words, a perfect example of what I like to call hard tech, or the ultimate level of science-based business.
Thanks to recent developments, there’s now a serious public conversation about deep BMI. Which means there’s enough noise already floating around that I should take a moment to say: this is my specialty. At NeuroPace, we made and commercialized the first and still the only permanently implantable, two-way brain interface. An embedded system we designed is currently running inside more than a thousand people’s heads and more are being implanted every day. So, I’m going to take the liberty of laying out a path toward deep BMI, based on the principles of successful science-based business. I hope this contributes to the public discussion, and I invite the other engineers who have tackled these things or have brought real BMI through a development cycle to chime in as well.
1. Innovation logjam
We’ll start with the good news. There’s enormous reason for optimism. Why? Because we currently have a backlog of twenty-plus years of neurotech innovation just waiting to be translated to clinical practice. 2017’s state-of-the-art implantable brain electrode is a chunk of platinum-iridium with a surface area of about 6 mm². Guess what was 1997’s state-of-the-art implantable brain electrode? You got it — a chunk of platinum-iridium with a surface area of about 6 mm².
There’s an amazing portfolio of neural interface tech being developed and tested in labs all over the world, including light-based signalling, ways to direct neural growth, and some micro-scale interfaces that have even seen use in humans. But none of these have yet become products for anything but lab use. This is in the nature of how medical-device projects get planned and funded. Industry is needed to make real products. Without clinical data showing that a new interface with an unknown risk-benefit profile performs better than existing tech, industry players are quite reasonably unwilling to assume the business risks of a large-scale project. But, these clinical data are unobtainable without great effort and expense, and the existing tech actually works very well in established fields like treatment of Parkinson’s disease.
Our firm belief — knowledge, even — that new neural interfaces will offer enormous commercial and human value isn’t enough to unblock the logjam and pull this innovation into real products. What’s needed is money, commitment, and truly long-range planning. The time has come.
2. Constantly build the business
A project as ambitious as deep BMI will encounter near-endless challenges, and no investor has infinitely deep pockets. Saving humanity is a brilliant goal but history has shown that whether you’re building a cathedral or trying to ship ten kilos of carbon nanotubes from Austin to Boston… the best motivator is cash.
So it’s safe to say that at some point a deep BMI project needs to make money, and the sooner the better. By now this is a firm tenet of the startup world. It’s not enough to build a great product. You have to be able to sell it and build some kind of viable business, whatever that means for your venture. This still isn’t the worldview of most scientists, but for those involved in planning a deep BMI development program, it has to be.
What does this mean? Every step of the project should be planned with an eye toward ensuring business as well as scientific success. Now, this means something very different for a well-funded, decades-long development program than it does for a two-person software startup. It’s perfectly reasonable, and probably necessary, to invest in diversified projects that may only pay returns fifty years from now. But this should be balanced by identification of shorter-term wins that might advance the field while transforming the project into a self-sustaining enterprise. This may begin on the medical-device side — spinal restoration or the like — and may advance into consumer products or enterprise-level human performance tools as the technology matures. Needless to say, this needs to be accompanied by constant, smart marketing and a truly long-range PR effort to ensure that the world is ready for what we build.
3. Go deep
I’ll say it: let’s discard any illusions that scalp-based tech is a path to real BMI. I’m an EEG guy, and my company makes neurostimulation headsets for athletes. I get it. Those signals are so tempting, and it’s amazing what can be learned from things like evoked potentials. But the language of the brain is not summed electric fields (“local field potentials”). The language of the brain is that of the individual activity of millions of neurons all doing what they do, and for deep BMI we need to speak to the brain in its own language. For this effort, EEG and its cousins fNIRS and MEG are a distraction, or at most a useful scientific tool for study of deeper methods. Even fMRI (and other flavors of MRI requiring larger and larger magnets to achieve finer and finer resolution) will eventually become irrelevant if the goal is permanent human augmentation. There’s an interesting possibility that interventions like tDCS or drugs can help the brain adapt to use novel interface technology, but again, it’s an adjunct, not the core path forward.
4. Start with the spine
There are plenty of things that make deep BMI tough, but one that stands out is just how intractable the information architecture of the brain is. In other words, we still don’t know how thoughts and percepts are encoded, and there’s a valid philosophical question of whether it’s actually possible for us to understand this.
There’s a good argument that we can eventually bootstrap around this problem — similar to how we might create powerful AI without fully understanding intelligence, by building systems with the right underlying rules that allow complex phenomena to emerge. But in the real world here, we have to be able to build effective systems to explore, de-risk, and productize BMI one step at a time. And working without a clear information architecture — a clear communications protocol, as it were — is hell for engineering.
When I’m debugging an embedded system, like a brain implant (it’s still fun to say that after 16+ years in this field), facial ganglion stimulator, or neuromodulation headset, the first thing I do is separate the problem into modules, and control inputs and outputs one-by-one. Trying to build and troubleshoot a deep BMI without the ability to intimately understand inputs and outputs will be like trying to fix a computer with no knowledge of how the microprocessor talks to the memory. We’ll undoubtedly find ourselves in this position, but it’s no way to get started and no way to map out the first principles of robustly growing artificial elements into and among neurons.
So, the first milestone on the path to deep BMI should be a prosthesis for treatment of spinal injury. The information architecture and neuroanatomy of the spine is as well-understood as anything in the nervous system, and there is a long existing history of surgical access to the spine. In fact, the first and still largest neurotech market is spinal cord stimulators for pain — using platinum-iridium electrodes with a few mm² area, of course.
A truly effective spinal neuroprosthesis solves a critical clinical problem. Nail this, and it’s a scientific win, an engineering win, a PR win (you’re curing paralysis and letting people walk again!), and a business win for the project. As the decades pass, keep driving forward, even including a consumer or augmentation product, and it will be an important basic part of the BMI tech portfolio.
That’s not to say it’s easy. There’s a huge gap between where we stand now and making this real. But it’s a tractable gap, and some of the lower-hanging fruit in neurotech. And, it’s work that de-risks the larger project in several essential ways. Simply building a product based on a next-generation neural interface — one that involves surgery, but incorporates a growing or self-replicating electrode design as a first step toward neural lace — will reveal all the problems that today we don’t even know are problems. If we start now and take this seriously, today’s paraplegics could be walking within fifteen years.
5. Diversify the science portfolio
While we’re captivating the world with an amazing spinal prosthesis, we’ll also be pressing on the longer-term science projects that lead to a true brain interface. The problem here is that there are huge gaps yet to be filled. We don’t know how we’ll safely and effectively grow interface fibers into the cortex. We don’t know how those fibers will speak to neurons. We don’t know a hundred other things that we don’t know we don’t know. For any given approach, the chance of success is low.
So the rule here is: diversify, support, analyze. Like venture investors and sports agents, a deep BMI program manager needs to pick a wide variety of early-stage prospects, and pick them so that the portfolio is diversified and each has a chance of contributing to the cause. Support and nurture each scientific approach with money, time, and strategic guidance. Then critically analyze the portfolio, pruning the tree to make sure resources are well-applied at every step.
This is familiar ground for scientists. It’s the same strategy NIH or DARPA uses to allocate resources between risky projects. It’s also familiar to entrepreneurs and investors — it’s how early-stage VCs diversify their portfolios.
But here on the path to deep BMI, nothing is easy. In a decades-long project with a tough endgame and real products at stake, each science project needs a plausible path to engineering success — while allowing for the intuitive leaps and pivots that move science forward in unexpected, yet hugely valuable ways. It will be uniquely hard to maintain this scientific portfolio, and I don’t have a great answer other than: diversify, support, and analyze. Over and over, with intellectual rigor and clear-eyed evaluation. Until all the gaps are filled and what remains is a boring old engineering and systems problem.
6. Move to visual and sensorimotor
After spine, what next — how do we learn to interface with cortex? Once more, we can keep the training wheels on in a zone with more tractable anatomy and information architecture. For the brain itself, these are the visual and sensorimotor cortices.
Neural interfaces in the visual cortex have some interesting background. Primary visual cortex (V1), located in the occipital lobe at the rear of the head, processes input from the eyes. The information is already highly processed by each step in the chain — from retina through the lateral geniculate nuclei (LGN), encoding things like edges and movements — but retains meaningful spatial organization and a tractable encoding scheme even when it reaches the cortex. It’s a favored target for both fringe and mainstream, although early, work in cortical interfaces. And why not? A successful visual prosthesis means the blind can see (or represents the truly immersive consumer VR experience of 2040).
As a side note, there are good arguments for moving earlier in the processing chain — the information architecture for vision is even simpler in the thalamus and optic nerve, and retinal implants are already a maturing technology. A LGN visual implant using early neural lace technology should be well within our grasp in a few decades, and offers another useful intermediate step toward deep BMI.
We may choose to make a sensorimotor BMI instead. Primary sensory cortex (S1) and primary motor cortex (M1), approximately located where the band of a pair of headphones would sit, are the other favored targets for early cortical-interface work. Even a decade ago, it was possible to give a monkey robust control of an artificial arm using an indwelling electrode array; the unsolved problems here are around how to interface pervasively and permanently with the cortex, including stimulation of sensory regions as well as output from motor cortex. Solve this one and you’ve effectively treated ALS, bypassed all spinal injuries, and created the final link in immersive, augmented input/output, while taking the last preliminary steps toward deep BMI.
7. Leverage plasticity
Our brains love to learn and optimize — in a word, neuroplasticity. This means one shortcut is constantly relevant, and it will save us multiple times along the path to deep BMI. In the same way an infant learns to use her body, or you internalize a new sport after constant practice — if the right framework of feedback and reward is provided, the brain will learn to use a BMI, even if we don’t understand exactly what’s going on under the hood.
In a classic experiment more than a century ago, psychologist George Stratton showed that the brain can learn to seamlessly reverse an upside-down visual input in less than a week of training. More recently, after only 70 hours of practice, blind people can learn to interpret audio information similar to echolocation as visual input — activating previously-unused visual cortex.
With such power available, BMI developers should constantly try to leverage plasticity, providing predictable responses and rapid feedback to allow the brain to learn. As the field progresses, drugs and/or electromagnetic stimulation might accelerate a period of controlled learning where patients, subjects, or consumers learn to treat a new interface as an extension of their bodies.
8. Risk and quality management
There are well-established principles for analyzing and managing risk in an uncertain technical environment. There are well-established principles for maintaining impeccable quality even in a vastly complicated project.
Use them! Start with a careful risk analysis, constantly updated and revisited. Treat it intelligently, accepting and gradually reducing the vast uncertainty that will doubtless be a fact of life here. And when it comes time to engineer and produce, build a solid foundation of medical-grade quality systems.
There’s nothing more sensitive than the brain, from every possible viewpoint. One safety disaster won’t be a mere disaster — it will potentially kill the entire deep BMI program.
9. Execute on cognitive
Finally, the main event. As I’ve said a few times now, we don’t know what we don’t know. But after we:
unblock the logjam of neural interface innovation…
commit to commercial viability…
hone our scalpels and craft our nanobots…
let paraplegics walk…
manage a vast portfolio of science projects…
let the blind see…
and avoid disaster…
…we can get there. With a plan like this, some good luck, and a crazy amount of work, we can execute on the greatest biological engineering project of our generation.
Again, I hope this contributes to public discussion, and I invite discussion from other engineers and neuroscientists. Please reach out in the comments or ping me at @wingeier. Looking forward to decades of fascinating work, and deep BMI under my Christmas tree in 2050.