Hacker News Reader

Comment by kevinskii | Hacker News Reader

I'd be happy to be corrected by an actual neuroscientist, but I don't think some folks realize just how far away the Neuralink team is from being able to do any of the things they are claiming. Inserting electrodes in the brain is the easy part. The reason no one else has yet bothered with this level of miniaturization is because it won't be of much benefit until you can figure out a way to decode the signals reliably. This is extremely daunting, and it's something that has occupied lots of brilliant people full time for decades.

View on HN

Well, in the update they showed themselves predicting the position of a pig's limbs as it walked on a treadmill.

If your bar is reading thoughts & dreams, you'll probably be right. But that's probably not what is needed to be useful in helping people with variety of ailments.

Not to take anything away from the Neuralink team, but predicting joint movements in a steady gait is one thing. Predicting join movements in starting, stopping, changing speed, and accurately manipulating objects is quite another.

Presumably they start with easy walking and build capabilities incrementally.

The model to push impossible tech has worked... cars and rockets.

That's kind of the thing though...those took 26 bil in subsidies to generate 20 bil in revenue to duplicate existing tech (spacex def gets credit for reusable, obvi, but Tesla man, yuck)

I’m not sure about the numbers, was thinking about the model to progress technology rapidly.

The reusable rocket tech seems worth it alone.

Not to take anything away from the Neuralink team, but predicting joint movements and angles is about the edge case of what can be achieved with semi-invasive and non-invasive bci.

Recording from the nerves predicting joint movements in a steady gait is definitely already achievable.

It can be predicted with the same precision/throughput, better than left/right:1/0 once per second?

Can you share some link?

Is your hypothesis that there is nothing they could do that would both be not too far from what they have now, plus valuable?

I believe that Neuralink has potential to do great things. However, like many other Musk ventures, they vastly exaggerate their own accomplishments while downplaying remaining challenges. Most brain-machine interface advances will probably be made elsewhere.

I think that's just how progress gets made quickly. As Larry Page would say, "have a healthy disrespect for the impossible."

For what they are trying to solve for initially, we’re talking about people that cannot even move a finger.

Not to burst your bubble but I don't think they're not predicting X,Y,Z limb positions in that demo. The traces on the left don't line up with the video, what would they actually represent?

It looks a lot more like the result of adding physical movement markers to a pig with an implant, and then mining the recorded neural data for signatures which correlate with the marker data. I'd be willing to wager that's exactly what is shown in that video.

As an actual neuroscientist (PhD student, in a past life, now a googler), I would disagree with this part: > Inserting electrodes in the brain is the easy part. The reason no one else has yet bothered with this level of miniaturization is because it won't be of much benefit until you can figure out a way to decode the signals reliably.

I performed and analyzed multielectrode recordings not entirely different from the ones used by Neuralink (Neuronexus is the brand if you want to look it up). These were pretty much the state of the art as far as electrode density in vivo goes. And we would absolutely have loved to have higher density. We even went so far as trying to simulate higher density by moving the electrode array around with a microdrive and repeating the stimulus.

I am quite sure the constraint is not on the demand for higher density arrays from scientists, it is more on the manufacturing side, the economics for further miniaturization are probably not there, in the absence of the Elon Musk cash spigot.

I fully agree that many of the claims Neuralink are making are quite a ways off though.

From some of the Q & A, it seems material design is going to be one of the biggest bottlenecks. Making the electrodes thin enough to give enough granularity, but also durable enough to provide a strong and reliable signal for many years.

Sounds like an engineering problem the can be solved with the right mix of experts, motivation driven by mission, and limitless funding by the likes of Musk to grease the wheels of innovation in this domain and problem space.

There are a half-dozen groups (companies & in academia) that are building neutral stim & recording platforms with thousands of channels & high bandwidth comms so that we can at least capture, analyse, & transmit signals to start learning & closing the feedback loop. Think of it like designing the first real integrated circuits so that we could eventually get the 8086.

This needn't be directly in the brain. E.g. there are a lot of groups looking at the parasympathetic nervous system (ie vagus nerve) for a host of organ control applications.

Been almost 5 years since I was working in the field... But there seems to have been a lot of progress already. Big limiting factor for iteration velocity is identifying viable test subjects -- elective participants aren't really an option (few volunteers, ethically questionable, & regulation); you have to find people with diseases or conditions that can justify such extreme interventions.

I’ve had parts of my body inactive due to peripheral nerve damage after an injury (peripheral nerves mostly always heal, so function came back), and I can tell you I would have easily opted into any experimental surgery that could restore function.

You don’t have to go far to find candidates, walk into any ER today and you’ll find trauma victims with irreversible nerve damage from a common car/motorcycle accident.

I agree with your assessment: it's easier to find subjects with peripheral nerve issues. It gets progressively harder to find subjects as you move up the chain, as the risks compound quickly: peripheral > central > brain.

As you can imagine, neural link is targeting some of the most invasive -- hence animal subjects.

There is a famous quote that we should all remember today " we overestimate what can be accomplished in the short term and underestimate what can be accomplished in the long term." - Arthur C Clarke

The demo in which they were able to read limb position was pretty impressive and shows that they should be relatively close to being able to help people that have lost limb control.

The brain is also quite adaptable and learns via feedback loops so getting the electrodes 'close enough' might be sufficient along with some physical therapy.

This has been there for a while, it is pretty simple to decipher impulses from the cortex. They are just publicizing years of research from other folks i.e. professor Jose M. Carmena from Berkeley and so on. We need more folks like Elon to publicize science...It is actually great work the team is doing.

It seems like the key advance here is the miniaturisation. Imagine one of these puppies paired up with an advanced prosthetic arm.

The key advance is the increase in order of magnitude and precision of placement of filaments embedded in the brain.

Combining the touch detection with other devices would already be useful on its own. A phone and neuralink could work together and figure out when I'm holding it and unlock automatically and then lock it when I stop holding it. The connection between them would have to be trusted and secured of course - it would need to verify the unlock signal actually came from my device.

Getting something like x,y,z coordinates of your limbs relative to the device would be a good primitive for many features (same usefulness as acceleration and orientation sensors in phones). If that was fast and accurate enough, that could be used to control robotics.

Pairing with other devices and issuing gesture commands to them would be another possibility. Sign language might be translate-able directly to bytes.

We're talking about building Pong at this point. There will be lots of creative solutions found within the constraints.

> A phone and neuralink could work together and figure out when I'm holding it and unlock automatically and then lock it when I stop holding it.

I mean, maybe, but I feel like you’re missing the potential of this device!

We're not :/ we just know our field

One of my favorite facts about neural interfaces... all electronics use negative charges for communications. But neurons use positive charge in the form of calcium flow. This makes it really hard to bridge that difference.

I wonder why is it so, in the end it's just electrons flowing from negative to positive end to create current, why do they behave differently inside neurons?

It’s because the flow of electrons driving current is mostly used only in one part of neurons: mitochondria. It’s used to generate ATP, which is the actual currency of energy in cells. This in turn powers systems that send and receive ions that are generally positively charged (why this is, I’m not sure) for neural communication.

So, while reading electrical signals is possible, returning them is tough because you’re so far removed from the language of the neurons when using electrons.

I actually really hope that is the case. I have a feeling of existential dread every time I imagine all the effort I am putting into learning and acquiring new skills all going to waste when those tasks become absolutely second nature to everyone, irrespective of effort or time investment.

Although, I am all for fixing the existing problems that it obviously could solve; I fear my mother will succumb to Alzheimer's, and something like this is dearly needed. It just seems like we won't know when to pull the brakes.

> I actually really hope that is the case. I have a feeling of existential dread every time I imagine all the effort I am putting into learning and acquiring new skills all going to waste when those tasks become absolutely second nature to everyone, irrespective of effort or time investment.

I'm not a neuroscientist, but with such a limited understanding of the brain, beyond just the anatomy and the nature of neurotransmitters, let alone consciousness, I think you can allay such fears for the time being.

I think this could have significant benefits for neural diseases like Alzheimers, Parkinsons, ALS and possibly anxiety and addiction/mania based maladies but the idea of simply being able to 'download' the ability to learn a new skill in real time seems entirely preposterous.

I say that because some of the skills that I have been able to turn into careers have actually required I learn something where my brain no longer interferes with the process so that it is done effectively.

The notion of muscle memory is real, meaning you cease thinking about it and it just occurs out without any effort due to continual practice and repetition: my question is how can your body recover from an unexpected/unpredictable error if you haven't already practiced that same scenario over and over, and you were only given the default normal operation of x skill? Most learning is adaptable and gained from observed behavior, be it our own or others, to mistakes.

Experience is a funny thing, and our brains are flawed in recalling many of our most cherished or dreaded events and experiences accurately; I honestly think this could be a medical boon for the aforementioned diseases, but the Matrix-like learning will probably be a quixotic pursuit, in my opinion. One that sounds awesome on paper but would be horrible in practice.

I agree with most of what you say; however.. if you consider some novel ability like speaking a foreign language, I would think that your aptitude in such a skill is a product of many factors, albeit memory and recall being paramount. If what Elon says about this device is true, and we can use it to mentally interface with the web, then I don't see what would stop us from speaking new languages "on-demand". This is a very naive and layman perspective but I believe I've heard Elon claiming ideas like this around the potential for this device. Not sure what the delta is between practical applications like disease treatment and this somewhat sci-fi scenario, but I hope it is large.

I don't feel existential dread, but I certainly do a feel a level of demotivation.

That being said, neuralink not being able to do this at the present moment is of no help to people who need to learn right now.

This is exactly how I feel about this. I hope this will help the people who are in need but I really don't think we'll know when to stop. I don't like the idea of this implant to be necessary to compete in the future. Granted it is still far in the future, but it is still kinda frightening knowing just like AI this could go out of our hands anytime.

Unfortunately, there is 0 chance that that this technology will be able to help with stuff like Alzheimer's (a disease that we don't even know the biology of - plaques are looking increasingly unlikely as the real mechanism) or any kind of higher brain function within our grandchildren's lifetimes. We don't even understand how computation happens in single-celled organisms, trying to jump to interpreting memory in humans is just so far off that it's like 10th century surgeons dreaming of re-attaching nerves.

However, there is probably hope for helping people who have lost control of their limbs through spinal chord injuries and similar effects within our lifetimes - this is a much more feasible problem that is already being worked on. Maybe even some fun gadgets for "thought"-based gameplay (simple movements).

everyone thinks it's pretty far away, i'm just glad someones taking a serious go at this

The people taking a serious go at trying to understand how the brain works are the ones trying to simulate a nematode brain, the ones working on understanding animal behaviors, the ones studying how infants acquire knowledge etc. These are all prerequisites for any kind of useful understanding that would allow us to build a working model of the brain and start imagining how to interface it more directly with other systems.

Not doing the fundamental research and trying to pick up useful signals by arbitrarily choosing to focus on inter-neuron communication in a few areas of the exterior of the brain is unlikely to even advance our understanding one iota. We're learning much more about thought (not neuro-motor transmission, mind you!) from the people training slime molds to navigate mazes than we are learning from this.

I'm glad though that there are more people working on perhaps helping people who have lost control of their bodies to regain it. That is an absolutely worthy goal that the team from Neuralink do have a hope of achieving in our lifetimes, and where I absolutely believe they can push the state of the art.

It's hard to build a model of a system without observing it - perhaps a reason we don't know how to decode neural signals yet is that we didn't have a way to read them. I'm sure that this vastly improved sensor will lead to better models of the brain and neural activity.

We have plenty of ways to observe neural activity, and have had them for years. This chip fab represents an incremental step. In fact if I had to choose for my lab this new chip or some other tech currently available, there are about 5-6 other things I'd prioritize over this Neuralink device.

Eg. what alternative?

1. Fiber photometry rig

2. A 1-/2-/multi-photon microscope with head fixation stage for GRIN lens imaging

3. New electrophysiology rig for single cell, paired cell, multi cell, voltage clamp, patch clamp, etc recording

4. An automated micromanipulator for molecular uncaging and optogenetics experiments

5. a rig for superres single molecule tracking experiments, PALM/STED/STORM/uPaint

6. A set of headmount Miniscopes (elon could actually help to vastly improve these)

Thanks.

Which of these could you use in vivo?

We’ve actually had ways of observing neural signals for quite some time. In order to identify the regions of the brain requiring surgery for epilepsy, patients get electrodes implanted and spend several weeks in hospital. Various researchers have recruited volunteers from this cohort for neural signal studies. A team recently managed to decode speech from neural signals from such a study!

https://www.ucsf.edu/news/2019/04/414296/synthetic-speech-ge...

Cool! This seems very similar to what Neuralink hopes to build out, aside from some technical details

That looks like a key point behind the existence of the Neuralink company.

I read we can decode signals reliably, but the hardware required is the size of a large desktop, so not something anyone could feasibly lug around. With that said, I think the use of GPUs is relatively new in the field, and I don't believe anyone is using ASICs, so there may be some significant gains possible.

I asked my neuroscientist friend about this a few months ago and his response was that inserting electrodes and interacting with the brain is the easy part: preventing the body from rejecting them or forming scar tissue over them after a few weeks is the hard part that nobody has been able to solve yet.

Reminds me of cyberbrain sclerosis from Ghost In The Shell.

Which is precisely what this has solved!

Ah, well that is exciting then (I did watch the video and know they said the pig had the unit for two months, but I wasn't sure if they had actually solved it or not, although two months is certainly a great sign).

Even among neuroscientists there will be disagreements about what is possible when.

Seems to me this is a circular problem, until somebody comes along and does the miniaturization and figures out all the software and systems you need, no serious progress will be made.

How about first trying to understand how the handful of neural cells in a jelly fish of worm can guide the animal to food and allow it to avoid danger, before trying to imagine we could help human beings recover their memory?

If you are so much better at solving these problem and your approach is so much superior go compete against them. If somebody is willing to dump 100M into making break-threw medical advances I'm not gone question their approach unless I am really, really sure I actually know better.

Well, unfortunately, unlike Elon Musk I didn't have rich parents, connected friends, or the luck to be part of a really great startup idea, so I don't really have 100M to spend on my own fantasies.

But the fact that someone is willing to put money into a fantasy will not make it a reality. As I said elsewhere, I am truly hopeful that much good will come out of Neuralink in the area of helping people with motor problems and prosthetics. I absolutely commend this goal and am happy to see more people invest in this area.

But that doesn't mean money will magically supplant the decades or more of research we still need to do to until we can meaningfully even think about what thought means at a physical level.

And just because Musk has delusions of grandeur about the way he will change the world, enough so as to put his money where his fantasies lie, is no reason to start believing in the same things.

Perhaps having more bandwidth would improve signal recognition and decoding?

I totally agree, I am totally unimpressed, Elon knows how to create hype and get smart people to do something. Smart people sometimes can't do it, it take a generation of research to access brain raw information. Codos to him to collect such smart minds but we are away 10-20 years from meaningful read and write to the brain.

> it take a generation of research to access brain raw information.

I don't see why this has to be the case. I'd more expect that the pace of research will accelerate as devices like Neuralink come onto the market and allow much higher resolution and more precise data to be collected across many more individuals.

10-20 years is reasonable for more advanced capabilities but they've figured out prediction of pigs' limb movements in ~1 year, so we could have neurally-controlled human prostheses very soon after trials begin.

i can imagine how seing the recent progress made by ML technologies in processing huge amount of information in real time to extract cognitive information, one could think the same kind of technics could be applied to other fields, provided you manage to reach the same amount of raw information.

I don’t see the point of this comment. Of course they are relatively far away. But they are also closer than anybody else.

No, not by a long shot. See, for example:

https://www.sciencedaily.com/releases/2019/06/190619142542.h...

Do they have a robot surgeon that installs it though?

It's non-invasive. No need. (Maybe you were joking.)

It’s an interesting article, but I’m not sure why you would imply this orthogonal approach is further along than Neuralink.

The novel research being done in the paper you linked isn’t even around the sensor suite, it’s around training AI to extract more precise signal from a noisy sensor — in this case an EEG.

The premise of the article is that EEG is being used because direct neural monitoring is too invasive and difficult.

Neuralink is essentially attempting to disprove the premise of that paper.

I’m going to guess the signal processing Neuralink will be doing is much more sophisticated than whatever that paper was doing on top of an EEG feed, just because the data rate from Neuralink is so much higher.

Disclaimer: I truly dislike this technology for my own personal reasons.

That aside, I think the problem you bring up is pretty much "solved" as the only thing that is missing are training sets. Pretty much like when Tesla just needed a bunch of driving footage to start improving its self-driving AI, and they got it.

The first Neuralinks (I believe) will be intended to gather as much data as possible in order to tackle this concrete issue. After that is going to be a quite vanilla "big data + ML" type of job.

It would be cool to see, if different people have different neuronal dynamics for the same sort of "activity". Who knows what it will be found. Regarding "the pig demo", we don't know for sure if they can do that to any pig wearing a neuralink (i.e. like installing software) or if the data/model was produced and meant to work on that particular pig only.

Agreed, ML will be able to extract signals from this data, no question. The electrodes really are the limiting factor I think, because you're going to need millions placed all over the brain and deep inside too for a truly useful implant.

I don’t think that’s entirely true because brains are quite adaptable. In similar sorts of studies, humans managed to train themselves to use the electrodes to direct things. It may never be ‘perfect’, but I think you’d be surprised how much could be achieved using such a method.

>because you're going to need millions

For sure, increasing # of electrodes + bandwidth must be on the roadmap.

You may not need to go as deep and broad as you believe. Most processing happens on the cortex anyway, the interior (white matter) is pretty much interconnection. Also, the action of one neuron could trigger thousands of others in quite remote regions, so you could probably get a decent picture of what's happening without having to cover it all.

I was honestly quite impressed by "the pig demo", I wouldn't have guessed that would be possible with the current state of their tech.

> That aside, I think the problem you bring up is pretty much "solved" as the only thing that is missing are training sets. Pretty much like when Tesla just needed a bunch of driving footage to start improving its self-driving AI, and they got it.

That is absolutely not the case, and we are absurdly far from anything like it. We haven't even been able to decode very much useful information from the "brain" of a nematode yet, and there we don't care in the slightest about its health.

Self-driving is actually an interesting example (though it is likely many orders of magnitude simpler than interfacing with the higher functions of the neural system of, say, an insect). We are still much farther away from real self-driving with the skills of an average human driver (such as not hurtling into static obstacles on a road) than anyone was thinking a few years ago. In fact, the CEO of Waymo believes that we won't achieve full, all-condition self-driving in our lifetimes (though he does believe that we will achieve useful self-driving in common conditions).

Regarding the brain, we don't even know yet where the computation takes place yet - to what extent does it happen at the cellular level, and to what extent and the neural network level? Further, the electrical signals are unlikely to be the only important part - there are many chemical substances that impact our judgement and reasoning, so I don't see why we would assume that even capturing all of the electrical activity of the entire neural network would be enough to decode thought patterns.

>That is absolutely not the case, and we are absurdly far from anything like it. We haven't even been able to decode very much useful information from the "brain" of a nematode yet

Huh? What?! Have you actually watched the update?

They decode the position and movement of the limbs of a pig as it is walking, in real-time, with uncanny accuracy.

Yes, motor signals. This is not necessarily entirely new, we have been able to take simple readings of motor signals for years, this is 'just' pushing the frontier (not to say that it isn't a great accomplishment!).

Thought processes though are a completely different beast, and we have no realistic hope of interpreting them at this time. We wouldn't even know what to look for. We're missing much more fundamental research in how thought works in even simple animals before we could have the first hope of finding something in a brain. Even thought processes such as 'that direction is more promising for food'.

Oh, I see your point. Yeah, we are not even sure how consciousness emerges. But, for sure, some sort of crude i/o with someone's thoughts could be established, and not far away.

I'm not even talking about consciousness, just basic computation in neural networks. How does the very simplest animal decide whether to move left or right in the presence of a stimulus? Can we influence that thought, not at the sensory/motor level, but at the computational level?

These are unanswered problems in worms. They are unanswered even in single-celled organisms, or at least in colonies of single-celled organisms, such as slime mold.

So I don't believe that any kind of IO will be achievable in the next hundred years, beyond possibly interpreting motor functions, and perhaps providing simple sensory input.

Reply on news.ycombinator.com