I agree that this seems farfetched. At the same time, I don't think we need to understand how the brain works in order to mimic or copy it, right? At a somewhat absurd level, we give birth to humans with brains without understanding brains.
The goal here seems to be to start by copying it without understanding in order to gain understanding. AFAIK we already simulate brains of silk worms or some such thing?
Yeah, agreed. I think it's a bit like working with a legacy system, where you might begin by testing and measuring its interactions with external interfaces, and then can start to gradually replace pieces with new components.
> AFAIK we already simulate brains of silk worms or some such thing?
There is research into mapping the complete neural structure of a nematode worm called C. elegans [0] - one of the simplest known organisms that has some kind of a nervous system (the entire worm has 959 cells for hermaphrodite individuals, or 1035 for males; of these, 302 are neurons - and yes, this is not a typo, these are not hundreds or thousands of cells, just cells - that's how tiny it is). Its entire nervous system has been mapped out, including all connections, in [0]. Note that its neural cells are not like mammal neurons, they lack many of the known processes present in even a single mammal neuron.
However, the project to simulate this simplest of organisms, OpenWorm [1], has not yet achieved its goals (which is also somewhat reduced - they only intend to simulate the motor system initially, not the entire organism).
Silkworms (which are moths) are waaaay beyond any imaginable simulation capability at the moment.
Right, I was just bringing that up as an example of simulating without understanding.
The paper is talking about taking another approach. Whether it will work or not is not something I could comment on.
Another example is how we have a very imperfect model for the universe, and yet we use our models to make strikingly accurate predictions. Huge gaps in our knowledge exist, but that doesn't stop us from using approximations to do interesting things.
> Right, I was just bringing that up as an example of simulating without understanding.
But that's just it - first they spent some time understanding how the system works,painstakingly mapping out each connection, the role of each individual cell (literally), they studied the chemistry inside each neural cell etc. The simulation effort started after all of this was done.
If anything, I would say OpenWorm is a counter-example to the idea that you can simulate such a system without understanding it.
Similarly, with our model of the universe, while gaps exist, the extent of those gaps and approximations is mostly known, and the areas affected by those gaps are exactly where we can't make predictions (e.g. the internal structure of a black hole). Not to mention, we had a massive piece of good luck there: it turns out that particles are described by linear equations, which have extremely nice properties for approximations. Already with GR we are in much worse territory with it's nonlinear equations, but at least we understand these well enough to create some linear approximations - or so we think.
Simulating without understanding isn't actually a thing. You neeed to know what to simulate and what to ignore.
For example: do we need to simulate all the DNA transcription to make a virtual neuron? Currently this is unknown. And if it is required it's pretty much game over already.
I don't believe you need that level of fidelity but that's more of a hope than a scientific standpoint.
> You neeed to know what to simulate and what to ignore.
And often the fastest and most economical way to find out is to try doing it.
Also simulations need not be perfect to be useful. I spent a large fraction of my life writing object oriented simulations of transformers. We simplified everything (reduced detail, reduced number of dimensions, rules of thumb where there was no analytical formula) but still succeeded in designing better transformers than before.
The next generation will be more precise and even more useful but you can't always get there in one jump.
Biology (and, indeed, physics) contains fractal levels of complexity. The existence of debate does not mean that we don't understand it well enough to mimic it.
At one time, people had no idea how the brain worked at all. Now we know that the brain is a collection of neurons, special cells which have filaments called an axon and dendrites that extend from it to other neurons at junctions we call synapses. At these synapses, electrochemical signals are transmitted to and from other neurons. When a neuron receives certain signals, it may transmit that signal on to other neurons. That's basically the Anatomy 101 chapter on the nervous system, and has been pretty well established for about a century.
It's true that we're constantly learning new things about diseases that affect the function of the nervous system. There remains some debate on how exactly neurons arrange themselves, repair themselves, and change. We're still investigating internal computations within the dendrites to see how those work to decide when to propagate a signal or not. The signals are analog in some ways, but mostly digital 'all or nothing' responses, there's time coding, different types of discharge patterns...fractal complexity.
If you set the goalposts at a comprehensive understanding of every bit of trivial minutiae that exists in the human body, yeah, we are (and probably will always be) far off. If you just want to predict the function of a nematode worm with 302 neurons, observing that if you engage this particular set of sensory neurons, then these few hundred synapses are activated with this pattern of impulses, and soon those other motor neurons are stimulated and the worm curls in that direction...that seems entirely within the realm of possibility. In my opinion, the goalposts ought to be "A better understanding than what we have today". We'll reach them by tomorrow, and move them forward again!
So, I keep reading responses such as this on the subject of neuron/brain emulation, and I have a question to ask back. Is it possible to understand a neuron at all? It appears that we have some understanding, given that we can decode some of the electrical and chemical signals coming from them and to them, but it seems that the zeitgeist is that these things are inherently impossible to understand. If so, as a layperson why waste the resources trying to advance this branch of science? (Note I personally fully support advancing this branch of science.) Why not spend the effort on something with positive tangible results, like space travel or clean power?
I'd say a single neuron should be understandable given that the network effects DO NOT have sufficient impact.
Given that "sufficient impact", then the "system understanding itself" issue arises.
However, if you believe the neuron is the same as the neuron in the worm, then the concept of how it works should (by my best guess and nothing else) be adequately simple for human understanding.
> Why not spend the effort on something with positive tangible results, like space travel or clean power?
Effort/interest is not really fungible or spendable like this. People work and study on the things that are interesting to them. We can't just flip a switch to get everyone who's researching brains to transition to researching space travel.
I do not think that they are. However, whenever this topic comes up, there are people who immediately jump out and state that is basically is, simply because we don't know enough about the underlying science yet. Which is strange, as studying the underlying science to understand the situation is exactly what we should be doing... It's just so backwards, and it confuses me.
There's a fair bit that is well understood, specifically the chemical processes that drive the signals (and that are present in many other types of cells), and there are mathematical models that have been tested in-vivo and reasonably replicate some dynamics (for example the Hodgkin-Huxley model). But neurons are indeed complicated, and there are a lot of them...
I mean you could theoretically copy some really simple animal brains (10k neurons or so) to see how it interacts with the real world in an actual 'in silico' experiment.
That's based on C. elegans which I think that only has 902 neurons. Also I don't think they've ever cloned the state of a living worm, only the connectome. That itself is pretty cool but it seems like what Samsung is talking about would go pretty far beyond that.
Even simpler at 302 for adults, their nervous system is completely mapped out and I'm still incredibly impressed nonetheless by how much technology has progressed in my lifetime to make something like that possible
Since worms usually survive electric shocks (and people usually survive seizures), the connectome is probably enough; there probably isn't anything vital in volatile memory. Less-volatile memory like DNA methylation has been proposed to be a significant memory mechanism at the per-cell level but so far, as I understand it, there is not good evidence for this.
Simulating 10k neurons is absurdly difficult. The only success in this area - and it is far from complete - is simulating C. elegans, a microscopic worm that has 959/1035 cells in total (1035 for males, 959 for hermaphrodites). Of these, 302/385 are simple neural cells (much simpler than a mammalian neuron) [0], and these along with the 95 muscle cells have been simulated mostly in their entirety - mostly, but not completely.
Even so, with 302 neural cells there are ~10k synapses to be simulated (9065/13,758 based on sex). The current status is a simulation taking 5-10 minutes of real time to simulate 15ms of worm time, on a somewhat regular PC[1]. Extrapolate that to 10k neurons...
Good sci-fi read on this is "Reamde" by Neal Stephenson.
Certainly any brain is just a biological machine. No magic.
But it's workings - even for simple brains - is still way beyond human understanding.
I remember an episode of Cosmos mentioned how we used to have gods for every little thing like rivers and trees, then as they were unserstood the remaining gods grew fewer and more abstract.
> I remember an episode of Cosmos mentioned how we used to have gods for every little thing like rivers and trees, then as they were unserstood the remaining gods grew fewer and more abstract.
That's a very convenient, attractive model, but not really even remotely accurate as a description of the evolution of religious belief.
What actually happened in a lot of the world is that people who called the “gods of every little thing” something else (“guardian angels”, “patron saints”) imposed their religious language on other people by force (or just exterminated them outright).
I will posit that there is intelligent life somewhere in this vast universe that does understand the brain (being so vast, there are places with similar brains), and therefore for some civilization, it's not magic ;)
There are certainly varying degrees of baseline consciousness.
I think you can pull many people out of their patterns into consciousness for a short time through the right engagement, but they'll just slip back into semi-oblivion when they go back to their daily life, social network, confirmation bias affirming news sources, etc...
If Joe met Jane and they had a baby together, everyone (except sibling commenter yosito, if the baby behaved irrationally during a time of political upheaval and global pandemics, and excepting maybe David Chalmers) would agree that Joe Junior would have consciousness.
If we cloned Joe, and grew him from an infant, then aside from Jane's genetics, he'd be much the same as Joe Jr. I think it's pretty undeniable that Joe the Second would have consciousness.
If we copied Joe's entire body down to the neuron - working molecule by molecule, 3D-printing him in a vat, and then defibrillated his heart into beating with a lightning strike, would you say that Dr. Joe's Monster had consciousness? I think the answer is probably yes. The human brain, like the heart, may require some initial pattern of activity - something stored in DRAM instead of FLASH, if you will - a self-perpetuating pattern of electrochemical action potentials that's not represented by the connectome of neurons; if you need to kick-start that in the same way you initialize the beating of the heart then that's simply part of the process of faithfully copying Joe. There may be some fast-moving chemical reactions and electrical potentials to duplicate, but they're measurable and physical; copy them and I think you'll have a consciousness.
The question is whether a high-fidelity but non-biological simulation would have consciousness. I think Joe 2.0 might be in a rough state, really testing the chip's adaptability functions, if his sensory neurons were not appropriately stimulated as they're accustomed to or if his motor neurons were unable to produce the motions that he's accustomed to. But those seem like small tasks compared to faithfully simulating a network of hundreds of trillions of synapses. If you can do that, then yes, I believe your chip would have consciousness.
I think that the cutoff where you cease to have a consciousness is when, rather than simulating Joe's connectome, you predict what Joe would have done. Consider the following conversation:
>> Hi, Predicted Joe, what's your name?
> My name is Eliza^W Joe.
That predicted Joe, I would argue, does not have consciousness, no matter the accuracy of the prediction.
I keep hearing this phrase "we don't know much about the brain" from people who've never opened a neuroscience textbook.
And at the same time everyone and their grandma has an idea how to build artificial intelligence.
I'm not sure exactly of what they mean by neuromorphic (paywall...), possibly something related to the Hodgkin and Huxley dynamical model. NTT recently released a paper on something similar (at least in name), here's a bit more details on their side:
52 comments
[ 4.8 ms ] story [ 118 ms ] threadFrom what I've understood, we are so far from understanding even how a single neuron works. There's still plenty of debate about that.
The goal here seems to be to start by copying it without understanding in order to gain understanding. AFAIK we already simulate brains of silk worms or some such thing?
I'm pretty sure OpenWorm has not achieved its goal.
My take on this would be, sure, copying it is an important step, but they don't know to copy it, so it won't work.
There is research into mapping the complete neural structure of a nematode worm called C. elegans [0] - one of the simplest known organisms that has some kind of a nervous system (the entire worm has 959 cells for hermaphrodite individuals, or 1035 for males; of these, 302 are neurons - and yes, this is not a typo, these are not hundreds or thousands of cells, just cells - that's how tiny it is). Its entire nervous system has been mapped out, including all connections, in [0]. Note that its neural cells are not like mammal neurons, they lack many of the known processes present in even a single mammal neuron.
However, the project to simulate this simplest of organisms, OpenWorm [1], has not yet achieved its goals (which is also somewhat reduced - they only intend to simulate the motor system initially, not the entire organism).
Silkworms (which are moths) are waaaay beyond any imaginable simulation capability at the moment.
[0] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6889226/
[1] https://github.com/openworm/OpenWorm/milestones
The paper is talking about taking another approach. Whether it will work or not is not something I could comment on.
Another example is how we have a very imperfect model for the universe, and yet we use our models to make strikingly accurate predictions. Huge gaps in our knowledge exist, but that doesn't stop us from using approximations to do interesting things.
But that's just it - first they spent some time understanding how the system works,painstakingly mapping out each connection, the role of each individual cell (literally), they studied the chemistry inside each neural cell etc. The simulation effort started after all of this was done.
If anything, I would say OpenWorm is a counter-example to the idea that you can simulate such a system without understanding it.
Similarly, with our model of the universe, while gaps exist, the extent of those gaps and approximations is mostly known, and the areas affected by those gaps are exactly where we can't make predictions (e.g. the internal structure of a black hole). Not to mention, we had a massive piece of good luck there: it turns out that particles are described by linear equations, which have extremely nice properties for approximations. Already with GR we are in much worse territory with it's nonlinear equations, but at least we understand these well enough to create some linear approximations - or so we think.
For example: do we need to simulate all the DNA transcription to make a virtual neuron? Currently this is unknown. And if it is required it's pretty much game over already.
I don't believe you need that level of fidelity but that's more of a hope than a scientific standpoint.
And often the fastest and most economical way to find out is to try doing it.
Also simulations need not be perfect to be useful. I spent a large fraction of my life writing object oriented simulations of transformers. We simplified everything (reduced detail, reduced number of dimensions, rules of thumb where there was no analytical formula) but still succeeded in designing better transformers than before.
The next generation will be more precise and even more useful but you can't always get there in one jump.
At one time, people had no idea how the brain worked at all. Now we know that the brain is a collection of neurons, special cells which have filaments called an axon and dendrites that extend from it to other neurons at junctions we call synapses. At these synapses, electrochemical signals are transmitted to and from other neurons. When a neuron receives certain signals, it may transmit that signal on to other neurons. That's basically the Anatomy 101 chapter on the nervous system, and has been pretty well established for about a century.
It's true that we're constantly learning new things about diseases that affect the function of the nervous system. There remains some debate on how exactly neurons arrange themselves, repair themselves, and change. We're still investigating internal computations within the dendrites to see how those work to decide when to propagate a signal or not. The signals are analog in some ways, but mostly digital 'all or nothing' responses, there's time coding, different types of discharge patterns...fractal complexity.
If you set the goalposts at a comprehensive understanding of every bit of trivial minutiae that exists in the human body, yeah, we are (and probably will always be) far off. If you just want to predict the function of a nematode worm with 302 neurons, observing that if you engage this particular set of sensory neurons, then these few hundred synapses are activated with this pattern of impulses, and soon those other motor neurons are stimulated and the worm curls in that direction...that seems entirely within the realm of possibility. In my opinion, the goalposts ought to be "A better understanding than what we have today". We'll reach them by tomorrow, and move them forward again!
Given that "sufficient impact", then the "system understanding itself" issue arises.
However, if you believe the neuron is the same as the neuron in the worm, then the concept of how it works should (by my best guess and nothing else) be adequately simple for human understanding.
Effort/interest is not really fungible or spendable like this. People work and study on the things that are interesting to them. We can't just flip a switch to get everyone who's researching brains to transition to researching space travel.
https://youtu.be/1wj9nJZKlDk
https://www.embopress.org/doi/full/10.1002/embj.201387637 seems to be an overview of the situation from a few years ago.
Even so, with 302 neural cells there are ~10k synapses to be simulated (9065/13,758 based on sex). The current status is a simulation taking 5-10 minutes of real time to simulate 15ms of worm time, on a somewhat regular PC[1]. Extrapolate that to 10k neurons...
[0] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6889226/
[1] https://github.com/openworm/OpenWorm
That's a very convenient, attractive model, but not really even remotely accurate as a description of the evolution of religious belief.
What actually happened in a lot of the world is that people who called the “gods of every little thing” something else (“guardian angels”, “patron saints”) imposed their religious language on other people by force (or just exterminated them outright).
It can instill terror in philosophers, computer scientists, and moviegoers all at the same time.
I think you can pull many people out of their patterns into consciousness for a short time through the right engagement, but they'll just slip back into semi-oblivion when they go back to their daily life, social network, confirmation bias affirming news sources, etc...
If Joe met Jane and they had a baby together, everyone (except sibling commenter yosito, if the baby behaved irrationally during a time of political upheaval and global pandemics, and excepting maybe David Chalmers) would agree that Joe Junior would have consciousness.
If we cloned Joe, and grew him from an infant, then aside from Jane's genetics, he'd be much the same as Joe Jr. I think it's pretty undeniable that Joe the Second would have consciousness.
If we copied Joe's entire body down to the neuron - working molecule by molecule, 3D-printing him in a vat, and then defibrillated his heart into beating with a lightning strike, would you say that Dr. Joe's Monster had consciousness? I think the answer is probably yes. The human brain, like the heart, may require some initial pattern of activity - something stored in DRAM instead of FLASH, if you will - a self-perpetuating pattern of electrochemical action potentials that's not represented by the connectome of neurons; if you need to kick-start that in the same way you initialize the beating of the heart then that's simply part of the process of faithfully copying Joe. There may be some fast-moving chemical reactions and electrical potentials to duplicate, but they're measurable and physical; copy them and I think you'll have a consciousness.
The question is whether a high-fidelity but non-biological simulation would have consciousness. I think Joe 2.0 might be in a rough state, really testing the chip's adaptability functions, if his sensory neurons were not appropriately stimulated as they're accustomed to or if his motor neurons were unable to produce the motions that he's accustomed to. But those seem like small tasks compared to faithfully simulating a network of hundreds of trillions of synapses. If you can do that, then yes, I believe your chip would have consciousness.
I think that the cutoff where you cease to have a consciousness is when, rather than simulating Joe's connectome, you predict what Joe would have done. Consider the following conversation:
>> Hi, Predicted Joe, what's your name?
> My name is Eliza^W Joe.
That predicted Joe, I would argue, does not have consciousness, no matter the accuracy of the prediction.
I keep hearing this phrase "we don't know much about the brain" from people who've never opened a neuroscience textbook. And at the same time everyone and their grandma has an idea how to build artificial intelligence.
https://group.ntt/en/newsrelease/2021/04/23/210423a.html
Could we ever track the transistor processes and determine the screensaver from first principles? (Can’t cheat knowing it’s W98 for example)
"Could a Neuroscientist Understand a Microprocessor?" by Eric Jonas and Konrad Paul Kording