The Brain is Not Computable
technologyreview.com
technologyreview.com
This is a fringe opinion, and I really wish the title reflected that. Ignoring the absurd 'linear' part of the article, I don't believe predictability is important to the brain. While it's possible quantum effects could explain unpredictability in the physical universe, there is no scientific evidence this is relevant to the brain; the brain operates at a much higher, more macro level than quantum mechanics. Besides, randomness can be introduced into silicon if it's that important. I hope the article is misrepresenting his opinion, but they seem dangerously close to 'the brain is more complex than I can comprehend, so it must be magic'.
I and many cognitive and neuroscientists I've spoken to consider this whole line of reasoning to be anti-scientific philosophy (although I may be biased because I studied AI, which rests on the idea that silicon can recreate intelligence).
http://arxiv.org/abs/quant-ph/9907009
This paper posits the brain can be modeled as a classical (e.g. Newtonian, "billiard ball"-style) physical system
The conventional neural network model neglects the interior of the neuron. Gene regulatory networks for complex eukaryotes are on the order of neural networks in complexity and involve quantum-scale interactions, which opens the possibility of quantum effects being significant. Gene regulation within the neuron affects neural firing behavior and, more importantly, profoundly affects neural growth patterns and thus learning and longer term forms of cognition.
This also neglects the possibility (now considered probable) that more cell types than just neurons are involved in brain activity:
http://en.wikipedia.org/wiki/Gliotransmitter
In short: the brain is not a neural network. Rather, those mathematical connectionist models are just that: models of aspects of the brain. We do not yet know to what extent these other mechanisms play a role, and what their role is. Given their nature it seems in both cases that their role might be more long-term, affecting long duration learning, planning, etc.
It really seems to me as if the most ardent and enthusiastic adherents of the Kurzweilian vision are computer scientists who don't really respect the domain of biology and like to hand-wave away its complexity as "background noise." You can't do that. I say this as a lifelong computer programmer who has studied biology. Studying biology really blew away any notions I had of simple, classical computer programs becoming movie-style AI.
The author is not making an anti-scientific "magic" argument. He is simply pointing out that biological systems are analog, embodied, electrochemical (and thus physical and possibly quantum), nonlinear complex systems, and he is being skeptical about the idea that such a system is going to yield readily to digital computer simulation. I agree with his skepticism.
Prediction: brain simulations will simulate superficial brain behavior but they will not become sentient. More specific prediction: they will get stuck in closed cycle loops. They will not exhibit the higher order motivation, creativity, or learning behavior seen in brains, which is probably because these behaviors emerge from all the real embodied biophysical stuff the CS people are ignoring.
Here's a pretty decent article:
http://www.nature.com/news/2011/110615/pdf/474272a.pdf
My point is that a neuron is not an equation in a connection model. It's a cell. You can't hand-wave away its identity as such or all the things that happen in cells.
Even if nothing quantum is involved (and I didn't say it was... we don't know), including significant aspects of intra-cellular activity and including other cell types (glia, etc.) and other types of interactions adds exponentially to the classical computation requirements.
I didn't mean that I dismissed the idea of classical computers simulating the brain or the implications. I just meant that I'm skeptical. Even if it is possible I'm very skeptical of it happening soon due to the absolutely insane computational requirements. My intuition is that it would require a leap on the order of vacuum tube ENIAC -> Intel Core i7. Meanwhile consumer demand for faster and faster chips is giving way to demand for slower but more energy efficient chips for mobile devices, which is subtracting from the economic incentive to continue Moore's Law. (AMD just bowed out of the x86 race for instance, leaving us with an Intel monopoly at the high end. Monopolies get lazy and stagnate.)
Personally I'm much more optimistic about "wet" transhumanism-- life extension, augmentation, brain hacking, etc., than I am about "real" AI in the near term or mind uploading. I also see computers stagnating a little for a while the way aerospace did from the Apollo era until about now... entirely for economic rather than physical reasons.
There's a huge philosophical difference between the idea that there is something going on in a brain that can't be computed (what the article suggests, and what you seem to suggest here) and the idea that a brain is hard to simulate (which is what you suggest in the next paragraph). As the posters above you have pointed out, the former is a fringe opinion unsupported by current evidence. The latter is a skeptical viewpoint to which many neuroscientists subscribe. Personally, I'm in the "I want to believe" camp.
> My intuition is that it would require a leap on the order of vacuum tube ENIAC -> Intel Core i7. Meanwhile consumer demand for faster and faster chips is giving way to demand for slower but more energy efficient chips for mobile devices, which is subtracting from the economic incentive to continue Moore's Law.
To a first approximation, speed, power consumption, and performance per dollar are all linear functions of transistor size. If you can make smaller transistors, you can make a more power-efficient chip with the same performance, or you can make a more powerful chip with the same energy consumption. I don't think people will stop innovating on this front until we hit a physical wall. There may only be one major desktop CPU maker, but there are plenty of competing fabrication plants. Since the brain is embarrassingly parallel, we only need to get within a few orders of magnitude of simulating a brain on a single chip, at which point we can just string a bunch of processors together.
> Personally I'm much more optimistic about "wet" transhumanism-- life extension, augmentation, brain hacking, etc., than I am about "real" AI in the near term or mind uploading.
If by "near term" you mean within the next 20 years or so, I agree. Beyond that, it's really hard to tell. If faster computers actually let us make scientific discoveries faster, the singularity people may well be right.
1) Be sufficiently large that quantum effects are irrelevant, like a billiard ball, and thus it can be modeled deterministically inside of a computer. The Tegmark paper proposes this is how the brain works.
2) Rely on quantum mechanical effects (which are non-deterministic, but can still be simulated inside a computer if you trust a computer's opinion of "random"). Penrose, Searle, etc. would argue otherwise, that there's a method to the madness, and that there's something "special" about the quantum mechanical effect on microtubules. We still don't understand quantum mechanics very well, so if human consciousness really relies on quantum mechanics, there is arguably a bit of wiggle room here, especially if you're a physicist of the same caliber as Penrose.
Microtubules are specifically of interest because they're one of the few brain structures small enough to arguably be subject to quantum mechanical events.
Tegmark argues that they're not small enough and that it's irrelevant.
The real question is:
Is it possible to simulate a brain within a computer?
- not -
Is it really hard? Is it too hard for modern technology? Is this problem simply too complex for us to understand? Do we lack the technology to build sufficient understanding today? Etc. Etc.
This article is claiming it's IMPOSSIBLE to simulate a brain within a computer. Proving that is a tall order.
This misunderstands the role of quantum theory in macroscopic systems. There is never a scale so large that quantum effects can be safely ignored. All that happens is that the specific effects, and probabilities, change.
The Heisenberg Uncertainly principle doesn't build a wall between the microscopic and macroscopic realms, it makes a probabilistic prediction about quantum events at all scales -- larger scale, lower probability.
But a macroscopic system that has a very low probability of exhibiting classic quantum behaviors as a whole, will nevertheless show quantum behaviors at some level. An easily understood example is a radioactive sample -- let's say a kilogram of uranium. The sample isn't going to behave like Schrodinger's cat, but its constituent atoms certainly will.
In one sense, the uranium is a classical mass with no contribution from quantum theory. In another sense, it's highly influenced by quantum theory -- were this not so, there would be no nuclear disintegrations.
An observer can examine the sample and, intent on demonstrating that it's a classical system, use the half-life equation to predict its future --
a' = a 2^-t/f
a = activity level at time zero
a' = activity level at time t
t = time, consistent units
f = half-life factor
-- And the outcome looks very classical, but only because the sample is large. But the timing of the next disintegration is quantum-deterministic. So the uranium is a chimera -- part classical, part quantum. This is an example that makes the point very clearly, but all classical systems (and scales) possess quantum properties, usually not so obvious as it is with a radioactive sample.
> Rely on quantum mechanical effects (which are non-deterministic, but can still be simulated inside a computer if you trust a computer's opinion of "random").
Quantum effects aren't merely random. What connects an event to quantum theory isn't its randomness, but the nature of the randomness -- its genesis. No matter how carefully I design a random number generator, I won't be able to imitate quantum entanglement unless the system is actually capable of this specific physical behavior.
I think what I was getting at was this: I'd take the Kurzweil crowd more seriously if I saw them really engaging seriously with the depth and breadth of biology... if I saw them talking about the sorts of things I was talking about seriously instead of hand-waving them away.
I get the impression (I'm knowledgeable in biology BTW, and a programmer) that they are making the frequent programmer mistake of failing to have respect for the problem domain they are entering. A similar error often occurs in creative product development. In this case you have a software effort to subsume/encapsulate an informal industrial process that is approximately five billion years old. There are lots of things going on that we don't understand yet, and that will not yield to a cavalier, dismissive approach.
The brain evolved in the physical world. It seems an odd economy to categorically exclude quantum effects because they are "too small", hard to measure, and/or poorly understood.
I didn't realize anyone had worked out the level at which quantum mechanics decoheres to Newtonian mechanics.
Well, yes, of course -- the Heisenberg Uncertainty principle is the border between them:
http://en.wikipedia.org/wiki/Uncertainty_principle
Equation: 𝝈x 𝝈p >= ℏ/2
I emphasize the "border" isn't a like a solid barrier, it's more a statement about probability.
Also, everyone should realize that the relationship between human brain function and QM is very speculative. It might be true, it might just be someone's pet idea. There's no real evidence for or against a role for QM in brain function, and a number of persuasive suggestions that it's not involved.
I would credit the argument only if they had an association, not solely because they're both unexplained. Otherwise we could assume a connection between Bigfoot and the Loch Ness Monster, on the dubious ground that they're both asserted to exist but with no evidence. :)
> I was thinking about Schrodinger's cat.
It's important to say that the Schrodinger's cat experiment is a useful thought experiment, but it can't actually be carried out -- too many ways for the system to decohere.
True, but it describes a theoretical property of quantum theory. Its value is only in showing the weirdness of the quantum world.
The problem is that many people read about it and conclude that it's a practical experiment.
I do work in molecular dynamics. To even simulate a million atoms requires huge approximations. You can get more accurate as you simulate less. If you want an almost perfect match with reality, simulation will get you about 2-3 helium atoms. Now consider how many atoms are in a human brain.
So it's hard for me to imagine fully simulating a human brain, although I don't see why it is theoretically impossible. Brains behave according to the same laws of physics as everything else in the universe.
On our current technological improvement path, I don't see a brain simulation occurring any time soon. If quantum computers were developed, it would make things much easier, but we would still need a new "kind" of technology. I wouldn't rule it out completely though. Who in the 1600s would have predicted microprocessors?
As for his talk about souls or consciousness, that just confuses me (and I'm religious too). Everything that we have thus far discovered obeys the laws of physics, so ruling out a simulation via some mystical "property" that human brains have seems sketchy to me.
Now, if you want to talk about things that really aren't computable, I'll direct you to Chaitin's constant: http://en.wikipedia.org/wiki/Chaitins_constant
Edit: I'll elaborate a little bit more. We currently simulate large proteins using force fields like CHARMM or AMBER. The problem we're trying to solve is what structure these proteins will fold into, and these force-fields work pretty well for that.
But consider this: these potentials are basically a handful of equations that describe stretching, bending, torsional, van der Waals, and electrostatic interactions. The parameters for these equations come from measurements of simple compounds that have similar structure, and these are used to extrapolate what will happen in a different substance. Good enough for folding, but if you want accurate energy levels? No way.
But saying the problem can't be solved ever seems dead wrong to me.
We humans excel at understanding and replicating what nature did - then at improving it.
Once we clearly understand how memories are stored, if they can be read and written that'll be half of the problem : accessing the data.
If the theory about memory being encoded in the microtubules is right, imagine some nanomachines that could read it from a "live" human - by broadcasting radio waves, or emitting photons (we started doing that for proteins with antibodies glued to radioactive markets, then we improved and glued them luciferase, now we do multiple colors and IIRC it's being experimented for DNA), whatever.
Now imagine other nanomachines that could rearrange the microtubules to match that - voila, you've got Matrix-style "uploading" of knowledge once we understand how the memory bits interact with eachother, how they can be accessed by the subject. Maybe it's like a SQL database fk/pk - we don't know. But something must exist to allow it. When we figure it out, there is no reason why it couldn't be done too.
My own predictions : after we confirm how memories are stored, if we have nanotechnologies to create nanomachines, we will start reading memories just like we did with DNA and proteins.
It will take a while, we will only have a read-only access at first- and with many bugs just like how introns and TATA boxes could be mysterious initially - but we will understand in the end, and that will be half of the problem solved.
Downloading will require additional advances in computer technology (at least faster cpus, in 3d instead of 2d to get more interconnections and raw computing power, and maybe some integration of processing and memory to match how neurons work), but it does not seem far-fetched to me.
Whether or not all physical processes are computable in the sense that they can be simulated by a turing machine is an open question, although my impression is that most folk who care to express an opinion think that they can. There's a little bit more on the Wikipedia article on the Church-Turing-Deutsch principle. http://en.wikipedia.org/wiki/Church%E2%80%93Turing%E2%80%93D...
Also, it would help creating silicon equivalent. Imagine we have something just as small (or smaller) than a biological neuron, which can interface and work in the very same way.
Computability or not, if it works in exactly the same way, I'd guess it could work the same on a higher level too - like, in a brain, especially if we have figured how to extract the stored information and if we can feed it back. (unless there are emergent properties we missed in the first place, but then if they can be identified, maybe they can be replicated too?)
That would allow an iterative trial/error (ex: if it doesn't work - why?) which might not resolve the computability question, but bring even more interesting issues about why it might not possible - something we can learn only from an experimental approach.
I like this blog's http://michaelnielsen.org/blog/interesting-problems-the-chur... take on it:
"Yet our ease with the CTD Principle is an ease brought by familiarity. One hundred years ago the statement would have been far more surprising, and, I suggest, even shocking to many people.
Viewed from the right angle, the CTD Principle still is shocking. All we have to do is look at it anew. How odd that there is a single physical system – albeit, an idealized system, with unbounded memory – which can be used to simulate any other system in the Universe!"
You say "once we figure how information is stored and processed", but how do you know that what is going on in the brain is just a matter of storing and processing information in a way that can be mimicked by a turing machine?
What I mean by this is that maybe we can't efficiently simulate a 3D human brain on any 2D computer, because there are vertical connections that are nearly instant in 3D that must map to a very very long route in a 2D "layers" simulation.
My hunch is we will efficiently simulate human brains once (1) we have finished the Human Connectome Project (2) we know much more systems biology about what happens inside both neurons and glia and (3) we have solved the heat-transfer and interconnect and yield issues in building 3D computer chips
Do we?
We have food optimized for tastiness. (I must say I'm waiting for food optimized for health benefits, but still we did improve it)
So yes we do.
Replace "cells" with "molecules" and "consciousness" with "fluid dynamics", and you can see what a vague, hand-waving argument this is.
>“You can’t predict whether the stock market will go up or down because you can’t compute it,” he says. “You could have all the computer chips ever in the world and you won’t create a consciousness.”
You can't predict the precise behavior of an analog amplifier, either, but you can still model it and produce a digital equivalent.
>“You can’t predict whether the stock market will go up or down because you can’t compute it,”
If you had an accurate model of the agents, you could easily compute the stock market.
However, that doesn't mean we cannot run a model. Exactly what the model's output 'means'... well, that's a different question. To use his examples, our simulations of weather or the stock market do not produce the same output as the future. But their outputs (hopefully) represent actually realizable states of the world.
In other words, as long as our model gives us 'human enough' output, then I guess it's sufficient? I mean, it really comes down to 'why do you want to simulate the human brain'. If you want to be able to upload your brain, then that probably isn't good enough. But I can imagine for various other uses, it could be enough.
I do think Kurzweil is at best... wildly optimistic though.
>"But Nicolelis is in a camp that thinks that human consciousness (and if you believe in it, the soul) simply can’t be replicated in silicon.
I'm guessing/hoping that the silicon reference was made by the author of the article and not Miguel Nicolelis, considering that silicon is extremely likely to be replaced by graphite or something else in the next decade or two, at least with almost full certainty by the next century. If it actually was Nicolelis who spoke about silicon, it automatically discredits him from having anything to say about the distant future of computing.
>That’s because its most important features are the result of unpredictable, non-linear interactions amongst billions of cells, Nicolelis says."
Even if that was true, which I doubt, so what? There has to be some kind of a system behind those "unpredictable, non-linear interactions" in order for the brain to have any functionality at all, and every system can be figured out and simulated. It might be incredibly complex and take centuries for us to gain the required knowledge and processing power, but even that doesn't make it impossible.
1) If we have to simulate it at a low level, the human brain is far too complex for any computer in any timeframe of our current lives to have enough power to simulate properly
2) Working backwards and simulating the high level processes (AI, etc) have been a dismal failure at actually replicating human thought processes, and will continue to be. While NN or the like can theoretically simulate any algorithm, we have no idea how to effectively train them in a way that produces high level thought similar to a human brain.
Generally when discussing this with people, I say this: if you disagree, give me a date by which you think I will be shown wrong, and then we'll reevaluate at that point.
I fully expect I could be proven wrong, and that would be an awesome world to live in, but my bold and unfortunate prediction is that I won't be.
You don't. The article argues that there is a theoretical barrier that prevents a brain emulation in principle, you argue that technology isn't ready yet and won't be in our lifetimes.
Opponents of your viewpoint argue that you simply can't imagine the state of technology in 50 years.
> That’s because its most important features are the result of unpredictable, non-linear interactions amongst billions of cells, Nicolelis says.
If he thought it was theoretically impossible, then the "billions of cells" would be redundant. It would only take 1 un-computable cell. Without a longer interview, we can't be sure what exactly he means.
My interpretation was "unpredictable, non-linear" (i.e. not computable by a simple algorithm, would have to be very complex, because of non linear interactions between inputs) amongst "billions of cells" = an obscene amount of computational data. I don't think he means unpredictable to mean strictly uncomputable.
> Opponents of your viewpoint argue that you simply can't imagine the state of technology in 50 years.
Yes, but at the same time there are things we thought we would be able to do 50 years ago that there is no way we can do now. There are physical constraints to the universe, and we can't just assume "technology" will overcome all of them. Nobody can imagine the state of technology in 50 years accurately, but I am still willing (and have done) to take bets on this 50 years into the future.
The concept that we're going to hit some moore's law style thing in science that will propel us to just automatically understand things, which we can barely measure currently, just doesn't line up. Just the process to understand how a single thing functions on a single channel seems to take forever now, and most neuroscience labs aren't limited by the speed of their desktops...
If we consider 1958 to be a good starting point for self-driving cars (http://technologizer.com/2010/10/09/google-self-driving-cars...), might we say that Google's car is orders of magnitude more powerful than those of 1986?
If we consider 1954 a starting point for machine translation (https://en.wikipedia.org/wiki/History_of_machine_translation), again might we find that Google Translate is similarly in a different league than SYSTRAN of the 1980s?
I do not really have the knowledge and sources to back those claims up, so I have to frame them as opinions for now. Could someone with experiences or deeper knowledge about these areas weigh in?
If you're doing animal-based research (the majority of real neuroscience currently), then its time spent with behavior testing, surgeries, waiting for the drug to be in an animal for 72 (or however many) hours, processing slides, pipetting, etc.
The time spent at a computer is mostly data analysis, reading papers, ordering supplies and grantwriting. A huge amount of time seems to be spent jumping through hoops, ordering things, working with vendors of equipment that doesn't frequently work as advertised, and dealing with broken stuff overall. The data-analysis they are doing again, isn't bound by the computer's speed. It generally is working with a few dozen (or hundred) samples of relatively computationally easy data.
There's not much for a computer to speed up.
But if so, I don't think it'll be with a standard von Neumann machine. Not that such a computer couldn't perform the required computations... it's Turing complete. But I think it would be a very poor fit for the problem domain. You'd want some kind of radically different incredibly highly parallel architecture. You also might want it to be analog or analog-like. There's been some interesting renewed interest in analog computers for a little while, and in probabilistic processors that can run incredibly fast by discarding the requirement of perfection.
Apart from that, there is this thing called ethnobiology, a sub-dicscipline of anthropology, that studies the way civilizations understand and represent the living things.
Ethnobiology reveals another constant in History : we tend to compare our brain to the most complex technology we know.
At the Renaissance, philosopher assimilated the brain to a very complex and subtle clockwork, Freud compared it to a steam engine, which pressure should be evacuated to avoid explosion. In the 40's, schoolboy and schoolgirls were told that brain was like a telephone exchange. Today, computers are the most advanced technology we know, so we tend to compare our brain to it. But like our predecessors, it's very likely that we are wrong.
Let just think forward, and admit that we are totally biased by the fact that computer are now inherent part of our life. Let's admit that there is a chance that our brain may never be modeled by a computer.
PS: for those who read french, a part of the above is largely inspired by a talk of Ted CHIANG, available here : http://www.actusf.com/spip/article-9802.html (sorry I can't find an English version)
> But the greater lesson lies in the vitalists' reverence for the elan vital, their eagerness to pronounce it a mystery beyond all science. Meeting the great dragon Unknown, the vitalists did not draw their swords to do battle, but bowed their necks in submission. They took pride in their ignorance, made biology into a sacred mystery, and thereby became loath to relinquish their ignorance when evidence came knocking.
> The Secret of Life was infinitely beyond the reach of science! Not just a little beyond, mind you, but infinitely beyond!
http://lesswrong.com/lw/iu/mysterious_answers_to_mysterious_...
Unless we're talking about souls here or something
Never say never.
The existing universe is only mostly described by simple mathematical laws. There are still lots of gaps in our understanding, just getting all the matter position and states probably wouldn't be good enough since we don't know the exact program yet to carry the simulation forward. However, that is a technical rather than theoretical limitation.
Besides which the Singularity has nothing to do with the 'soul' and consciousness. It's about super intelligence, this is possible without being self aware. IE Deep Blue I assume is not considered 'conscious' but it can solve the chess problem better than us, why would you think a super intelligent machine that can solve general problems has to also be 'self aware', bizarre.
I say: Is too! Dialogue complete.