Simulating 1 second of real brain activity takes 40 minutes and 83K processors
gigaom.com
gigaom.com
My research is in molecular dynamics. I simulate systems up to 1 billion atoms on Kraken & Titan. HUGE approximations are made in simulating these systems, but depending on what exactly it is you're studying, these simulations still provide useful results. That is the key to all these studies: how well does your approximation reproduce whatever it is that you're attempting to model? In some cases, very well. For instance, I'm not going to get exact energy levels of a large system, but the system will qualitatively evolve in the same fashion that the experimental system does (which then guides the experimental counterpart to the research). I don't know the details of this brain simulation, but there is certainly some aspect of it that is not being reproduced anywhere close to real-life, and hopefully this isn't what they're interested in (and I'm sure they know that, but I don't think the article author does).
The very best simulations of reality that we can perform handle at most just a few H/He/C atoms. And an issue called the fermion-sign problem means that the computational power necessary to simulate larger systems scales exponentially with the number of particles. Unfortunately what that means is -- short of developing quantum computers (which are polynomial order for the sign problem) -- we aren't ever going to simulate more than a few atoms with near-perfect accuracy, and certainly nothing like a human brain.
EDIT: Didn't mean for my comment to sound so negative. Obviously, the researchers know exactly what they're doing. I was just trying to dispel the impression that we're close to simulating the human brain.
To date nobody has come up with a really compelling disproof that they are more or less correct- thought and mind seem to be best explained as "emergent behavior of a complex system".
Now, whether can we bootstrap a thinking mind by running complex simulations on computers remains to be seen. I'd like to believe there is some interesting physics going on in brains we can't simulate with straightforward physical models, but there really isn't any good evidence for that.
Can you elaborate on this class of 'something'? There are intelligent behaviours that can be reproduced this way, so without more specifics I feel like you're making the same mistake you're condemning.
Worse, I don't think you're prepared to prove that a mostly pure-neuron simulation is incapable of giving rise to that certain complex quality apparently far removed from the level of the simulation.
No, questions can be answered by anyone, if they provide a relevant answer.
What the person making the claim thought is irrelavent. The parent gave you a valid reply.
So, basically, you're the only one being an 'asshole' here.
From my observation, it appears people assume "reductionist" means "simplistic" as opposed to "representing a complex system as no more than the sum of its components". To be honest, I think it's just an etymological problem: "reduction" sounds bad, so "reductionism" sounds bad.
The grandparent here follows this pattern. Using the actual meaning, it would be paradoxical for an approximation to be reductionist. In fact, it's quite the opposite. A reductionist would hold that the most accurate high-level model of a brain would be the most low-level model of a brain. If we could model each atom precisely then we would, necessarily, model the whole brain precisely.
The whole thing is pretty amusing when you take into account that anyone who manipulates software or systems in any serious way needs to engage in reductionism to be able to work. It's not as if "first, we write module A, then module B, then a system magically appears from the ether" is a viable architecture. At least, not since we stopped taking neural nets seriously.
Otherwise Sperm Whales would simply be much smarter than us.
About conscience itself, my guess is as good as yours.
On a deeper level, we still have trouble simulating even a single proton.
Hmm. I need to research this some more.
I mean, my doctorate involved calculating tiny, tiny corrections (8 or 9 decimal places) to hydrogen-like energy levels, but even there I could still treat the proton as point like!
An overwhelming amount of neuroscience research data points to the neocortex reusing the same functional unit composed of 100's-1000's of neurons called a cortical column. For example one patient was able to regain her balance by rerouting the signal from her vestibular system to her tongue. So the cortical column is the layer of abstraction we should be shooting for.
Hopefully we'll move away from the 1950’s style perceptron, whose mathematical simplicity has attracted researchers but has been a major distraction to AI as they share almost nothing in common with their biological counterparts. If we don’t want to see a 3rd AI winter, we need to move to more biologically inspired approaches.
This matters, for example, if we want to simulate neuronal plasticity or the effect of antidepressants. We're an incredibly long way from a full simulation, which would also require integration with the rest of the nervous system.
To me, this mostly shows that naive beliefs about difficulty aren't really accurate ('but it should it harder to simulate a brain than a cell, because it's bigger right!').
But of course, progress is being made on simulating cells: http://lesswrong.com/lw/drk/paper_simulation_of_a_complete_c... It's a high-level enough model that you can do it with only 1 core per cell.
Yes, you can simulate anything if you abstract away enough. But if the goal is a convincing replication, you'll need to include lots of intracellular effects.
In the end, I believe that practical simulation of a whole brain that you can have a conversation about its existence with is more difficult than the development of quantum computing powerful enough to break all existing crypto.
The work is still interesting, if only because it pushes the limits of what we can do. But I might want to start with an ant brain.
>it is perfectly possible to have a high-level simulation of a 'big' thing which is easier to run than a low-level simulation of a 'small' thing
While this is true, it's not being shown here until we know that this particular high-level simulation actually works. If it doesn't work then it's more of a failed attempt at creating a simulation.
Also note that: "The model accounts for previously observed gene essentiality with 79% accuracy." This is REALLY bad. Considering they are 'seeding' the model using a biased structure that automatically "knows" that certain genes are essential (for example, they have a 'module' for replication that probably zeroes out if you are missing an important replication component, as opposed to simulating how those genes work at a low level) this is probably somewhere close to as good as flipping a coin.
I think an interesting number would be how many atoms does it take to simulate one atom reasonably well. Then you could get correct results for some ( really ) small biological functions.
It's hard to say which is the better approach. On the one hand you have MD from which you could technically build an accurate model "from scratch", folding all of your own proteins as part of the simulation, but it is exceedingly expensive to do that. On the other hand, you have a model built at the level of biochemical reactions and molecular biology, which is computationally feasible, but then you don't have enough information from bench work to flesh it out.
Massive waste of cycles. Never gonna make a useful prediction.
Reduced representations that contain the minimum of complexity to express the phenomena of interest are far more useful.
I'm no molecular biologist, so i'll happily agree if you say there are some fraction of water molecules broken down, or formed by various cellular processes, but that must be a tiny fraction of the total, and could just be special cased.
It just seems like you're willfully ignoring everything physics has to say about how the universe works so you can get a big number for your calculation.
My understanding of "solution" is stuff floating in water. If there's some value in breaking out each atom, i'm all for it, but i kind of suspect the spherical cow is good enough in that case.
Even if i'm wrong there, each atom would introduce at least 6 degrees of freedom. IIRC, water is slightly polarized, and that's likely relevant to at least some reactions. Replacing the 18 degrees of freedom with 6 to represent a rigid water, or perhaps 10 or so to represent a flexible version, modeling as the molecule will dramatically lower complexity.
In that case, don't even bother with trying to work out what it would take to mathematically model biochemistry yet. Go study some basic physics and chemistry for a few years instead.
I am not attempting to be insulting, I am attempting to point out to you that you need to go and learn a hell of a lot more about the sciences than you seem to know at the moment if you are going to have any sort of productive discussion about the technical aspects of modelling biochemistry.
http://mgl.scripps.edu/people/goodsell/illustration/mycoplas...
I don't understand why it's easier to simulate each individual atom than it is to simulate things with structure. I would assume a simulation would regard the long yellow strands of dna as dna rather than the component parts.
I'm used to a world where having structure implies constraints. With constraints, simulation can avoid a bunch of special cases and is therefore faster. I find foobarbazqux's approach of simulating each atom upsetting. But that's ok. You guys seem to think it's the way to go, and you all get results. shrug
-edit-
I don't think simulating water is easy, but i believe it's likely easier to simulate water than it is to simulate all possible states of the constituant atoms.
That's not to say that you can't get useful results by treating solvent as a continuum (or even ignoring it entirely), they'll just have a different kind of explanatory power.
[1] http://www.sciencedaily.com/releases/2013/07/130728134055.ht...
Perhaps a more interesting question is whether or not the brain can really be simulated by a classical computer at all. If quantum mechanics plays a role in the function of the brain we will need a quantum computer to do the job. A classical simulation of a quantum brain would be a very odd beast indeed! It might exhibit behavior that seems like intelligence but somehow falls short.
(http://watarts.uwaterloo.ca/~pthagard/Articles/quantum.pdf)
Massive array at 8nm and there's your positronic brain.
See http://www.stanford.edu/group/brainsinsilicon/about.html
Maybe the simulation algorithm is the thing that needs improvement. Some day one finds a way to reduce the complexity an order of magnitude, then the era of human brain is over.
Not at all. A classical computer can accurately simulate a quantum computer, albeit slowly.
Before you answer, please consider how certain you are that you are not a simulation.
Edit: I don't know the answer, and there probably is a large gray area. But I do know this starts to make me uncomfortable the more accurate it gets. Time to go reread Egan's Axiomatic again, I guess.
Considering we experiment on animals to reduce human suffering (heck, we kill and eat them just because they're tasty), running a tiny fraction of a simulated brain for a few subjective seconds doesn't worry me much. Of course, as computers get faster and cheaper this could become a huge issue.
It's only tangentially-related, but Robin Hanson has some good talks and papers about emulated minds. The shortest summary of his views is probably this talk: http://www.youtube.com/watch?v=9qcIsjrHENU. A more in-depth version of that talk is at http://vimeo.com/9508131.
Insane maybe, but why in pain? Why would you add pain signals without also communication?
Even currently existing lower animals (lobsters and down) don't feel pain - why would you expect a simulation to do so?
If we are in a simulation, then ipso facto, we have no idea what 'reality' outside our simulation is actually like. We can only surmise that it probably resembles our simulation, in as much as the creators probably used their own reality as a template.
Even more reason why there is no functional difference, subjectively speaking, whether we're living in a simulation or in actual reality.
Only when given the choice to enter or leave a simulation does the distinction becomes meaningful.
I'd rather err far on the side of assuming sentience and personhood if there is a shadow of a doubt, simulated or otherwise.
But such a test does not work in a possibly-tainted environment. A neural net that seems to communicate might just have a hidden encoding of ELIZA.
You can err on the side of assuming sentience if you want, but don't think it's out of self-preservation-empathy-logic.
Isn't that only true if you assume that our definition of sentience is universal or that there is no higher development of consciousness? It seems to me that it doesn't even require a very skilled science fiction/fantasy writer to develop such a concept or ability.
Sure there might be higher levels, but we're not talking about higher levels?
On the topic of sentience, I just think it's exceedingly unlikely for there to be an intelligent organism that does literally nothing with its intelligence. Some kind of self-aware rock. And on top of that the mechanism of the intelligence would have to be almost impossible to study or plug wires into to try to force it to communicate.
See also: http://terrybison.com/page6/page6.html
It's not like this is some crazy threshold designed specifically to apply to humans, it applies to a bunch of animals too.
At a certain level it's kind of like checking for turing completeness.
I find it difficult to see how this could ever be unethical (given the status quo).
From the press release (emphasis mine):
"The nerve cells were randomly connected and the simulation itself was not supposed to provide new insight into the brain"
We could do it now (hardware-wise), if we really wanted to.
I would be very surprised if it took us another 26.6 years to simulate a brain in real-time. Maybe half of that time. And a few years later you will be able to cheaply backup your own brain in the cloud in case you have an accident.
How many gears from a mechanical computer does it take to simulate a micro processor?
If we were to one day to closely simulate the brain, I am sure the method will not involve computers as we know them today.
Imo brains themselves are like mechanical gears compared to the potential of processor.
Mathematically, we could figure out how much computing power we would need to match the human brain (24 bytes per synapse * number of synapse, etc).
I think I would be more interested in what was the specific of their experiment.
- What kind of software were they running?
- Was the software bug-free(yeah right)?
- Was the software optimized?
- Caching?
Simulating a lump of tissue is definitely useful, it's when they start talking about simulating a complete brain that I get lost.
A brain doesn't happen overnight. It is the result of a long developmental process encompassing embryology, learning and experience acquisition.
How do they plan to wire the whole thing?
Micro- and macroscopic connections are well understood, it is the so-called meso-scale (in between) that is troublesome.
There will never be a non-destructive way to do extract the information of a live brain, and I'm not sure there will ever be a way to extract it at all. The current methods[0] allow to extract either the wiring or the genes expression.
If they simulate cortex, they wire arrange it into cortical columns (size of each column is between 50,000 and 100,000 neurons and there is over two million of them). Those columns have many different layers and neuron types. Understanding how cortical columns and minicolumns work is very important.
> So they wire it like the brain is wired.
As if anyone knew how to do that. Even assuming that you can do both accurate micro-tractography and extract cell-by cell epigenetic information out of a single column, you'd still be missing other cricial information like individual synaptic strength.
I'm not talking about cortico-thalamic connections (required to model epilepsy) inter-columnar connections, and whatever regulation happens in the white matter.
Relying on the probability distribution of connections as a crude approximation is useful, but far from the real deal.
Your real deal has nothing to do with current brain science and these simulations.
With current brain science, call me perplexed. Could detail were you think I'm off base?
Disclaimer: this is completely out of my field - just a layman speculating.
Is there any research on what other chemical/electrical interactions might be necessary to simulate the brain?