Linespace
Blue Brain
Linespace

3 July 2005

home 

columns

Bitbucket Well, it looks like the beginning of the end of the argument so far as AI is concerned. A true brain simulation is on the way: it's only a matter of time.

There's always been a kind of last resort argument for those of us who think the brain must ultimately be capable of simulation on a computer. OK, you don't accept that any of the existing theories is any good; OK, you think when computers play chess or whatever, they're actually doing it in a completely different way. But at the end of the day, the brain is just another physical object. If it comes to it, we can just simulate the whole thing. We can model every single neuron in a real brain. What then? We don't even need to understand how the brain works; if we copy it exactly, it will work, just as if we copied all the parts of an old-fashioned watch without having any ideas about horology. If we do that, we just will have consciousness functioning in a computer - there really is no denying it short of some hopeless retreat into dualism or whatever. Now you've always said that such a model was impossible, not just in practice, which seemed obvious, but also in principle. Guess what? Somebody's actually going to do it.

The Blue Brain project, (a joint venture of the Brain Mind Institute, EPFL, Switzerland and IBM, USA) has the eventual aim of applying terrific computer power to the simulation of an entire brain. It's not going to happen overnight, but it's going to happen. 

Linespace

BlandulaHang on a minute. Aren't you exaggerating? So far as I can see, all they plan to do is simulate one of the columnar structures in the cortex. That's a tiny fraction of a whole brain.

Linespace

Bitbucket That's the initial stage. It's a significant step forward in itself, but the ultimate aim is a total brain simulation. What I'd like to know is what you think makes this impossible? What's going to stop them, given enough raw computing power? If you accept that the brain obeys the same laws of physics as the rest of the world, I just don't see what's to stop a simulation. Really, you're just betting against the progress of technology. That's always been a bad idea. You're putting yourself in the same camp as those people who said heavier-than-air flight was impossible in principle.

Linespace

BlandulaI really need to know more about the project, I think. A lot depends on the level of the proposed simulation. Your view has always been that there is a relatively high level of interpretation on which we can find a working functional analysis of the brain - we can identify the essential working parts without having to go down into fine detail, more or less the way we can identify the way the cogs in a clock work without having to worry too much about what they're made of. But it ain't necessarily so. My guess is that these people are going to use a standard model neuron and put a whole lot of them together. But there's a tendency to under-rate the neuron. It isn't a simple switch. We know now that there are lots of different neurotransmitters that work in different subtle ways and are influenced by the chemical environment in all sorts of ways. These fine details may well be important, and it's not unlikely that to get a working simulation you'd actually have to model each individual neuron down to the molecular level, which is just beyond any conceivable technology. Even then, you might well find that you needed to model the surrounding glial cells at a similar level of detail. People tend to think glial cells are just sort of stuffing that fills up the gaps between neurons, but there's been some research which suggests their role in getting neurons ready to fire again may be quite significant. The complexity of all this is mind-boggling.

Linespace

Bitbucket I very much doubt if it's as bad as that. You have to remember that however sophisticated the behaviour of neurons may be, they ultimately only do one thing: fire - or fail to fire. I actually think that modelling all the vast number of connections is going to be more of a challenge than worrying about neurotransmitters or glial cells, but in any case get this - they are modelling on a molecular level.

Linespace

BlandulaWhat! I... I actually don't think I believe that.

Linespace

Bitbucket Well, it actually isn't completely clear what the whole-brain project will eventually be like, but certainly some of the existing work is at a molecular level. And in the current columnar project, I believe each processor is supposed to handle only a few neurons - plenty of capacity for really fine detail, I should have thought.

Linespace

BlandulaHmm. Well, I think they're going to run into some problems. Actually, in some ways, doing an entire brain is a more hopeful project than doing part. I assume they picked on a column from the cortex because, as well as being an interesting structure which probably does have something to do with higher mental functions, it looks a bit like a component; a single unit that can be unplugged and run in isolation. But that isn't really true. Granted, the cortex has this columnar structure, but it also has a definite laminar structure - it comes in layers. These columns are strongly wired up to other structures as well as internally; simulating one might end up being as useful as simulating a single cog out of a clock. But even doing an entire cortex might not work; we know that other parts of the brain are essential to certain kinds of memory, and to emotion, for example. It might easily turn out that you need almost the whole brain before the simulation will really work. I believe people can get by fairly well with most of their cerebellum missing, as a matter of fact, but apart from that it may all be essential.

I have to ask - don't you suspect this is a just a marketing ploy on the part of IBM? 

Linespace

Bitbucket Oh, it clearly is, but so what? Doesn't mean it won't work. I see your point about the column in isolation, but I don't think anybody is naive enough to think it's an unpluggable component. It's just such an obviously salient structure that it's clearly worth studying.

Linespace

BlandulaThe other thing is, whose brain are they going to be simulating? There are lots of detailed differences between the way one person's brain is wired, and another's. Now it's quite likely that some of the properties of that pattern of wiring are essential to consciousness, while others are just individual variation. If you don't have some theoretical understanding already, how are you going to be able to tell which is which?

Linespace

Bitbucket You really are clutching at straws now, aren't you? I can see it must be traumatic when events force you to reconsider your whole iintellectual outlook.

Linespace

BlandulaThe point I'm really making my way towards is that a simulation is a simulation, and reality is reality. Searle pointed out that you can simulate a thunderstorm on a computer, but nobody gets wet; surely it would be the same with a brain: you might simulate the way it worked, but you wouldn't get actual consciousness.

You see, it's important that the causal relations operating in an artificial intelligence are the same (in the relevant respects, whatever they may be) as the causal relations operating in a real brain. But if it's a computer simulation, they never will be. In the real brain, the firing of neuron A causes the firing of neuron B: in the computer, both events are caused by the instructions in a lot of code. It's the difference between someone who walks because of the muscles in their legs, and someone who walks because a puppeteer is pulling the relevant strings.

Linespace

Bitbucket No, that's false. It's more like this. Suppose you had a universal shape-mimicking device. Now you set some of these little machines to mimic the shape of cogs in a clock. When you put it together, it all works: now they're just simulated cogs, and internally they're doing all sorts of calculations and monitoring in order to keep the right form: but does that mean it isn't a clock? That it doesn't tell the time? Obviously not. The causal relations in the simulation are more complex, but they've been set up in such a way that the overall pattern is the same.  Of course no-one gets wet from computer rain: all that means is that our brain simulation won't generate sloppy piles of white and grey gunk. It's not the physical nature of the thing that matters; it's the ability to perform in certain ways - yes, ultimately to have a coherent conversation, for example. You know this is true really - you don't recognise consciousness by the physical qualities of someone's brain, but by their behaviour - whether they answer you and what the quality of their answer is. This idea about some magic property of biological brains is an idea that would never have occurred to you if you hadn't needed a bolt-hole to hide in. When a real simulation comes along, that bolt-hole is going to become pretty uncomfortable.

Linespace

Earlier

home

later