Greg Fish: Against Causal Functionalism Tuesday, Nov 24 2009
AI and philosophy 1:49 pm
Greg Fish, a science writer with a popular blog who contributes to places like Business Week and Discovery News, has lately been advancing a Searleian criticism of causal functionalism. For instance, here and here. Here is an excerpt from the latter:
A Computer Brain is Still Just Code
In the future, if we model an entire brain in real time on the level of every neuron, every signal, and every burst of the neurotransmitter, we’ll just end up with a very complex visualization controlled by a complex set of routines and subroutines.
These models could help neurosurgeons by mimicking what would happen during novel brain surgery, or provide ideas for neuroscientists, but they’re not going to become alive or self aware since as far as a computer is concerned, they live as millions of lines of code based on a multitude of formulas and rules. The real chemistry that makes our brains work will be locked in our heads, far away from the circuitry trying to reproduce its results.
Now, if we built a new generation of computers using organic components, the simulations we could run could have some very interesting results.
On his blog, he says:
The actual chemical reactions that decide on an action or think through a problem don’t take place and the biological wiring that’s the crucial part of how the whole process takes place isn’t there, just a statistical approximation of it.
This is just another version of vitalism. Computers lack the “vital spark” necessary to create the “soul”, even if they implement the functions of intelligence and self-reflection even more effectively than the biological entity that inspired their creation. But those functions are what create intelligence and self-reflection, not magic chemistry-that-can-never-ever-be-simulated-even-in-principle.
There is quite a bit of fuzziness in chemical reactions themselves, and not all this fuzziness is necessary to implement intelligence or “self-awareness”.
Say we have a molecular dynamics simulation of the brain in complete and utter detail. It behaves exactly the same as the intelligence that it is “simulating”. You can say “it’s just a simulation”, but it can achieve all the same things that the original can, including be your friend or even possibly kill you. In such circumstances, “it’s just a simulation” is quite pointless hairsplitting. Certainly, some atomic configurations are conscious and others are not, but there is no vital force that biological molecules possess that high-resolution simulations of those biological molecules would not also possess.
If it walks like a duck, and quacks like a duck, it’s still possible that it’s not a duck, but if it has a perfect emulation of a duck brain and can walk around in a duck body, then it may as well be a duck.
November 24th, 2009 at 2:20 pm
So if you have the body a man and the brain of a duck would duckman be born?
November 24th, 2009 at 2:33 pm
“This is just another version of vitalism. Computers lack the “vital spark” necessary to create the ’soul’…”
I think I was probably being a little too vague or too obtuse in my explanation because nowhere do I say that computers lack any spark or that there’s some sort of supernatural essence to the human brain. My big argument is that the way computers work and the way neurons work produce very different frameworks (something on which Michael Vassar and I found common ground very quickly) and all we do by getting computers to run simulations of working brains are pretty pictures and bursts of data. Therefore, we should be trying to bridge the two by combining the flexibility of the brain with a computer’s raw processing might rather than try to force a square peg into a round hole.
It seems unfair to say that I’m against causal functionalism when I’m talking about the technical aspects of creating an emulation of organic brains in the digital world. If I really thought we needed a divine spark to make intelligent machines rather than playing around with hybrid organic/digital models to merge the two, I would’ve left it at that, as some of those who actually ascribe to vitalism do. Just see the comment section of my Discovery Tech article for a shining example of that.
“There is quite a bit of fuzziness in chemical reactions themselves, and not all this fuzziness is necessary to implement intelligence or ’self-awareness’.”
Technically, yes. The concepts of awareness and intelligence come from certain parts of the brain and not every neuron firing away is a vital component of making us who we are. It’s actually an argument I make often to counter creationist claims of “irreducible complexity.”
“… if it has a perfect emulation of a duck brain and can walk around in a duck body, then it may as well be a duck.”
To us, maybe. To some people, it would just be a terrifying impostor. To me, it would be a neat trick demonstrating our knowledge of neurology and robotics. Unlike living things, it’s not going to evolve or reproduce. We have to tweak its software and hardware, and make copies of it in our labs. It’s not going to be self-propagating like living things.
But again, this may very well be more of a question for philosophers than anything else. I’d rather stick to the realm of functional requirements and basic architecture since this is what I’m actually qualified to address.
Aside: I contribute to Discovery News rather than Discover Magazine. The only work of mine on Discover comes in the form of links to my blog from Bad Astronomy.
November 24th, 2009 at 4:45 pm
A digital process can potentially simulate any analog process if it has enough details, so I’m not sure why we have to create organic/digital hybrid computers.
Neurons operate based on Maxwell’s equations, chemical gradients, neurotransmitter concentrations, and the like. It may be that a digital computer cannot simulate these by what you would consider efficient means, but I don’t see why they couldn’t be simulated in principle.
You say that such a thing would be an imposter, but if I could remove everyone’s brain overnight and replace it with a silicon version that did exactly the same thing, I think that no one would notice. Do you think that someone would notice?
A human being with a robotic brain could reproduce just fine, as could a duck with a robotic brain. The brain is a replaceable part of the body, just like the leg, only the brain is much more complex. Not mystically complex, not irreproducible-in-silicon complex, just complex. No qualitative difference from a leg. The functions of a leg can be simulated by a prosthetic leg, and the functions of a brain could be simulated by a prosthetic brain.
I am interested in the in-principle question of whether a square peg could be put into a round hole. Practicality is a separate question.
I’m happy to hear that you’re distanced from vitalism, but I still don’t understand whether you have an objection to emulation in principle or just in practice in X years. I don’t understand why nanocomputers wouldn’t dissolve your concerns over practicality. A cost performance increase of ten orders of magnitude could be achievable over the next 50 years.
The tone of your articles very strongly imply that you are objecting to causal functionalism. Your tone sounds exactly like Searle’s, appealing to the intuitive notion of the simulation as a fundamental conterfeit, when such a notion is not coherent with what we know about the interchangeability of functional processes given varying substrates. (For instance, different animals use different protein structures to slice away at enemies.) Only in your comments here do you seem to be indicating that it’s a matter of “technical” difficulty. Will you acknowledge the in-principle possibility of perfect emulation?
Visiting the Whole Brain Emulation roadmap, you see a hierarchy of different levels of functionality that various people in the brain emulation workshop considered necessary for brain emulation. I believe that people should have to pick a level — it doesn’t make sense to reject the notion altogether. Even Hameroff and the like support the idea of quantum computers being used to emulate minds.
November 24th, 2009 at 6:43 pm
“Your tone sounds exactly like Searle’s, appealing to the intuitive notion of the simulation as a fundamental conterfeit…”
Maybe it does, but that’s not at all what I’m trying to state. The notion between “genuine” and “counterfeit” depends on your frame of reference and personal opinions. As a philosopher, Searle oversimplifies some of the key issues in AI and his Chinese Room analogy - while valid for your everyday computer - quickly breaks down when we consider the idea of simulating the process of learning and brain formation, which is very difficult but has already been done at a very rudimentary level.
All my posts and articles focus on the difference between the processes of a neuron and those of a computer chip, and how the differences in processing chemical signals in one and data in the other, ultimately reduce the outcome of computer simulations to statistical constructs.
I always point to the technical details and say that the outcomes are not the same, not that the process is somehow “suspect” or “not you anymore” just because it’s in another format. Though I see how in my Discovery Tech article, the bit about the chemistry of the brain came off as confusing or Searle-ian when taken in a technical sense.
“Will you acknowledge the in-principle possibility of perfect emulation?”
Sure. I acknowledged it a back and forth with Michael Vassar on my blog before we talked about what to do with this perfect emulation. Get enough details into a computer with enough processing power, and you will get a perfect simulation of a human brain.
My problem starts when people say that it will become conscious and self-aware. Computers see data constructs as collections of bits and bytes tied together by indexes, or as the result of specific algorithms. Biological constructs are more flexible and while they do work by certain rules, they have much, much greater leeway than we can allow CPUs to have.
Let’s keep in mind that we designed computers to be calculators and collators first and foremost. Making them too imprecise would be to abandon the initial concept behind computing devices as we know them today. It’s not a bad thing, nor is it impossible with enough time and money. But is is a pretty radical change.
November 24th, 2009 at 7:59 pm
Michael is so close to admitting the truth there is no intelligence just dumb algorithms.
November 25th, 2009 at 2:14 am
indeed a leg can be simulated by prosthetic leg but when implications come with brain, just get more complex and weird, Michael! Replacing programmed organic brain by as equally complex silicon brain, which works in same way, seems just as simple as we see in movies. You can’t say that prosthetic leg is alive either.a prosthetic leg works in same way not because it’s replica of leg but it’s controlled same complex organic brain. Here I’m leaving a question for you. Would that simulation be able to replace organic brain that is mystically complex? Would then human be alive or just terminator like cyborg?
November 25th, 2009 at 2:15 am
indeed a leg can be simulated by prosthetic leg but when implications come with brain, just get more complex and weird, Michael! Replacing programmed organic brain by as equally complex silicon brain, which works in same way, seems just as simple as we see in movies. You can’t say that prosthetic leg is alive either.a prosthetic leg works in same way not because it’s replica of leg but it’s controlled by same complex organic brain. Here I’m leaving a question for you. Would that simulation be able to replace organic brain that is mystically complex? Would then human be alive or just terminator like cyborg?
November 25th, 2009 at 2:18 am
indeed a leg can be simulated by prosthetic leg but when implications come with brain, just get more complex and weird, Michael! Replacing programmed organic brain by as equally complex silicon brain, which works in same way, seems just as simple as we see in movies. You can’t say that prosthetic leg is alive either.a prosthetic leg works in same way not because it’s replica of leg but it’s controlled by same complex organic brain. Here I’m leaving a question for you. Would that simulation be able to replace organic brain that is mystically complex? Would then human be alive or just terminator like cyborg?
November 25th, 2009 at 6:55 am
I could see creating a conscious brain emulation being a practical if too much detail is needed. For example the IBM neural net that you mentioned used a lot of hardware and a lot of power.
Also consciousness is tricky. I could see it depending on something like the precise order operations are carried out. In that case you could have two systems with the same output where one is conscious and the other isn’t.
November 25th, 2009 at 12:22 pm
Brains see data constructs as collections of synaptic weightings and neural interconnections tied together by axons. Is there a fundamental difference? Behind all that neural “flexibility” is a lot of rigidity that computers don’t have, like the necessity of only having less than 100 neural processing steps (because neurons spike only about 200 times a second) to get anything important done in realtime, like ducking a shoe being thrown at you.
Measured by what areas of information space are accessible to computers vs. brains, I’d argue that computers are actually much more flexible. They can represent a far wider variety of data structures and process them faster and in fundamentally different ways than neurons can. I also don’t see why a computer couldn’t simulate the leeway embodied in neurons.
We don’t know much about consciousness. Assuming that it works reliably on messy proteins but not on digital computers capable of simulating 100 billion parallel processors implementing the information flow that underlies intelligence (and likely consciousness) seems to be fundamentally biased towards biology just because it is a status quo.
A Turing machine with enough tape can simulate any other Turing machine. If an “organic-digital hybrid” model in your mind would include a digital machine designed to simulate organic parallel neurons, then I wouldn’t really consider that a hybrid.
It’s this sort of unreasonable demand for bio-centric computing models that creates the sort of public relations catastrophes like the IBM situation. Instead of actually trying to implement intelligence with what we have, which seems like a worthy first step towards “consciousness”, they create extremely vague homunculi of “how the brain works” that consume huge amounts of computing power but don’t actually do anything interesting.
The digital-organic hybrid approach is partially a red herring because no matter how organic-like we make computing structures, someone will always come forth and say they’re not enough like us. It’s hard to call that anything but anthropocentrism.
We know tremendously little about phenomenological consciousness. You cannot say with any degree of confidence that a conventional Turing machine cannot be conscious, because we know close to nothing about what causes it. We can’t even tell if other humans are conscious or not besides guessing.
November 25th, 2009 at 5:01 pm
It’s been a while since my bio-psych class but I think that’s a pretty major oversimplification of how things work in the brain. There’s more to it than that.
It could, but you would be effectively creating a three teared architecture which uses your basic von Neumann design, a layer which simulates biological structures and an interface which then uses the cognitive patterns to perform actual tasks. The problem is that computers do have a lot of leeway in what data they can model, but their leeway relies on very precise algorithms. For example, if my code looks like…
printf(”Hello World!”);
… the computer knows it needs to display the phrase “Hello World!” and nothing else. Now if I tweak it to say…
if userid=”Anissimov” {
printf(”Hello Michael!”);
return 0;
}
else {
printf(”Hello World!”);
}
… it knows to greet you by name but to say “Hello World!” to everyone else. And we can go on and on with this, making our algorithms ever more complex and bringing in new data structures and concepts into it. But we need to lay the ground work for them with code. (Speaking of which, pardon any of my potential syntax errors here, it’s been years since I worked with C or C++).
On the other hand, neurons use a number of chemical signals and pathways to store memories and process stimuli, then a network of constantly changing pathways which decide what to do with them. It’s like the computer code we typed out above is constantly shifting and adding on to itself. And I’m sure you’ll agree that’s what AI research is all about. Making the machine write its own algorithms and decide how it will handle situations on its own. But again, we need that groundwork. The plasticity of neurons makes biological entities able to do this by default.
I don’t think that the IBM claim was any sort of a PR disaster. In fact, it was a PR bonanza, though for all the wrong reasons as a recent post on my blog points out. They basically claimed to have taken a huge step towards cognitive computing when in reality they just simulated lots of neurons and said because it was a big number, it exceeds the scale of a cat brain and is thus on its way to becoming cognitive.
Honestly, that’s my last possible concern about this technology. AI isn’t about creating “something like us.” It’s not about people’s opinion of the technology from a metaphysical and standpoint, it’s about the potential of these devices in medical, military and supercomputing applications.
Very true, which is why this is the crucial problem with the Turing test. It assumes the “walk like a duck, quack like a duck” approach is good enough and can easily be assailed by the Searlean Chinese Room Analogy. But if we can demonstrate the framework by which the computer understands language and conversation, that thought experiment fails.
I learned how to speak and hold conversations as a child. I then learned English on top of my native Russian and I now understand the words in both languages and what they mean. If we can dig down to the fundamental need for language and communication, then we could truly ace the Turing test rather than cheat our way through it.
November 25th, 2009 at 6:21 pm
Well, it looks like we’re at least getting somewhere here, that’s good, but I still think you have an overly rigid conception of what “algorithms” and “code” are. They can be flexible, dynamic, fluid, uncertain, nonlinear, etc.
It still seems like your primary reason for believing that organic brains can be conscious and silicon transistors can’t is based on intuition, and we both know that intuition can be wrong.
The duck would have the frameworks of understanding you’re talking about, so it would swipe away the Chinese Room objection. I believe that the Chinese Room objection is wrong on so many levels, not least of all the fact that no individual neuron “understands” the global activity of the brain anyway.
The basic operating principle of the brain has to do with electric signals passed between neurons. There’s more to it than that, but that really is the basic idea. Synaptic spine shapes, neurotransmitter concentrations, etc., all take a back seat to that basic dynamic, in terms of the workload quantity.
You should also recognize that it is possible to create probabilistic algorithms that are only discrete at the lowest possible level. You can write algorithms that deal well with fuzziness and uncertainty, just like the human brain. There are programs that run on conventional computers that have a lot more in common with how the brain works than arbitrarily large networks of FPGAs (reprogrammable chips that some people claim embodies the flexibility of neurons), for instance.
It’s also worth pointing out here that neither me nor Michael Vassar are computer programmers or even studied programming in school. He studied biology and business in school, I skipped school and wrote about 1,600 short science articles for the website WiseGeek.com.
Causal functionalism implies that any system that replicates the functions of the brain will have all the properties of the mind, including consciousness. Whether or not a system that replicates some but not all of the functions of the brain will be conscious is another question.
November 25th, 2009 at 7:35 pm
Wait, I thought the issue was cognition rather than consciousness. Let’s pick one topic and stick with it for the sake of not drifting all over the place in the discussion.
My conception is based on actually creating them in real computers.
Again, that’s a gross oversimplification. It’s fine for a conceptual brainstorming, but not for an actual project.
They’re difficult as hell to write and highly prone to bugs. You sort of end up giving the machines options and hope they pick one that works. The problem is there’s only so many options you can give computers in the real world because if you don’t box them in somehow, they’ll probably crash or get stuck in an infinite loop trying to figure out what to do. Eventually, if they can’t solve the problem with the options available in their code, they have to turn control over to a human.
I think at this point, I should point out that my background is that of an IT systems analyst. So if you need a virtual brain built, I would be one of the people drawing up the flowcharts and programming requirements to do them.
November 25th, 2009 at 7:55 pm
The issue is both cognition and consciousness. At this point, since you said that it would be entirely possible to simulate intelligence on a computer with enough computing power, I have no more quibbles on cognition. Now, I’m simply trying to uphold my accusation that you are rejecting causal functionalism by saying that consciousness is uniquely proteinaceous with high confidence. You are free to define what degree of certainty you have about that.
I agree with you that point neurons, or even neurons that do not take into account all the complex neurological dynamics may not be conscious or even intelligent. We basically have to try it and see. With IBM sending out bullshit press releases on the issue, the whole world is going to turn against the entire field in no time.
November 25th, 2009 at 10:52 pm
And since this is not an assertion I’ve made, you would be talking right past me. My quote about brain chemistry vs. simulations is only true for von Neumann machines.
Quite possibly. This is why I tore into their methods on my blog, using their technical paper to perform a basic requirement analysis. But to be fair, the media took their already hyped up press release and made it even worse by presenting it as a simulation of a cat brain.
As one of the people who cited those articles and wrote about the implications of the simulation, I would have to take responsibility for that as well…
November 25th, 2009 at 11:11 pm
You said:
My contention is that a sufficiently accurate simulation on a Von Neumann machine would be more than “just” a “very complex visualization controlled by a complex set of routines and subroutines”, and that could still produce consciousness. It could be self-aware and conscious, given a fine enough simulation. I furthermore think that it makes the most to sense to regard our own brains as “just” a “complex set of routines and subroutines”, specifically adaptive routines which evolved to meet specific evolutionary challenges. This follows from physicalism. Our brains operate in a series of discrete states and certain rules govern the state transitions. The most important transitions are probably on a scale of 100 ms.
Causal functionalism implies that the “real chemistry” is not essential to consciousness. Only the functions are. To imply otherwise is acceptable, surely, but it requires violating causal functionalism. Casual functionalism implies that a Turing machine can be conscious given the right programming, because it’s the pattern that matters, not the implementation. Are you saying that a Turing machine can’t replicate the pattern because it operates serially instead of parallel or something? If so, what about a parallel network of Turing machines?
November 26th, 2009 at 10:19 am
I really appreciate the debate going on in the comments here. I have to say it seems like there might be something to Greg’s argument about data structures, but I don’t know nearly enough about computer science or neuroscience to really evaluate that claim. Where he confuses me is in saying a perfect emulation of a brain wouldn’t necessarily be conscious/self aware. Maybe I’m missing something, but it seems that if you adopt a naturalistic view of the universe it can’t NOT be conscious. The way I see it, if we ever develop a sufficiently accurate simulation of the brain (at the neuronal, chemical, or even atomic level, pick your fundamental unit), there are 3 possible outcomes: the simulation becomes conscious, or we’ve proven that we aren’t conscious, or we’ve proven that human brains run on magic. I think that gets right to the heart of the philosophical issue. The engineering issue is a much more debatable point, but I’m not sure it’s possible to even have that debate unless you agree on the basic philosophical point that a simulation of a human running on a silicon substrate would be equally conscious to one running on a biological one. Thinking otherwise is the same sort of convoluted logic that leads people to accept vitalism and the concept of p-zombies. I don’t see how Greg could disagree with Michael on that point, but maybe I’m missing some subtle distinction that allows him to accept John K Clark’s argument that we are nouns, not adjectives (nothing more than the way matter arranged a certain way acts), and still think a simulation wouldn’t be self aware.
January 11th, 2010 at 5:14 pm
Discussion above is quite interesting and helps to understand two major veiwpoints on AI. To me Greg’s veiwpoint makes more sense, beacause he stresses that “mathematical” approach towards AI is somehting different from “material” one. You can have two similar or even exact results, but trough using different means. The thing is that, as I understand Greg’s veiwpoint, is that the “mathematical” AI will be “thoughtless”, whereas “material” AI will not. It doesn’t matter, really, if our interest lays in repeating “what” brain does, not “how” he does. “How” is interesting to some degree, but we don’t care about copying brain-we would like just to mimic his functions. To put it simply: it’s like flying. Birds fly, airplanes fly too. And they do that hell better than birds, but can we call them birds just beacuase they fly? Is their flight the same? Of course not, but it’s not the point why we have airplanes. We wanted to fly like birds. We didn’wanted to build birds. Beacuase it’s much more difficult and honestly quite pointless. No, we wanted to fly-by the same token we want AI’s to be inteligent, and we don’t want them to really think. It won’t matter as far as functions of inteligence are concerned. Hell, I think that even counciousness can be done in the same way, producing outputs as if AI was really self-aware. It doesn’t mean that there is no difference between human brain and “mathematical” AI. What it truly means is that world is far more wierder than most of us are able to accept. And THAT’S scary, not the fact that AI’s can wipe us out. A thoughtless genius, how cool is that? But in order to really see that, we have to have far more better knowledge on both AI and human brain. Up to this moment wheather “functionalist” apporahc towards inteligence is right or not is just a speculation.
And by the way I was educated to be a lawyer, so that’s why it matters to me how things are done, not just that they are done (basics of juridic logic anyone?).