The Computational Theory of Mind (2015)
plato.stanford.edu
plato.stanford.edu
Edit: My comment is basically asking what saying "the mind is a Turing machine" really means. It seems like the main implication someone would take away from that statement is that it's reasonably practical to describe mental processes with a computer but given that the question is formulated as a binary is/is-not. The mind "is" a Turing machine given ordinary physics which I believe say everything can be approximated by a large enough Turing machine. But this sense of "is" doesn't imply the Turing machine mind model is useful.
To the details: We don't want just one tape on the Turing machine. It's fine but slow, like using unary Church numerals instead of binary Cantor numerals. We usually assume at least two tapes. Similarly, we often ignore tapes and use a random-access memory instead.
Edit: I guess that this is how we're talking today? From the article:
> It is common to summarize CCTM through the slogan “the mind is a Turing machine”. This slogan is also somewhat misleading, because no one regards Turing’s precise formalism as a plausible model of mental activity. The formalism seems too restrictive in several ways.
The article then lists senses, finite memory, concurrency, and determinism as four ways in which the brain might differ from idealized Turing machines.
Concurrency doesn't make Turing computable processes capable of anything beyond what they can already do. At best it buys them speed, maybe some expressiveness for the programmer, and bugs.
Nondeterminism is useful for some algorithms but it also doesn't make it possible for Turing computable processes to go beyond Turing computability. See for example https://arxiv.org/pdf/cs/0401019.pdf.
The most interesting item in that list would be the senses of organisms, as some of those might indeed involve processes that are not computable.
Also, some aspects of the brain's activity may not be computable. See Roger Penrose's books for details, such as p. 377 of Shadows of the Mind (a section about possible noncomputability in some physical processes).
Well, that's the Church-Turing Thesis, which most people accept but isn't being discussed by me or the OP - the topic being "theories of the mind", I think.
I don't believe this is correct. If the brain's model/the brain itself is Turing-complete, that doesn't mean a Turing machine can simulate; it means that the brain's model can simulate any Turing machine qua Turing machine. This does not imply the reverse, that any Turing machine can simulate it.
Maybe someone can clarify whether this is correct.
If you phrase it that way, then it seems more intuitive that the brain could be described as a Turing machine. The definition of a Turing machine is deceivingly simple and cause us to question what we perceive of as complex. The transition function (and the tape) actually make magic happen if you compare it to any "dumb" mechanical or electrical device.
Further, there is a conflation among models and implementation from the parent commenter. Minsky propounds in his automaton theory that a machine may be implemented in any way and it is only its formal history of signals and responses to states which determine the machine. In this sense a biological machine made of organic bases can be considered the same to that of gears and another of semi-conducting transistors assuming their states are described by the same histories.
More so, in the same automaton theory any environment which interacts with a machine in itself must be a machine by symmetry. The classification of each is arbitrary for consideration of the states you’re interested in.
The author of the article remains correct as we must be able to model at least one faculty of the mind as Turing complete because we can also compute by the nature of computation itself. This remains useful as then we can project the guarantees of a Universal Turing machine onto that specific faculty of the mind which is Turing complete. This then allows us to explore those novel guarantees of the mind, those which are also of the Universal Turing Machine since by that faculty having the properties of UTM, it is to be considered a UTM, as opposed to having no guarantees.
The article clarifies that "the mind is a Turing machine" is not what the Classic Computational Theory of Mind claims. See section 3 "The classical computational theory of mind":
https://plato.stanford.edu/entries/computational-mind/#ClaCo...
To summarise, the article suggests that, while: "It is common to describe CCTM as embodying “the computer metaphor”. This description is doubly misleading." (first because nobody claims that the mind is a computer and second because the comparison to Turing machines is not a metaphor).
The article concludes that: "CCTM claims that mental activity is “Turing-style computation” [snipped for relevance]".
In short, nobody really claims that the mind is a computer, a digital computer, a Turing machine, a metaphorical Turing machine, etc etc. The idea, if I'm not misrepresenting it myself, is that the mind can think like Turing machines can compute.
It means that if we internalise this idea we can use all the language we have developed to say precise things about computers/software/hardware to also say precise things about our minds.
After all “computer” was a human job description...
All descriptions (theories) are, fundamentally symbolic/linguistic.
Because that's what your interpretation of basic materialism implies.
Computability theory demands repeated combination/selection and memory. It doesn’t otherwise care how this is accomplished.
In practice? No chance.
Computation is just an abstract description of the behavior of some classes of systems. But it doesn't exhaust the set of possible descriptions for a system. Some subset of computational systems might also have a description in terms of consciousness.
I wonder: is this information outdated?
A Neural Turing Machine (first described in 2014 by Alex Graves) is a recurrent neural network architecture with an external memory store. Reads and writes from and to the memory are controlled by an attention mechanism. A newer version is the Differential Neural Computer (first described in 2016, also by Graves).
The setup is not fundamentally different to the Elman networks or Long-Short Term Memory networks other than the mechanism by which "memory" is manipulated and storage, retrieval or discarding of "memories" is decided, although the mechanisms are very similar too (for instance, in LSTMs, you could say that training a network to decide when to "recall" a weight value is essentially similar to the "attention" mechanism).
Is there a significant difference between an LSTM-based neural architecture with a "reverberatory" memory and one with an external storage, both controlled by similar mechanisms?
I would say- yes.
Fodor was a snappy writer and talker. I urge you to view his videos if they can be found on YouTube. Unfortunately he passed away recently.
The argument goes something like this:
1. The computational theory of mind is the only remotely plausible theory of mind we have
2. A remotely plausible theory is better than none at all
A TOM needs to explain "mental life", at best the CTOM provides a model of a very narrow sort of cognition (inference over propositions).
There's a gigantic (and in my view, deeply implausible) leap from "hey this kinda works for modelling inference in animals" to "hey this is how The Mind! works".
Not only do all CTOM models fail for actual inference in animals where we can be reasonably sure inference is taking place (due to the frame problem), it clearly fails for non-inferential processes and states (eg., emotions/environmental-action/...).
These non-inferential processes are regarded by CTOMists as "black boxes" that just "plug into" the "Real Mind" (ie., inference over propositions).
I dont think Fodor's argument holds here: it is tantamount to saying, "hey we've explained light with waves, why dont we just explain everything with waves!" -- the cost to that approach should be obvious.
The problem this field has is that its being lead by computer scientists not neurobiologists. You ask a computer scientist what the right model of anything is, and they'd reply with a logic.
We do not, however, model causal reality with logic. Temperature isnt computed from a 'logic of molecule motion', it is done via a causal model which relates causal variables to one another.
As a theory, not only is it incomplete, it is only remotely plausible. He concedes that up front!
It’s like the old joke about capitalism - worst but for any other.
I would expect a response to give a better alternative, not to lay out admittedly deep problems with it.
If there are fatal problems with Theory-A, Theory-A isn't true. No matter how much Theory-A might help with some other problem you have.
The CCTM is the research paradigm for modern AI, parts of cognitive science, etc. -- and insofar as it provides a clear set of assumptions to arrive at useful models, so be it.
Alexa can turn the lights on, for sure. She may even be able to reason a little (if-then-else, etc.). I doubt she will ever know what a "light" is, or what she is doing when she turns it "on".
That would require Alexa to have lived a human life, and to have lived in a deep and complex social/physical environment. There is no "logic" which can specify such things in a limited set of propositions: the effect of the world on animals is not merely to add propositions to their "set of beliefs".
Rather, animals are first traumatised by the world: their emotions, physical memory, instincts, etc. are all unmindfully coerced by their environments. Only with a peculiar sort of frontal lobe are those things expressible as propositions -- but they arent propositions, as evidenced by the infinite number of them required to capture the effects.
What we need before understanding inference, is to understand on what inference operates: the mental life created by the effect of the world on the whole mind of the animal.
You mean Alexa may never know what we mean by "light" or "turning it on". Neither would an intelligent alien that doesn't rely on sight. That doesn't entail that such a creature isn't intelligent, or doesn't have a mental life, or its operations doesn't operate on a model consisting of a set of propositions.
> There is no "logic" which can specify such things in a limited set of propositions: the effect of the world on animals is not merely to add propositions to their "set of beliefs".
That's conjecture, although I think the way you've framed it is misleading. Instincts are also "beliefs" in this model, and the operation of a mind can have multiple layers with inconsistent sets of "beliefs" that sometimes drive seemingly inconsistent behaviour.
But your argument with Alexa is in my view in the wrong "direction." Alexa doesn't know what it means to be a light, and perhaps a computer "never" will.
But the real question is, how do humans "know what a light is", or what do you mean when you say a human knows what a light is.
My intuition is similar to yours, that our living in a "deep and complex" environment has something to do with it, but what?
The deep and complex environment might explain how we learn what a light is, but what is it? To put in Fodor's terms, what is the representation?
When you or I "think" (what is thinking?) of the light (what does it mean to think of the light?), what is going on in our heads?
I suspect whatever theory of representation you come up with will look something like a computational theory. The notion or concept of the light will be "stored" and have "relationships" and so forth.
edit: readability
I think you're actually agreeing here with mjburgess' point that therefore theory-A should be taken as the basis for a research paradigm.
Question is influenced by this idea originally from the 80s: https://en.m.wikipedia.org/wiki/Computational_irreducibility
We know it doesn't. There's no metaphorical linear tape in the brain; it's a network of neurons. But a Turing machine can simulate a network of neurons (see Machine Learning), just as a brain can model a Turing machine (see Programmer). There are currently things a brain can think about that a Turing machine cannot, but the question is whether that will continue to be true despite the steady advance of (computer) science.
The article above has a good summary of the problems with the idea of neural networks as simulations of biological neural networks:
These appeals to biology are problematic, because most connectionist networks are actually not so biologically plausible (Bechtel and Abrahamsen 2002: 341–343; Bermúdez 2010: 237–239; Clark 2014: 87–89; Harnish 2002: 359–362). For example, real neurons are much more heterogeneous than the interchangeable nodes that figure in typical connectionist networks. It is far from clear how, if at all, properties of the interchangeable nodes map onto properties of real neurons. Especially problematic from a biological perspective is the backpropagation algorithm. The algorithm requires that weights between nodes can vary between excitatory and inhibitory, yet actual synapses cannot so vary (Crick and Asanuma 1986). Moreover, the algorithm assumes target outputs supplied exogenously by modelers who know the desired answer. In that sense, learning is supervised. Very little learning in actual biological systems involves anything resembling supervised training.
https://plato.stanford.edu/entries/computational-mind/#ArgFo...
Such as? Are these things forbidden by theory, or just simply beyond current engineering practice?
You say this as though we understand what "thinking" is. We do not.
If that includes "have a conscious experience," then you're the one who's going to have to descend into hand wavy nonsense to explain how that's possible. Unless you've solved the hard problem of consciousness.
If there is nothing inherently remarkable beyond scale then no hand waving needed.
The remarkable aspect is that of conscious awareness. You (presumably) have an experience of the world, in a way that a computer does not. Paraphrasing Nagel, "there is something it is like to be you."
Most people don't think that this is true of an executing computer program, for example - it executes whatever its instructions are, and even a self-modifying program, as you described, doesn't change that.
There is no known way to write a computer program which has conscious awareness, and no plausible reason that scale should affect this. If you scale up a computer program, or a computational neural network, there's no reason to believe that it wouldn't just be a very big machine, blindly executing its instructions with no conscious awareness.
The proposed explanations that do exist are all nothing but handwaving at this point, hence my original comment. The burden of explanation here is on those who claim that the brain is nothing but a computing device, since our current models of computing devices can't explain consciousness.
Primates, which are very similar to us, do not have a comparable subjective experience. We know that because we can communicate with them and they don't have that much to say.
There are also brains in other animals that share a lot of similarities and are larger than ours.
Yet no beings (that's were aware of) have a subjective experience as rich as humans or can self-referentially communicate about that experience.
But turn up the clock speed on my gaming rig far enough and I get to debate the meaning of existence with it?
You have to ignore so much that is obvious to come to a reductive, materialistic conclusion like that.
[Edit]
Not sure if it's worth adding this but I do think we can build a computer that passes the Turing test. I also believe (sooner rather than later) we'll have a Siri-like AI that will provide enough companionship that a relationship can be formed with it.
We could even teach that AI to discuss subjective experience in a believable way.
I'm saying that your hypothetical gaming rig because it can simulate any type of classical computation including running a simulation of your brain and thus would be sufficient given a programmer.
There is no special need for a specific model of "learning" in a classical setting. Given an inference procedure, such as induction, adbuction or deduction, that can derive new facts and rules in a logical language from observations and a pre-existing theory (i.e. a pre-existing set of facts and rules), all it takes to "learn" is to store the newly derived facts and rules to a database.
I mean "learning" in the sense of Mitchell's definition of _machine_ learning, as (informally) the ability of a system to improve its performance from experience. In this sense, a system that starts with a database of logical facts and rules and adds new facts and rules derived from new observations is "learning".
You can find many examples of learning in a classical, logic setting in the early ('70s and '80s) machine learning literature, particularly with propositional logic learners such as decision list and decision tree learners, the most famous of which are J. Ross Quinlan's ID3 and C4.5 decision tree learners. The field of Inductive Logic Programming studies learning in First-Order Logic languages, especially logic programming languages such as Prolog and Answer Set Programming, and includes early systems such as Shapiro's Model Inference System, Quinlan's FOIL (First-Order Inductive Learner, essentialy a relational version of ID3), Muggleton's Progol and Srinivasan's Aleph (based on inverse entailment), and more recently ASP learners such as ASPAL (Mark Law), or Statistical Relational Learning techniques, e.g. by De Raedt, Kerstig, Getoor, Taschar and others; etc etc.
Bottom line- there is a huge body of work on learning in a classical, logic setting. There is no serious objection that "there _are_ good classical models of learning". Such models are all over the place in machine learning. In fact, they tend to be the most carefully characterised models of machine learning.
It is worth mentioning that there are at least good reasons to reject nearly all of these theories (including physicalist and material thesis). Many aspects of each theories turn out to imply highly unintuitive effects. But I would recommend you read for yourself, since the mind-body problem is immense and very interesting at every step. The SEP articles on both of these[1][2] are quite good.
Elsewhere in this thread some have pointed out that the ontological assumptions of philosophers fade away as we approach the mind in our inquiry. This is at least partly because the mind stretches our understanding of knowledge and matter themselves, and to an even greater degree, our intuitions thereof.
[1] https://plato.stanford.edu/entries/intentionality/
[2] https://plato.stanford.edu/entries/consciousness/; see especially §4.
The first related article linked from your link was "Humans cannot consciously generate random numbers sequences: Polemic study." (https://www.ncbi.nlm.nih.gov/pubmed/17888582)
> 3. The classical computational theory of mind
> The label classical computational theory of mind (which we will abbreviate as CCTM) is now fairly standard.
> Turing computation is deterministic: total computational state determines subsequent computational state.
Astrology itself is not scientific but, if you use a version of it that's falsifiable, you absolutely can do a valid scientific study on it by making astrological predictions and then testing to see whether they come true. I think it's pretty widely agreed that we've already done this and the (valid scientific) result was negative.
This is beside the point. The point of contention is whether or not falsifiability is what makes science what it is, or should be. Genuine science (such as exploratory papers) very often does not start by specifying a falsifiable hypothesis. Bad science, such as astrology, does often propose falsifiable hypotheses. Therefore, astrology can be falsifiable. Therefore, according to Popperian demarcation, astrology counts as science, or it's scientific (useful to remember that Popper counted Darwinian evolution as non-science).
Falsifiability isn't enough for something to be science; it's not necessary and sufficient - because otherwise astrology is science, and exploratory research, popular in many scientific fields, isn't science. The fact that astrology's claims have been falsified does not discount it as science, since a great number of genuine scientific papers also successfully falsify their hypotheses - finding a null result is an example of falsifying a hypothesis.
There is plenty of useful work which doesn't specify a falsifiable hypothesis, but it's not science until it does so.
> Therefore, astrology can be falsifiable. Therefore, according to Popperian demarcation, astrology counts as science, or it's scientific
No. Again, being falsifiable means that astrology can be a subject of scientific study. It doesn't make it science in and of itself.
Science is work that follows the scientific process: Choose a question to answer, formulate a hypothesis, make testable predictions based on the hypothesis, test the predictions, analyze and report the results. We can come with a new term (maybe 'pondering'?) for trying to answer questions without testing hypotheses, but by definition it won't be science.
That's quite a bold statement which is not supported by current work in the philosophy of science. Would you be willing to claim that most papers submitted to Nature don't count as science?
>No. Again, being falsifiable means that astrology can be a subject of scientific study. It doesn't make it science in and of itself.
The claim was that falsifiability is necessary and sufficient to count as science - so really we're in agreement. Making falsifiable claims is not necessary and sufficient demarcation of science and pseudo-science. You need something more than falsifiability to distinguish science from pseudo-science. The question is: what is that thing?
>but by definition it won't be science.
By whose definition? You're sending mixed messages - why is physics a science, rather than merely capable of being a subject of scientific study? We can make claims in physics that are just as falsifiable as the ones in astrology.
It's also unwise to paint an idealistic vision of science (falsificationism) in contrast to how it's actually practiced; from SEP:
>Popper’s focus on falsifications of theories led to a concentration on the rather rare instances when a whole theory is at stake. According to Kuhn, the way in which science works on such occasions cannot be used to characterize the entire scientific enterprise. Instead it is in “normal science”, the science that takes place between the unusual moments of scientific revolutions, that we find the characteristics by which science can be distinguished from other activities.
That's not entirely correct. We don't know that a Turing machine can have conscious experience. There's no substantiated "in principle" explanation for that.
Either way we certainly don't have any indication that it is impossible in principle for a Turing machine to perform the same kind of calculation that gives rise to conscious experience in humans, or that this calculation would for some (supernatural?) reason not have the same outcome if performed by something other than a human brain.
I do know that I have a conscious experience, so that's quite meaningful - to me. I cannot check that any other being has similar feelings, so the doubtful question would be whether "an X other than me can have conscious experience"; but people with good manners make the polite assumption that it's also true for other similar beings.
How would you write a computer program that "tells that story to itself," such that it actually has an experience of the world, as opposed to just being a machine executing a program without any conscious awareness?
Edit: also, whether we're in charge of our thoughts is a separate question from whether we possess consciousness. Even if we're not in charge of our thoughts, we still have a conscious experience of them.
In fact, there's pretty good evidence that what we call consciousness is a post-facto rationalization of the subconscious brain processes that determine an automatic answer of your brain to stimulus (not that it makes them deterministic, but certainly they're not "rational" in the classic sense).
A perception that, divorced from all other facts, entails a false conclusion.
So your perception of conscious subjectivity could indeed be an illusion.
Put another way, how would you program a computer to have a perception of conscious subjectivity, as opposed to just blindly and unconsciously executing its instructions?
No it doesn't! Assuming by "consciousness", you mean a phenomenon that's not reducible to unconscious particle interactions, which is typically what is meant in philosophical discussions of this topic.
We have some mechanistic theories for consciousness [1]. It basically amounts to the same sort of illusion that your single core CPU uses to achieve the illusion of parallelism, ie. context switching between internal and external mental models produces the illusion of consciousness.
[1] https://www.frontiersin.org/articles/10.3389/fpsyg.2015.0050...
I'd say that's an unfounded assumption, which doesn't come up in the argument you're responding to - even if it's somewhat 'popular' elsewhere.
The argument made is that consciousness is (or includes) a form of perception; not that this perception is independent of mechanistic components. With this definition, you assertion that 'conscious subjectivity is an illusion' is inconsistent, as an illusion is a complex form of perception that requires a consciousness to perceive it.
Following your CPU example, there is parallelism from the point of view of the program being executed, even if it's simulated from a single-core mechanical basis (threads and context-switching).
It's not really. Consciousness quite literally does not exist in mechanistic/eliminativist conceptions of consciousness like the link I provided, just like cars don't really exist because they aren't in the ontology of physics. My clarification of "assumption" is simply because many people don't know this.
> Following your CPU example, there is parallelism from the point of view of the program being executed, even if it's simulated from a single-core mechanical basis (threads and context-switching).
No, there is concurrency but not parallelism.
If I understand you correctly, that's a pretty harsh criterion for existence, isn't it? Even though a car is just a composite of metal atoms under a precise configuration and not a metaphysical entity on itself, you can still use it to drive you home. I suppose that makes me an utilitarian.
> No, there is concurrency but not parallelism.
You're right, my bad. I've forgotten my precision from my college days. Still, that's good enough for the program, just like my consciousness is good enough for me, even if it's entirely mechanistic and doesn't exist in the same way that cars don't exist.
If it looks like a duck, swims like a duck, and quacks like a duck...
For example, I'm assuming you're conscious, because you posted a reply that was on-topic and coherent with the conversation above it.
To come back to the original claim:
> Simple: Find something that the brain does that could not, in principle, be emulated by a Turing machine or equivalent. So far we don't know of any such thing
We may not "know" of such a thing with certainty, but we have a strong candidate in consciousness.
There are two possibilities here:
One is that Turing machines are conscious (and we're monsters for what we do with them), in which case we still have an unexplained panpsychic phenomenon which we would need new science to understand.
The other is that Turing machines are not conscious, in which case there's an unexplained phenomenon in how an object like the brain can give rise to consciousness. In that case, the question of whether a Turing machine could in principle emulate consciousness depends on what the cause of consciousness is. It's certainly possible, and doesn't even seem particularly unlikely, that we find that Turing machines cannot do this, and that something other than "computation" is needed.
I don't have no reason or explanation that suggests that you have a consciousness either. You could be a very elaborate chatbot that posts coherent replies at online forums. Also, I do know whether I'm a chatbot myself or not, but you can't tell about me just from the replies written here.
> We may not "know" of such a thing with certainty, but we have a strong candidate in consciousness.
The problem with that is, you don't have a test for consciousness. There's a strong candidate in MRI brain scans (at least for humans), but you can't really be sure.