The Computational Theory of Mind (2015)

https://plato.stanford.edu/entries/computational-mind/

148 90 7 years ago

plato.stanford.edu

It seems like basic materialism to claim that the brain could be simulated by a large enough Turing machine. But that claim would not necessarily imply that a Turing machine is a useful model for brain. A Turing machine isn't necessarily even used as the working model of a computer though it conceivably could be.

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.

p1esk• 7 years ago

Your comment makes more sense if we replace “useful” with “practical”

lidHanteyk• 7 years ago

Nobody claims that Turing machines in particular are well-suited to the task. What we claim is something more open-ended: Turing-complete languages can simulate each other given a universal encoding (a program we typically call an "emulator") and the simulation is only polynomially slower and larger.

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.

joe_the_user2• 7 years ago

Nobody claims that Turing machines in particular are well-suited to the task. What we claim is something more open-ended: Turing-complete languages can simulate each other given a universal encoding (a program we typically call an "emulator") and the simulation is only polynomially slower and larger.

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.

it• 7 years ago

Finite memory just means a limited subset of what a Turing machine can do.

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).

proc0• 7 years ago

I'm still reading through the paper/article, but a Turing machine would be the substrate in this case, not the model. I think the question is whether we can model the brain's way of computing in a Turing machine, and then it will follow we can also model the brain. Turing machines are a model of computation, and if the brain's computational model is Turing complete, it will follow that we can simulate a brain inside a Turing machine.

jblwps2• 7 years ago

if the brain's computational model is Turing complete, it will follow that we can simulate a brain inside a Turing machine

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.

mikorym2• 7 years ago

I think a Turing Machine is an infinite state automaton with the capability that you can remotely access state information. [1]

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.

ethn• 7 years ago

This is correct w.r.t the formalization of Minsky.

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.

mikorym2• 7 years ago

[1] https://en.wikipedia.org/wiki/Finite-state_machine

kaffeemitsahne• 7 years ago

It might seem like basic materialism, but a lot of philosophers suddenly loose their simple and clear-cut ontological stance as soon as the mind gets involved.

TheOtherHobbes• 7 years ago

Are you absolutely convinced that you can build a sentient mind out of clockwork parts?

Because that's what your interpretation of basic materialism implies.

PixyMisa• 7 years ago

In principle? No question.

In practice? No chance.

BadThink66553212• 7 years ago

Clockwork parts? No. Parts that select from a combination of inputs? Yes.

Computability theory demands repeated combination/selection and memory. It doesn’t otherwise care how this is accomplished.

YeGoblynQueenne4• 7 years ago

> Edit: My comment is basically asking what saying "the mind is a Turing machine" really means.

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.

ukj• 7 years ago

Edit: My comment is basically asking what saying "the mind is a Turing machine" really means.

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.

DanielleMolloy• 7 years ago

Please let me apologize for the lack of understanding, maybe someone can add to this: If biological intelligence works like a Turing machine, is it even possible to know what or how certain brain areas compute without simulating them bottom-up?

Question is influenced by this idea originally from the 80s: https://en.m.wikipedia.org/wiki/Computational_irreducibility

mLuby• 7 years ago

biological intelligence works like a Turing machine

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.

michaelmrose4• 7 years ago

I don't think the idea that the brain can do anything a sufficiently fast computer can't deserves any credence. Those who argue this point always descend into hand wavy nonsense.

mathgenius• 7 years ago

things a brain can think about that a Turing machine cannot

You say this as though we understand what "thinking" is. We do not.

YeGoblynQueenne4• 7 years ago

Machine learning doesn't "simulate a network of neurons". I think by "machine learning" you mean "artificial neural networks" but (artificial) neural networks also do not simulate a network of (I assume you mean:) biological neurons.

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...

BadThink66553212• 7 years ago

There are currently things a brain can think about that a Turing machine cannot...

Such as? Are these things forbidden by theory, or just simply beyond current engineering practice?

yters8• 7 years ago

How can such a theory be falsified?

taneq5• 7 years ago

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 (since quantum mechanics is computable and everything including the brain is ultimately quantum mechanics).

antonvs7• 7 years ago

So far we don't know of any such thing

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.

yters8• 7 years ago

If there is nothing that is uncomputable, then doesn't that mean the hypothesis is unfalsifiable?

finghin• 7 years ago

Restricting brain activity to computation it becomes more difficult to find exceptions to the argument. I think this is because computation is fundamentally quite alien to human minds, insofar as it is not an elementary, inescapable mental state. Many philosophers of mind (though not all, particularly not those who adhere to reductive physicalism and eliminative materialism) would characterise phenomenal consciousness and intentionality as these unavoidable mental properties.

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.

tcgv2• 7 years ago

True random number generation can't be done by a deterministic computer, and it appears that human brains can, although not conclusive yet, and the precise mechanism is unclear:

- https://www.ncbi.nlm.nih.gov/pubmed/15922090

omarhaneef4• 7 years ago

In a different lifetime I got to take a few classes with Jerry Fodor, and although this publication references him extensively, one of his most succinct arguments for the computational theory of mind is only alluded to: the lack of alternatives.

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

mjburgess2• 7 years ago

The problem with that argument is that it "the computational theory of mind" isn't a "theory of mind" -- so I don't think we have any.

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.

omarhaneef4• 7 years ago

You’re right in the sense that the CCTM (classical computational theory of mind) isn’t a complete theory. There are lot of problems. Studying it is essentially memorizing all the problems and the problems with various proposed solutions.

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.

bolzano2• 7 years ago

He is fantastic, I've only been able to find a couple of videos of him on Youtube. If you have any links at all, I would really really appreciate the share, even just plain old audio recordings would be awesome. Thanks!

omarhaneef4• 7 years ago

I also rely on youtube. If I knew then what I knew now I would have been more diligent in recording what he said.

YeGoblynQueenne4• 7 years ago

> The first argument emphasizes learning (Bechtel and Abrahamsen 2002: 51). A vast range of cognitive phenomena involve learning from experience. Many connectionist models are explicitly designed to model learning, through backpropagation or some other algorithm that modifies the weights between nodes. By contrast, connectionists often complain that there are no good classical models of learning [4.2 Arguments for connectionism].

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.

YeGoblynQueenne4• 7 years ago

> However, the mechanisms that connectionists usually propose for implementing memory are not plausible. Existing proposals are mainly variants upon a single idea: a recurrent neural network that allows reverberating activity to travel around a loop (Elman 1990). There are many reasons why the reverberatory loop model is hopeless as a theory of long-term memory. For example, noise in the nervous system ensures that signals would rapidly degrade in a few minutes. Implementationist connectionists have thus far offered no plausible model of read/write memory. [4.3 Systematicity and Productivity]

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.

logicallee• 7 years ago

For anyone here who thinks consciousness and all thinking aren't an emergent property of "computation" in the mind, what do you think would fail to happen if you simulated it in silicon, trying to emulate it exactly?

The Computational Theory of Mind (2015)

Related Stories