Thinking Machines and The Semantic Quagmire



What Is The Semantic Quagmire?

I use the term "semantic quagmire" to refer to a peculiar state of mind which must ultimately be reached when thinking about thought itself, and cognition generally, if one cannot avoid infinite semantic regression.  By semantic regression I mean the sort of inquiry that children make when they ask "What does word x mean?", "OK, so what does word y mean?", and so on indefinitely, trying to make sense of words by dint of pure semantics.  The final question in the chain is not "What does 'mean' mean?", because 'to mean' can only be explained semantically in terms of yet other words.  There is no final question, the chain is infinitely long, infinitely coiled and infinitely interconnected.

Too much focus on semantics leads to a dream-like state of running through glue, trying to reach an ever-receding goal.  Cognitive science and philosophy, each by a different route, currently lead the inquirer only into the semantic quagmire.  The reason is reason itself.  Reason, cognition, thinking, intelligence, all such words must be understood intuitively.  They are not things which can be seen, weighed or located in space.  Even if we agree or accept that all cognitive states are subjective, and the 'real' world objective, we are strictly circumventing the question.  The question, of course, is "What is intelligence / cognition / thinking / perception etc.?"

The root of all possible states or types of cognition would seem to be perception.  Without perception, there can be no knowledge, no observer.  No observer, whether subjective or objective.  To say that perception is subjective, and the thing perceived objective lands us back in the semantic quagmire.  But we have to start somewhere, so it is useful for each of us to assume that there is an 'I' which uses perception and an 'everything else' that is perceived.

By dividing 'everything' into these two component sets we begin to understand ourselves and our place in the universe.  It seems logical to assume that everything we know or can know comes from our inherent ability to categorise, to place into sets.  (There appears to be some support for this idea from the demonstrable fact that the whole of mathematics can be derived from set theory.)  

We can generally agree amongst ourselves about set memberships, but not to a high degree of linguistic precision.  A classic example of imprecision is colour groupings.  What is, to me, yellow, may be orange or green to another.  But the greatest area of imprecision is the set of all words with denotation or connotation of cognition and perception.  This is a prime cause of much woolly thinking about 'thinking'.

Is the self real? 

For most people, the 'I' is taken for granted.  But "I think, therefore I am." is neither good science nor good philosophy.  The 'I' is not an absolute but a relative term.  'I' is a word we are taught to use from childhood, together with 'you'.  If there were only one sentient thing in the entire universe, it could have no concept of sentient 'I' because there would be no sentient 'other' as a point of reference.  The being could thus have no concepts relating to perception or cognition.

The only way that I can accept the proposition "I think" is by external observation.  Introspection leads only to the semantic quagmire.  In my universe I observe creatures which seem to be so much like me that I group myself with them into a single set, the set of animals which think.  How do I know that they think?  They exhibit behaviours to me which I can exhibit to them from a choice of several behaviours.  We have a range of behaviours far greater than any other animal.  I call the selection of behaviours 'thinking'.  I can only understand the commonality of behaviour between myself and others by agreeing that everyone thinks.  'Thinking' is thus an eclectic process postulated to account for shared behaviours.  I can say "I think" without fear of rational contradiction because it is a well-established consensus truth.  I am reliably informed that I think, therefore I think.

What is reality?
    Having divided 'everything' into 'I' and 'everything else', I can adhere to four different theories of reality:

    1.    Everything is real.
    2.    I am real, everything else is illusory.
    3.    I am illusory, everything else is real.
    4.    Everything is illusory.

If I accept any theory other than the 1st, I am forced to accept that I can never 'really' know anything at all, in which case my quest for knowledge, indeed, my life itself, is futile.  I must, for my sanity, accept only the 1st proposition.  Which leads to the proposition: "Everything is real, therefore 'I' am real."

Acceptance that everything is real does not imply that everything is tangible.  'Solubility' is a real property of matter.  I can locate either a solvent or a solute in space and time, independent of any other material thing, but I cannot do such with 'solubility'.  Nor can I do it with 'intelligence'.  That is because those words are of a class (ability or property) which relate to potential causation, hence imply an observer.  

The obvious observer of my thinking processes is 'I'.  But as with so many things, the common-sense observation is not necessarily true.  I can never be an objective and impartial observer of my own 'I' because my perceptual experience of my 'I', and my 'I' itself co-exist.  All that I can do as an objective and impartial observer is to deduce from the behaviour of others that they each have an 'I' substantially like mine.  In like manner, others can observe my behaviour, as for example my writing of this document, and deduce that there is an equivalent  'I' at work like them.  That is the only way in which a sentient being can ever 'know' that another is sentient.

The other minds problem.

Even if telepathy were a given fact, the mere reception of signals in a different form is not proof of thought.  It is only proof of mental states that come to be shared in both  transmitter and receiver from whatever cause, presumably thought.  Since the mental states are shared, there is no independent and objective observer.  It follows that we can never know that any sentient thing or being is or is not 'really' thinking.

This leads to the question "Could a machine think?".  There can be no 'real' answer to that question.  But I would say, as with human thought or intelligence, that it might be possible to establish by consensus truth whether or not a particular machine was 'really' thinking.  By proposition 1, above, everything is real.  It follows that the process I deduce as cause of behaviour in others, and they deduce of me, is real.  

If a machine could exhibit anonymous remote behaviours such as writing a document, such that most people would not question the source of such behaviours, then it would be appropriate to infer intelligence in that machine.  Suppose you were told and believed that the document you are now reading was written by a computer in response to the command "Write an essay on intelligence".  Would it be rational to say "Well, I thought the writer was intelligent, but I must withdraw that, now that I know it was compiled by a computer"?

The Turing Test (TT), in whatever specific form, has to be a valid test for computer intelligence if we are to remain rational.  It is not, however, the best test.  A better test than TT would be TTR, the reflexive Turing Test.  This test requires of an intelligence the ability to judge a Turing Test and to determine whether the other party is human or machine.

I now come to three ideas about intelligence which I propose to refute.  These are:

        The Illusion Argument (IA)    ................    John Gray and others.
        The Chinese Room Argument (CRA)  ..      John Searle.
        The Syntax Argument (SA)    ................       "    "

These arguments must be refuted because their acceptance would be 'proof' that a machine, whether mechanical, electrical or biological, could never think.

The illusion argument purports to show that the 'I', the identity of self is illusory. If it is accepted that, as stated above, we cannot have thought without perception, then it follows that since an illusion can only occur in thought, it must be accompanied by perception.  If my sense of self, my 'I' is illusory, then it follows that there must be an observer of that illusion.  I do not subscribe to humunculus theories, so I deduce that either my 'I' is real or 'I' am observing an illusion.  Alternatively, my own 'I' is an illusion observed by others who, being observers, are real.  The illusion argument is subject to infinite regressions into the semantic quagmire which can best be avoided by choosing the only other option: 'I' am real.

The Chinese Room Argument, reduced to basics, 'proves' by reductio ad absurdum that a machine could never 'really' think.  In essence, Searle's argument is this:

Given:

    1.    A black box (room) can pass a Turing Test in Chinese.
    2.    In the black box is nothing but a book of rules in English and John Searle.
    3.    John Searle does not speak, read or understand Chinese.

it follows:

Neither in Searle, nor the book, nor the fabric of the room is there any intelligence which understands Chinese.

hence:

The Chinese Room does not 'really' understand Chinese.


There is an apparent paradox.  Obviously it cannot be true that the room both does and does not understand Chinese, both is and is not intelligent.  But a paradox can only arise from a set of axioms which are only prima facie true.  Once the deeper truth is discovered, the paradox is shown to be a fallacy.

We are forced to accept axiom 1, else the whole discussion of machine intelligence is futile.  Axiom 2 must be accepted: a Turing Test capable machine must have at least a program, a means to execute that program, and data to manipulate with that program.  This leaves only axiom 3 to be examined.

We can begin to resolve the paradox by analysing axiom 3 into its components.  We know intuitively what it means to be able to speak or read a language.  But what does it mean to 'understand' a language?  Does it imply a full knowledge of the syntax, semantics and pragmatics of the language?  Does it imply an ability to 'think' in that language.  If John Searle became however proficient at his program executing task, he could never learn to 'understand Chinese' in either of these two senses.

Does the CRA hold, then?  I suggest it does not.  Understanding is not a single entity.  It has various levels.  My dog understands various command words.  But I am entirely confident that he does not 'think in English', nor does he have any conscious knowledge of semantics, syntax or pragmatics.  Yet my use of English causes him to exhibit various behaviours, and this implies understanding or intelligence.

The CRA exhibits behaviours in response to an input of Chinese language.  It must, then have a source of those behaviours.  The only thing in the CRA which is capable of behaving is Searle.  The fact that he is inherently incapable of reporting his gross behaviour in terms of understanding Chinese does not mean that he has no understanding whatsoever of Chinese.  The paradox disappears once it is proven that Searle (or any mechanism performing the same function ) understands some essential components of Chinese.

In the CRA, the book of rules is in English.  This is irrelevant.  The rules could equally be in binary code, DNA or any other 'language'.  The fact that Searle understands English is proof that he has at the very least some rudimentary linguistic ability.  It can be shown that he, as with every human and every digital computer has the components of general linguistic ability.

At the most primitive level of all cognition lies pattern recognition.  The simplest pattern is represented by an object which can maintain one of two mutually exclusive states A or B and which can be made to switch states.  Whether we view state A as true and state B as false or vice versa makes no ultimate difference.  A digital computer, even a primitive form of Turing Machine, must be able to execute the simplest possible instruction: determine the binary state of object x.

For any machine to determine the state of even such a simple thing as a binary object requires perception.  But as stated above, perception is necessary to thought.  Hence it is reasonable to accept that the ability to recognise alternative binary conditions is the most primitive possible element of intelligence.  For the sake of simplicity, let us call this simple unit of intelligence a BInary Percept, or BIP.  We now come to the fact that in the Chinese Room, Searle is matching input symbols to symbols in the rule book.  This requires perception of points of correspondence between symbol and symbol.  The more complex the pattern to be matched, the greater the number of BIPs per operation to be performed.

As soon as we come to deal with a sequenced set of symbols (syntax) we have a time-dimensioned array of BIPs, hence for each phrase or sentence there is an intelligence level measured as BIPs per character and BIPs per chain of characters.  The more complex the sentence, the greater the number of BIPs needed to recognise the sentence.  'Understanding' a sentence can be viewed as the perception (semantics) that a particular sequence of symbols (syntax) has a correspondence with a particular state of reality.  Finally, there must be perception that of various possible results of the application of production rules, one output behaviour is the most apt in all the circumstances (pragmatics).

Accepted that the smallest unit of intelligence is the BIP, the proper question to ask of the Chinese Room is not "Where is the intelligence?" but "Are BIPs essential to the operation of the room?  If so, where are the BIP operations performed?".  Manifestly, since the room 'understands' Chinese there must be BIPs in the room, and these can only be due to Searle.  So the intelligence of the room is in Searle. By implication, the intelligence would be in any machine that might replace Searle.  Searle's lack of awareness that he is the source of the intelligence is not proof that the Chinese Room as a whole unit must be entirely lacking in intelligence.

The computing brain.

In an address to the APA: "Is the brain a digital Computer¹", John Searle stated that there is a well-defined research question: "Are the computational procedures by which the brain processes information the same as the procedures by which computers process the same information?".  In discussing that question, he makes the following major claims:

1.    Syntax and symbols are not defined in terms of physics, they are observer-relative.
    
2.    Syntax is not intrinsic to physics.

[it follows:]

The brain, as far as its intrinsic operations are concerned, does no information processing.

The 1st proposition leads to an absurd conclusion.  The only way in which we can perceive or understand the universe is through symbols.  If symbols do not exist in physical reality and are observer relative, then for each symbol found to exist there must be an observer.  Ultimately, that observer must be an intelligence, an 'I'.

According to physics and biology, an object in the real world is illuminated by light.  Some of this light falls on my retina and forms an image.  The image is not the object.  It is a token, image or symbol of a real object.  If Searle is correct, the symbol only exists because there is an 'I' to interpret it.  That implies that in a dumb camera, the pattern of light cannot be called an image except in terms of observation / interpretation by an intelligence, presumably human.

A proper question now is "How does an intelligence assign meaning to a symbol?".  There can be only one answer: behaviour.  I can recognise a no-smoking sign and refrain from smoking, or perhaps not.  An appropriate symbol can trigger any sort of behaviour.  A melancholy tune can trigger tears.  Symbols trigger behaviour.  

Imagine a camera set up to start a recording device if it 'sees' a particular symbol, say a human face.  It exhibits behaviour in response to a symbol.  If a symbol can only have meaning with reference to a sentient observer, then automatic machines would not work.  To agree with Searle's 1st proposition, above, in an implied context where only a human can 'observe' is to agree that automatic symbol sensing machines do not 'really' work.

Given the truth of the 'bare-bones' proposition that symbols require intelligent observers, and the truth that symbol detection machines exist, it follows that a symbol detection machine cannot operate without at least some minimal form of intelligence.  Just as there is a smallest possible unit of information, the BIT, so there is a smallest possible unit of perception, the BIP (see above).  But any quantity of BIPs does not, indeed cannot, alone exhibit intelligence.

Intelligence pertains to the detection of and reaction to information.  This implies that the smallest unit of  intelligence, displayed by the simplest possible intelligence implementing machine, requires a simplest symbol, the binary digit, or BIT, a simplest perception, the binary percept, or BIP and a simplest reaction to the perception.  Behaviour requires an effector.  

Suppose a machine which can recognise an absence of light and that it can accordingly switch on a lamp.  In binary terms this is an invertor.  It can detect '1' and output '0' or vice-versa.  This is the simplest possible effector, as is any two-state device.  This gives us a smallest possible unit of behaviour, a change of binary state triggered by a perception of a binary state.  Effectors have effects, so it may be well to use the abreviation 'EFF' to signify the least possible unit of effect.  

Even if we accept the foregoing, that intelligence consists of BIPs and EFFs, it is counter-intuitive to accept that human-like intelligence is just a matter of scale.  Binary operations can be parallel or serial, or a mixture of both.  Incidentally, only a basic Turing machine is purely serial.  In all real digital computers there are registers which can hold a number of digits.  An arithmetic operation  on the contents of a register is serial, because of the 'carry' requirement.  But logic operations can be parallel.  For example, the 'XOR 1' operation can invert all digits at once.  Such operations have no inherent sequence.

All serial operations, and program execution itself, imply syntax.  According to Searle, proposition 2, above, syntax is not intrinsic to physics.  "The ascription of syntactical properties is always relative to an agent or observer who treats certain physical phenomena as syntactical."  If this proposition is true, then it has grave implications for cognitive science generally.  Any non-trivial serial program, even as simple as two instructions, is inherently syntactical.  If syntax is observer dependant, then computers, being syntactical objects, can never 'really' think.  And humans, being presumably non-syntactical, cannot 'really' be machines.

Is syntax real?

Of course, if syntax is observer dependent, then the argument against machine intelligence, and impliedly for a special place in the universe for humans must stand.  So, is it true that syntax has no physical reality?  Firstly, what do we mean by syntax?  Of course, in language it means word order.  In computing, as applied to a Turing machine, it means program execution order.  Program execution depends on a mechanism which can in some way detect that a condition is satisfied and then act upon that condition.  The question of physical syntax is thus: "Is there any physical mechanism, independent of an observer, that detects a condition and acts upon it in a selective manner?"

In logic terms we are looking for a natural phenomenon which exhibits non-trivial sequential behaviour and is observer independant.  Trivially, any cause - effect relationship is sequential, hence syntactical.  But I want to give a complex example to prove the point more thoroughly, that there are in fact syntactical sequences in nature.  Any sequence A - B - C - D which leads to outcome X is syntactical if a different sequence would not produce the same outcome.

There is an obvious example in the real physical world: DNA.  The operation of DNA is syntactical.  Protein synthesis depends on the sequence of operations.  Change the sequence and the protein is different.  The process is observer independent, else all processes are observer dependent, and a falling tree makes no sound if there is no-one there to hear it!

It could be said that DNA, being an essential part of living things, has some inherent intelligence.  That cannot be true, hence a gene cannot be selfish.  Certainly, DNA can be viewed as a language.  There is a one-to-one correspondence between DNA sequences as words with proteins as objects, so there is semantic content.  The sequence of bases is crucial to proper function of a cell, so there is syntax.  The purpose of DNA is to allow a cell to 'select' for production, particular proteins at any one time out of the many different proteins that a cell might be capable of producing.  There is an eclectic or pragmatic process.

Syntax + semantics + pragmatics =  language.  DNA can well be said to be a language.  But a language is not an intelligence.  Intelligence can only arise when perception of a state causes an effector to act.  In the living cell, DNA provides the information only.  The effectors are not in the DNA.  It is the cell that expresses intelligence through its behaviour.  So is the cell the observer of DNA syntax?  Is Searle right?

Assume that Searle is right.  An observer is essential to syntax.  In the example of a living cell, the cell can function syntactically without needing the presence of any external observer, so the cell must be the observer.  But there is no 'mind' postulated in the cell example, and the model is purely mechanistic.  To argue that the cell itself is the observer of DNA syntax is to argue that any mechanism which can act in a like manner is an observer.  Turning this around, wherever there is a syntactical mechanism which expresses perception of syntax through behaviour, there is an observer.  The fact of an inherent observer is implementation independent.

Searle's use of the term 'observer' seems meant to imply 'intelligent observer'.  The obvious conclusion is that any mechanism capable of expressing behaviour in response to syntactical perception must be intelligent.  It is not a matter of 'either - or'.  It is a matter of degree.  

Asking the right question.

To ask if the brain is a computer or if a computer is a brain is to ask the wrong question.  The right question is: "Is there a class of mechanism which includes living brains and non-living 'brains'?"  The answer is: "Yes, the class 'intelligent things' whose members are capable of perception of physical states and are capable of expressing behaviour in response to those states."  

If a black box could respond to any natural language input sentence by expressing the behaviour impliedly required of it by that input sentence, then it would be impossible to determine whether the black box contained a living brain or a computer.  But it would contain an intelligent thing, and the intelligence would be as real as you and I!

Reference:
[1] John R. Searle Presidential Address to the APA.

Further reading:
The Chinese Room Argument
The Turing Test
I Think, But Who Am I?

Related Articles in my blog:
Intelligence Made Simple
A Journey to the Centre of the Universe
Digging Beneath the Surface of Grammar