Can Machines Think?
By M. V. WILICFS FROM time to time headlines about mechanical or electronic brains appear in the newspapers. The reference is usually to one of the large automatic calculating-machines or to some automatic-control device which includes a computing element. What is one to make of the suggestion advanced in these headlines that machines can think ? Is it to be dismissed as sensationalism or is there something behind it ? Undoubtedly there is a great readiness on the part of many people to accept the •idea that large automatic calculating-machines are some- thing more than machines. I have found people who .were quite disappointed when the action of a machine was explained to them and the mystery resolved. There is some evidence that this attitude already existed during the last century in the time of Charles Babbage (1791-1871). It is to be compared, as Sir Frederick Bartlett has pointed out to me, to the ifiterest taken in performing animals. Perhaps it is flattering to human vanity to see something which humans can do easily being done as a tour de force by an animal or by a machine.
Here the matter might well be left, were it not for the fact that the more recent automatic calculating-machines possess features which differentiate them from other automatic devices and which have the effect of enlarging enormously their sphere of application. Before discussing these features, however, it is desirable that we shobld replace the question: " Can machines think? " by one which is less vague and which does not depend to the same degree on the meaning to be attached to a word in common use. A. M. Turing, in a recent article in Mind, suggests that a suitable question is whether the machine could be made to exchange a series of questions and answers with an examiner and to deceive him into thinking that he was dealing with a man and not with a machine. The examiner would be in a different room from that containing the machine and would communicate with it by means of a teleprinter circuit. Let us be quite clear at once that no existing machine is in the least capable of practising a deception of this kind. The question to be con- sidered is whether this will always be the case.
The first interesting thing about modern automatic calculating- machines is that they are universal machines, that is, they perform a strictly limited set of basic operations—add, subtract, multiply, &c. They can perform only one of these operations- at a time, but they can be set up to perform a complicated sequence—or programme, as the technical jargon has it—of them one after another. A machine is thus logically extremely simple, and the complexity, if any, is in the programme. Provided that the basic operations form a logically complete set, a universal machine can be programmed to do anything which could be done by a specially built machine. The tendency nowadays is, there- fore, to ask whether a universal machine could be programmed to perform a particular function, rather than to ask whether it would be possible to design a special machine for the purpose. The universal machines which have been built so far have been designed for performing arithmetical calculations rather than the logical operations which would be involved, if they were to simulate human behaviour. This is not, however, a matter of fundamental importance.
An automatic calculating-machine can'be programmed in such a way that, when specified points are reached in the calculation, the action of the machine will depend on the results which have been obtained so far. Such a feature is not peculiar to large machines ; it is possessed by some commercial accounting- machines which will, for example, treat a total in one way if it turns out to represent a credit balance and in another way if it represents a debit. Another, and more revolutionary, feature of large automatic calculating-machines is that they can modify their own programmes in ways laid down by the programmer. Suppose, for.example, that the original programme calls first for 240 to be subtracted repeatedly from a given total until any *further subtractions would give a debit, and then for the number' of times a subtraction has been performed to be printed. The programmer could, if he wished, arrange for the programme to. be altered automatically at this point so as to call for the number 12 to be subtracted instead of 240, and for the cycle of sub- tractions to be repeated. Such a programme might be used to convert a number of pence into pounds, shillings and pence. From the present point of view the important thing to notice is that both the number of shillings and the number of pounds are obtained by the same sequence of operations, except that one operation is changed. But if one operation can be changed so can others, and there is no reason why the machine should not progressively modify its programme until little or nothing is left of the original programme. The nature of the modifications carried out could be made to depend on data received through the input mechanism of the machine.
It is here that believers in the possibility of mechanical " thinking " begin to talk about the ability of a machine to learn. And indeed it is easy to see how a machine could be programmed so that it appeared to learn. For example, if the machine possessed- a suitable input device it could be programmed to translate and print a message presented to it in morse code. A key to the code in the form of a list of characters with the corresponding letters would have to be stored in the machine. An incoming character would be compared with the entries in the list in turn until the character was identified. This operation would be on the average most rapidly performed if the order in which the comparison was made was such that the most frequently occurring letters were tried first. There is no reason why the machine should not be programmed to keep a record of the number of times the various letters are encountered and to adjust the order in which the key is consulted accordingly. Thus when a few sentences had been handled the machine would be able to deal with subsequent sentences more quickly. If a message in another language (in which the relative frequency of the letters was different) were-then presented to the machine, it would decipher it more slowly to begin with, speeding up when it had " learned " the letter frequency of the new language. Such\ a demonstration might be quite impressive until the secret was' explained. This example illustrates how a situation fully envisaged in advance may be provided for by the programmer. The machine would be defeated, however, if, for example, alternate sentences were in different languages, although if this possibility had been envisaged by the programmer he would have had little difficulty in allowing for it.
The construction of programmes of the above kind would not present-any difficulty to anyone with experience of programming an automatic calculating-machine. The principles involved might even be developed to allow more complicated situations to be dealt with. For example, it is not impossible that a programme could be constructed which would enable the machine to learn by trial and error to play a simple game, like noughts and crosses, although this has not yet been done. One could even with a greater stretch of the imagination conceive of a programme being made which would enable the machine to learn to play a more complicated game, such as draughts. Note that we are concerned here with programmes in which some kind of learning is involved. It is possible to programme a machine to play simple games by giving it a comprehensive set of rules'which will meet all circumstances ; there is nothing very interesting in this, since the machine is merely acting in an executive capacity and is not adding anything to the instructions it was given in the first place. An example of a machine acting in this way can be seen at the Festival of Britain Science Exhibi- tion at South Kensington, where there is a machine, known as Nimrod, which has a built-in programme for playing the game of Nim.
We now come to the much more speculative question of whether it , would in principle be possible to construct a generalised " learning " programme which would enable an operator, if he had sufficient patience, to " teach " the -machine any subject he chose. Presumably the process would take the following form. A piece of information or a question would be presented to the machine through its input mechanism. The machine would, under the control of the programme, respond by printing a group of symbols. If these were of a kind approved of by the operator he would include a statement to that effect in the next input message ; otherwise he would include a contrary statement. The machine would continually modify its programme in such a way as to obtain the maximum possible number of messages indicating approval.
The problem of constructing a learning programme of this general kind is altogether of a different order of difficulty from that of constructing simple programmes of the type described earlier. As far as the writer is aware, no detailed suggestions have yet been made as to how the problem might be tackled, and whether it will ever be solved must, therefore, remain for the present a matter of opinion. It is possible that the method, if any, finally evolved for teaching a machine will not bear any close -resemblance to that used for teaching human beings. Any great progress will have to wait until much bigger and perhaps faster machines are available, but it should be noted that these machines will only differ from the present ones in their size and not in their complexity ; they will be capable of storing large quantities of information and of performing long programmes made up of simple basic operations similar to those performed by existing machines.
The extent to which it will ever be possible to make machines simulate human behaviour can also be discussed from the metaphysical or theological points of view. Arguments in this class are most often advanced to show that a machine can in no circumstances be said to " think " and consideration of them would be outside the scope of this article. One word of warning may, however, be given in connection with arguments based on a machine's supposed lack of consciousness ; the argument, for example, that a machine could not think because it could not in any circumstances know that it was thinking. Such arguments- : like other arguments about the existence of a consciousness outside one's own—are apt to lead all too rapidly to solipsism.
Previous page
