J. J. Sparkes
Pattern Recognition and Scientific Progress
Mind, New Series,
Vol. 81, No. 321
Jan. 1972, 29-41.
This paper falls into two quite distinct parts. In the first part I am concerned to show that there is nothing peculiarly scientific about scientific theorising. In the second part I deal with the question which necessarily arises out of the first, that if the logic of scientific discovery is not especially scientific then what is it about science that distinguishes it from other branches of knowledge and understanding?
I shall take it as no longer a matter of dispute that science progresses by a series of hypothetico-deductive steps at various levels, from the deepest and most general, such as that which led to Relativity, to the relatively trivial, such as that a T.V. picture is noisy because of a neighbour’s electric motor. The question I want to discuss further is how rival theories, each of which may be imperfect, are judged relative to each other in the light of experimental evidence relating to them.
I accept first, that as Popper  pointed out, it is not possible to prove a theory true; second, as has been forcibly argued by various authors recently (notably Peetz,  Lakatos,  Kuhn,  Swinburne,  Agassi,   Polanyi,  and Braithwaite ) that, contrary to Popper’s view, it is not in principle possible to prove a theory false either - that indeed, to quote Peetz, “It looks as if the process of refuting a scientific hypothesis is too complex a matter to allow a purely formal criterion.” 
1. K. R. Popper, The Logic of Scientific Discovery, London, 1959.
2. D. W. Peetz, “Falsification in Science”, Aristotelian Society Proceedings, 1969.
3. I. Lakatos, “Criticism and Methodology in Scientific Research Programmes ”, Aristotelian Society Proceedings, 1969.
4. T. S. Kuhn, The Structure of Scientific Revolutions, University of Chicago Press, 1962.
5. R. G. Swinburne, “Falsifiability of Scientific Theories ”, Mind, July 1964.
6. J~ Agassi, “Sensationalism”, Mind, N.S. 75, 1966.
7. J. J. Sparkes, “Scientific Method”, Bulletin of the Institute of Physics, November 1962.
8. M. Polanyi, Personal Knowledge, Routledge & Kegan Paul, London, 1958.
9. R. B. Braithwaite, Scientific Explanation, New York, 1960, p. 19.
10. D. W. Peetz, ibid. p.29.
The various arguments in support of Peetz’s conclusion may be summarised as follows.
Lakatos is impressed by the fact that experimental results have to be interpreted by what he calls “touchstone theories” before they can be seen as providing falsifying evidence for the hypothesis under examination. But how do we decide to cling to the touchstone theories, for perhaps it is they that are at fault? Did Galileo observe the planets of Jupiter? He relied, according to Lakatos, on a “virtually non-existent optical theory” to interpret his results, so he might well have been mistaken. Actually Lakatos oversimplifies here. Galileo needed no theory to support him; he merely had to turn his telescope on a nearby steeple to demonstrate beyond reasonable doubt that he had nothing more than a magnifier.
Peetz himself begins from the total incompatibility of the wave and corpuscular theories of light; (experiments can readily falsify either, yet both survive) and moves to a general concern with the extreme difficulty of finding criteria for choosing one imperfect hypothesis in comparison with another.
Swinburne, Polanyi and Sparkes all give examples of falsifying experiments which have not succeeded in falsifying theories which incorrectly predicted the results of measurements. Sparkes goes on to list the various strategies the scientist may adopt to reconcile his theory with new results. He may reject the evidence, or introduce new concepts which, within the given theoretical framework, predict the new result. He may introduce new ad hoc hypotheses, quite apart from modifying the structure of the theory itself.
Kuhn, I think, sees a staircase-like evolution of paradigms. He emphasises the difference between “normal science”, when theories are more thoroughly articulated, and “crisis science”, when paradigms compete and advances are made. But he says in his preface that this distinction is “much too schematic”, and I suspect he would not quarrel with a blurring of the edges of his crisis situations and would agree that modifications occur to theories and hypotheses from time to time even during near-normal science. For the moment, however, his point is that the modifications bear no clear relationship to experimental data.
Polanyi deploys both the argument that evidence is not in practice always significant in rejecting or even modifying a theory, since it is often ignored in the hope that it will subsequently be shown to be mistaken, as well as the argument ascribed above to Lakatos. He says, for example, “within two different conceptual
frameworks the same range of experience takes the shape of different evidence.”
Braithwaite puts a similar argument in a different way. He notes that experimental results can be explained in more than one way, using various “high level” hypotheses. In general the data can neither confirm nor reject any particular approach.
These arguments seem to me to be incontrovertible and I therefore take it as no longer in doubt that the relationship between observational data and hypothetical or theoretical explanations is complex; that clear-cut confirmation or falsification of a theory by experiment or observation is not possible. The question, then, of how we do decide between rival theories and relate them to experimental data becomes a much more difficult one to answer.
Lakatos addresses himself to this problem and clearly rejects the view, which he unfairly attributes to Kuhn, that “scientific change - from one paradigm to another - is a mystical conversion which is not and cannot be governed by the rules of reason.” He then proceeds to adduce rational criteria for planning and conducting research programmes. The position I wish to argue in this paper is that, whilst it is clear that the idea of mystical conversion must remain unacceptable if argument is to continue, it does not follow that we must be able to delineate the rational criteria we are to use in bringing about scientific change. Briefly, my view is that we use our somewhat innate capabilities of pattern-recognition to find our way through the maze of facts and hypotheses. Unfortunately, although much research is being done on the problem of pattern recognition, the rationale of the process is by no means clear yet. However, I agree with Suppes’  view “(That) a convergence of effort on the most difficult cognitive problems, those of perception and concept formation, has been building up at least since 1960... Many scientists working on these problems feel we are getting close to hitting on the one or two fundamental ideas needed to move rapidly forward.” Indeed, we are moving towards a crisis situation in this particular branch of science.
In the first part of this paper I wish to show that the acceptance of a paradigm in science is very similar to the recognition of “patterns”, an activity we perform every day of our lives.
1. P. Suppes, “Information Processing and Choice of Behaviour “, in Philosophy of Science, A. Musgrave and I. Lakatos (Eds.). North Holland Publishing Co., Amsterdam, 1968, pp. 298-299.
Here I am using the word “pattern” to describe any distinguishable interrelationship of data. The recognition of a face, a chair, a printed or handwritten word, a locality, etc., are examples of pattern recognition. So, too, is the recognition of a tune, or a theme in a symphony or words in speech. The remarkable thing is that we are able to recognise a chair, for example, from any direction, at any orientation, at any distance, whatever its colour and so on. The concept “chair”, which seeing one stimulates, seems to bear no clear, or as yet rational, relationship with the actual data entering our eyes. Furthermore we can perceive patterns even when they are grossly distorted or severely affected by ‘noise’. (Here I use ‘noise’ in its technical sense. Random visual interference on a television screen is noise, just as much as hissing and crackling on a telephone or the clattering of cutlery in a restaurant.) It is this capability which I believe we use in deciding upon and accepting a particular hypothesis in science. Consider a simple example.
Suppose we are presented with a printed character Ć [HHC: figure intentionally distorted in original] we might hypothesise that it is a new type of character or that it is an imperfect H, or an imperfect A. How do we decide which? We might ask where it came from, and on being told it was typed by a cheap English typewriter, we might assume a Roman font and conclude it must be an A, since an H distorted thus would be most improbable. But we could be wrong. Alternatively we could try to find examples of the symbol in context. Thus if we found, in an English text, ĆAT or TĆE we could conclude it was an H. But if we found HĆT or CĆR we would probably conclude it was an A. Context here, if we know we are dealing with English, is crucial.
Suppose now we play a slightly different game and try to decide whether a set of printed words made of strange characters such as ǚ, Ń, Ô, Đ [HHC: figures intentionally distorted in original] constitute a page of English. We might, for example, hypothesise that they are distorted letters intended to be A, N, O, D, or perhaps H. Y. C. G. If one set of interpretations of the distorted patterns leads to the build-up of an English-like sentence, but including a few mistakes, then anyone knowing English might settle for the conclusion that this indeed was the correct solution, especially if he had a reason to expect to find English. But a different set of interpretations could equally lead, say, to an imperfect Latin sentence. Thus it is quite possible for two sets of hypotheses about what the letters stand for to
lead to two further hypotheses about what the sentences are, from which it should fairly clearly follow what the language is. But the decisions are not clear cut and the hypotheses are plainly hierarchical.
Now this pseudo-problem is quite typical in character both of scientific problems and of the pattern recognition problems which human beings, and many less intelligent animals, encounter and solve every day of their lives. We are continually having to recognise people, objects, handwriting, etc., using distorted or unfamiliar and inconclusive data. Often, possibly almost always, it is the context of the events which enable us to solve the problem.
Similarly, we have to recognise the speech of people we may not have heard before, saying sentences we have never heard before; yet usually we understand what we hear. Somehow we find sufficient of a recognisable pattern in our perceptual input for us to identify the incoming signal in terms of what we know and have learnt, so that it can then impart some new knowledge or intelligence to us. In just the same way we interpret scientific data using any relevant touchstone theories we can bring to bear on them, and use them in order to confirm or deny or modify another scientific hypothesis. Thus my thesis is that the mental processes which lead us to select theories and advance science are the same as those which enable us to acquire knowledge, or form concepts at all times. The procedure adopted by human beings when they recognize patterns, such as speech waveforms, is not yet known, but it is possible to conclude from studies of human behaviour some minimum essentials without which we could not do what we do do. I will consider two such features of the procedure, and will discuss them in connection with speech recognition. Since patterns of sound are spread out in time the operations are necessarily sequential, and seem easier to think about than are patterns distributed spatially.
First it appears to be a fact that it is just not possible in general to identify simple words spoken by an unknown speaker without knowing some of the characteristics of his speech. But, on the other hand, in order to discover the characteristics of his speech we have to know what he is trying to say. If we don’t know he is trying to say ‘make’ rather than ‘mike’ we cannot discover that he is a cockney. Where do we start? It is in fact like trying to solve two simultaneous equations. If you know the appropriate algebra it is easy otherwise you guess the value of one unknown and find a consistent solution for the other - -a simple iterative problem.
In a practical situation, when we meet someone new, if he does not say something conventional like “Hello” or “How do you do” which we can easily recognise and use to identify his speech characteristics, we have to use this iterative technique. We attempt to recognise his words and, as soon as possible, put up hypotheses both of what he is saying and how he is saying it. We then analyse his speech signal using these hypotheses, improving them continuously as more “experimental” data flow in.  
Thus the putting up of hypotheses and clinging to them whilst trying to force the data to fit is a typical perceptual strategy, just as it is a typical scientific activity. It is well known that we tend to see or hear what we expect, and it is also a common experience that we can completely misinterpret sights and sounds when we are prepared for something else. Only context or a reprocessing of the signal can put things right again.
The second point I wish to make about pattern recognition is concerned with the part played by memory and learning in the recognition process. For the present purpose it is necessary to distinguish two aspects of memory. First the part which holds the information we can recall, a face, a poem, our vocabulary, theorems, algebra etc. Second the information which we have learnt and which presumably is stored somewhere because we make use of it, but which we cannot retrieve; We cannot describe how to recognise speech, how to balance when we walk, how to turn an idea into sentences expressing that idea, how to drive a car, or any other aspects of pattern processing, although the information must be stored in our brains somewhere.
Thus, in computer terms, one form of memory can be addressed, and its contents recalled, the other form cannot.
The present significance of this is that rational aspects of thought and knowledge are in the addressable memory. We can describe the steps in a logical or mathematical argument because we can recall them from memory - they have been memorised. We cannot recall the steps of how we process a speech signal in order to recognize a word or phrase, any more than we can recall exactly how we came to formulate the theory that, for example, comprehensive education is best. We can all describe aspects of the evidence but our decisions are based on irrational weightings of this evidence.
1. For a lucid introductory exposition of speech processing see: P. Denes, and E. Pinson, The Speech Chain, Bell Telephone Labs., 1963.
2. See J. J. Sparkes, “Pattern Recognition and a Model of the Brain “, International Journal of Man-Machine Studies, 1, no. 3, July 1969.
Thus Lakatos’ determination to find rational criteria with which to guide scientific progress, though laudable, since it would tend to reduce arguments, is itself irrational, since many of our most important mental processes, particularly those most akin to scientific theorising, are not rational, or even, as yet, knowable.
In summary then, let me be quite specific about this analogy between pattern recognition and scientific progress.
In speech recognition we receive an acoustic signal which our ears analyse into acoustic elements in a particular way. In science we have an environment from which we extract data in a quite specific way (to be discussed in part II).
In order to achieve the recognition of speech we have to have learnt how to split up the coded sequence of acoustic elements which passes from our ears to our brains into various phonetic or linguistic features. We have to interpret variations from what we expect either as new information or else as peculiarities of the speaker, whichever seems appropriate. In science we have to arrange and select our data to articulate the touchstone hypotheses we have read about or discovered, and thus augment or comment on the more fundamental theories we are considering. Or, if it is more appropriate, we use the data to revise our touchstone hypotheses, and thus obtain quite different information regarding the fundamental theories.
In speech if we recognise a sentence, say, but mishear a word, we use the sentence to interpret the word. In science if a mass of data seems to bear crucially upon a fundamental theory we may conclude that if some of our data do not fit properly they must be wrong - we begin to look for a revised, low-level hypothesis capable of interpreting this rogue data satisfactorily.
The whole of speech, whether its recognition or its synthesis, is constrained by our rules of grammar, of phonetics, of semantics. It is extraordinarily difficult to invent a new language other than a re-ordering of existing ones. In science it is extraordinarily difficult for most of us to look at data from a new point of view. The Newtons and Einsteins are rightly distinguished for their achievements, even though nowadays their viewpoint seems somehow inevitable. Similarly, if someone speaks only one language it is difficult for him to imagine what it would be like to speak a second one. But if he speaks two it is difficult for him to see what the difficulty is!
The only real difference between pattern recognition and scientific theorising is that speech or visual recognition takes place naturally on a time scale of milliseconds, whilst with scientific data we have to deliberate and it takes much longer.
Thus my thesis is that the major problem at present confronting us in scientific methodology, leaving aside creativity, namely how we choose between various hypotheses and data and come up with an “improved” scientific edifice, is just another aspect of a fundamental human capability, that of pattern recognition and processing. Indeed I would go further and maintain that, since pattern processing is so characteristic a human activity, then if the logic of scientific discovery is something different from it, we need to find an explanation for this difference.
But this thesis entails two further problems. Firstly, since we do not yet know how pattern recognition is performed, my thesis does not solve the problem of how science advances; it merely proposes a “progressive problem shift” to use Lakatos’ expression, away from the limited study of science towards a general study of how brains deal with complex interrelated information. Second, since I am now saying that there is nothing peculiarly scientific about this part of the scientific method, I must face the question of what it is about science which distinguishes it from other areas of knowledge and understanding. For certainly the adjective ‘scientific’ means something, even if we do not know quite what. This is one of the tasks of the second part of this paper which now follows.
I want now to discuss what it is that distinguishes scientific method from other branches of learning so that we can understand its remarkable power, and so that particularly we can see to what extent rationality is an essential part of science.
If we understood the purpose of science we might find it easier to isolate the methods by which it reaches out for its objective. This is hardly a new problem to discuss, but that it has a new implication and that various answers have been given to it renders it non-trivial.
For Ziman,  for instance, “the goal of science is a consensus of rational opinion over the widest possible field.” For Hagstrom  the goal is conceived in psychological terms; a scientist’s aim is to achieve social recognition by providing to the community
1. J.M. Ziman, Public Knowledge, Cambridge University Press, Cambridge, 1968. p. 9.
2. W. 0. Hagstrom, The Scientific Community, Basic Books, London, 1965.
information which it values. Neither viewpoint seems adequate, for some scientists certainly defy the consensus of scientific opinion, whilst others seem to have no desire to communicate their results and ideas.
That science seeks out the truth is a very common formulation of its purpose. It suffers, as is well known, from the fact that we have no way of knowing whether our theories are true or are merely successful. Consider the conflict between the scientific view of the world and that of common sense which, though it is with us all the time, is merely a matter for mild amusement nowadays, for our surrender to science as regards the way we describe our perceptions is almost total. Most of us know about optical illusions and agree that parallel lines when suitably cross-hatched appear to converge, or that circles can appear as spirals or that we can misjudge the lengths of lines. Furthermore, we would not normally quarrel with the kind of description which refers to parallel lines in the above illusion and to our impression being illusory. That these lines are parallel is established by scientific measurement; that they are not parallel is an observation, which enjoys a full consensus of rational observers, interpreted by “common sense”. We believe science and ascribe the error to ourselves; we call the phenomena illusions. Similarly we all know that time passes slowly if we are bored and rapidly if we are absorbed, but this too is regarded as an illusion; suitcases get heavier as we carry them, or so it seems, and so on. Who knows the extent to which our interpretations of what we observe have been influenced by different, non-natural, orderings of our perceptual impressions. Thus qua scientists we are not in a position to claim that we discover the truth; if we claim that we do, are we stating a personal belief for which there can be no evidence, only argument. However, it is important to enquire why it is that science does seem to be able to justify this special claim to truth and objectivity, for by so doing the pattern of scientific achievement and progress becomes clearer.
There are two facts we should recall about scientific evidence and scientific theories, which, although they are not peculiar to science, are characteristic of it. These are:
First, scientific evidence is only admissible if all conceivable precautions have been taken to prevent the judgement and fallibility of the observer from affecting the results. We are all quite capable of estimating time or weight or height or colour or temperature, but our estimates are so variable that they are far too inaccurate to test quantitative theories. Instead, as
Eddington  emphasised, all measurements in the exact sciences are reduced to pointer readings or their equivalent. The only pattern which an observer is allowed to incorporate as evidence for his judgernent is of the coincidence of a pointer with a scale or ruler calibration, or the coincidence of two events in time, or of the matching of two frequencies or hues. Even better, nowadays we use instruments to do our observations for us. We are also “permitted” to count events, but we have much more trouble here since we can argue about classifications of events. The purpose of this, as Ziman emphasised, is to obtain a consensus of rational opinion, for a “scientifically” acquired result of a measurement can in principle be repeated by anyone and is therefore very difficult to reject. In practice, however, some measurements, such as the position of planets etc., are inevitably unrepeatable!
In the physical sciences an additional part of the design of a proper experiment involves sufficient and diligent control of any unwanted influences, for example, temperature variations, unwanted static electricity, etc. In the biological and social sciences control of some of the influences is impossible, so statistics and “controls” are used instead. Indeed, by the proper application of statistics it is hoped that not only the unwanted external influences but also the variability of the observer will be removed. In practice, however, few physical scientists seem to believe that they are removed, for statistics and controls are no substitute for a properly designed experiment unless the influences which have to be eliminated, and their effects, are linear and the ensemble is truly random. In practice the quantitative effects of various influences are not normally known and it is not possible to obtain random ensembles, so these techniques are not strictly justifiable. But no better method has yet been worked out.
The first point then is that every effort is made in science to ensure that the results around which theories revolve, albeit often very elliptically, are as near as possible independent of the observer.
The second point is that as far as possible general and fundamental theories, especially those of the physical sciences, are normally presented in a time independent manner, so that the actual period of time occupied by a set of events does not affect the validity of theories applied to them. The reason for this is of course quite clear. One of the primary functions of scientific theory is to predict future events, or explain past ones, and,
1 A. S. Eddington, The Nature of the Physical World, Cambridge University Press, Cambridge, 1928, pp. 251.
since science is not clairvoyant, it can only achieve prediction by making statements which can be expected to be valid in the past, the present and the future.
The hypothesis that, for example, a bridge collapsed as a result of a sudden, localised increase in gravitational force is not acceptable scientifically; indeed I doubt if even a non-scientist would have the temerity to suggest it.
One interesting exception to this general time-independence of theories is the theory that the universe was created with a “big bang” 10,000 million years ago. That this is not rejected out of hand as unscientific is itself quite surprising. Why allow a vast quantity of hydrogen to be created a long time ago, when all the universe might have been created yesterday together with us all, all our memories, records and undiscovered facts? Since both creations are scientifically impossible - they violate the law of conservation of energy - it is a moot point which is the more plausible. That we accept the former and not the latter illustrates the strength of our conviction that science discovers the truth about the world - it is too much to accept that it was all planted there yesterday!
Thus it is clear that the very properties which all of us would concede must characterise the ‘real’ world, namely independence of who observes it and when or where he does so, are built into scientific results and theories by its method. So we cannot know if it tells us the truth about the world, or just seems to.
But if we cannot be sure that science helps us make contact with reality, we are simply left with the aim of gaining confidence, through consensus of rational opinion, in its predictions and explanations. In the end the whole edifice of science seems to serve our perhaps irrational desires for confidence, understanding and predictive power. Its aim therefore is not unlike that of many other human activities, it’s just that scientists select their data in a peculiarly irrefutable manner.
Thus there emerges a not unfamiliar picture of a science, whose objective is prediction and explanation in terms of theories and concepts arrived at by hypothetico-deductive procedures. All the aspects of science are turned to this end - controlled experiments, pointer readings, the creation of a consensus of rational opinion, the use of mathematics and reason in the deductive part of the process and so on. These, in my view, and not the way scientists handle the interrelationships of hypotheses and evidence, are the features of science which distinguish it from other types of human thought.
Now the purpose of retracing these arguments is to arrive at a
decision on how important is the use of rational criteria, or the rules of reason, in giving precision to the logic of scientific discovery. Lakatos, is not alone in taking it for granted that the rules of reason must be used. This is one of his “touchstone theories”, even though it is evident that what is thought to be regarded as rational varies as time passes. The question remains, then, how we decide between these two views of science - that it is for prediction or that it must be rational - when they come into conflict. Obviously it is irrational to insist on rational criteria to decide this issue! Fortunately there are enough examples of irrationality in science to convince me, in what I believe to be the usual kind of way, that the rules of reason take second place.
Once it was thought irrational to conceive of unaided movement of heavy bodies, or to believe in heliocentric planetary motion. Nowadays both seem wholly reasonable. Today the idea that electromagnetic radiation is both corpuscular and undulatory is the outstanding monument to irrationality in physics, though the idea that massive bodies can exert forces on each other through intervening empty space, not even containing an aether, comes a close second. Both ideas appear to contain genuine internal conflicts, but few of us now worry about it. We put up with the irrationality, since both ideas have great predictive and explanatory power, and perhaps hope that it all will eventually be resolved rationally.
This is precisely the point of view I am advocating with regard to the logic of scientific discovery. It is important to describe the way scientists actually proceed, in order to guide our future scientific work, and this has been Kuhn’s major contribution. That some of our actions appear irrational, as do at present some aspects of human pattern recognition, is relatively unimportant. In due course we will find a rationale, as we have of electromagnetic radiation, and the problem will have been solved for the time being. But resort to this kind of position is not mystical; it is, I hope, like all good scientific hypotheses, the best platform from which to take the next forward step.
The question raised in this paper is whether we should look for rational criteria, as at present understood, with which to assess the relative merits of rival imperfect hypotheses, or whether the assessment is a much more complex process. The
partial answer proposed is that it is more useful to think of the process as an aspect of pattern recognition. If the rationale of pattern recognition, as performed daily by many living creatures, were understood, this could be a full answer. But the process is not yet understood, so the suggestion is, therefore, of a paradigm change or of a progressive problem-shift towards a hypothesis of higher content. We know that simple rational criteria apply to deductive arguments but we also know that they do not apply to many aspects of induction, especially to hypothesis formation; we know that they do not apply to the formulation of a number of important scientific theories. I am now suggesting that they do not apply to the complex process by which we construct the scientific edifice. Pattern recognition is undoubtedly a deeply ingrained human capability, and that it should be used for the kind of information processing which goes on in science seems beyond reasonable doubt.
The Open University