Content and Consciousness

by Daniel C. Dennett

  This advance in outlook for deprivation experiments is just a special case of a general advance in outlook for behavioural theory provided by the evolutionary hypotheses sketched in this chapter. The difficulty with behaviourism is, tautologically, that one’s subject matter is limited to behaviour, and the difficulty with this is that behaviour does not allow itself to be divided into the right sorts of parts. The strong point of S-R behaviourism is its recognition, in vague intuitive form, that there must be something in the nature of a carrot and stick, reward and punishment, survival and extinction, if learning is to be explained. Somehow sensory ‘feedback’ must be distinguished as at least positive or negative if the creature is to make any headway at all. And, since there is nothing intrinsically positive or negative in any stimulus, there must be something like the evolutionary conflict sketched here. But so long as the evolutionary conflict is dogmatically asserted to be entirely overt, manifested in trial-and-error behaviour – and this is what is involved in the behaviourist creed that behaviour must be a function of past behaviour and stimulation – the evidence will just not support the theories proposed. Animals do not engage in enough trial-and-error behaviour to ensure the development of their behaviour along appropriate lines. Except at the lowest evolutionary levels of life, where S-R behaviourism fits the facts quite well, animals need not in every case run their behaviour all the way to painful consequences before learning that it is inappropriate. So long as it is behaviour, and not behavioural controls in the form of afferent-efferent interconnections, that is deemed to be reinforced or extinguished, one’s theory is stymied by the fact that one cannot divide behaviour into fractionally appropriate, fractionally ‘rewarded’ bits and fractionally inappropriate, fractionally ‘punished’ bits. The absence of overt reward in the form of, say, food, and the absence of overt punishment in the form of pain, in cases where the experimental animal approximates appropriate behaviour or partially completes an appropriate or inappropriate response has led to the postulation of such theoretical monsters as ‘fractional anticipatory goal responses’, ‘expectancies’ and ‘partially reinforcing stimuli’.7

  Once one recognizes the need for a carrot and stick in learning, pleasure and pain stimuli offer themselves as the obvious candidates for these roles, and at a low enough level they prove adequate to the task. But when the appropriate behaviour requires extended motion and control, pain and pleasure do not suffice as directors since their force cannot be transmitted from the whole behavioural response to its parts. How can ‘a step in the right direction’ or ‘a step in the wrong direction’ be recognized by the organism if it is not immediately rewarded or punished? A covert, internal carrot and stick must do the job, and here the impossible notions of fractionally appropriate and fractionally inappropriate motions can be replaced by the notions of afferent-efferent functional structures either compatible or incompatible with the overruling pre-wired structures. A ‘fractionally inappropriate’ structure could be discriminated and extinguished not by virtue of any overt semi-punishing stimulus but by virtue of its being blocked by internal programming. Trial and error must be required for learning, but there is no reason why it must all be external and overt, all in terms of behavioural trial and error.

  VIII GOAL-DIRECTED BEHAVIOUR

  The principles of evolution proposed to explain learning and discrimination in the brain have the capacity to produce structures that have not only a cause but also a reason for being. That is, we can say of a particular structure that the animal has it because it helps in certain specified ways to maintain the animal’s existence. It is a structure for discriminating edible from inedible material, or for finding one’s way out of danger.

  Here we must be careful to distinguish between there being a reason for the animal’s having a structure and the animal’s having a reason for having the structure. Even a human being could not normally be said to have a reason for having a particular neural structure (although a human being might, for example, have a reason for having an electrode implanted in his brain). Thus when there are reasons for the presence of neural structures, it is not the case that the animal or person has these reasons. Nor is the raison d’être of a neural structure just like the raison d’être of, say, a can-opener, for a can-opener’s existence depends on the recognition of its raison d’être by its maker, whereas no one and no thing (e.g., Nature) need recognize in any way the raison d’être of a neural structure for it to exist for the reason it does. The cash value of saying that a neural structure exists for a reason is just this: were the necessary conditions for the survival of a particular animal and the environmental circumstances in general other than they are, such that the neural structure in question would not have the role in survival it has, the structure would not exist.8 It is through observing and evaluating circumstances and behavioural efficacy that we are able to frame the raison d’être of a structure, but no such activity of observation and evaluation is a pre-requisite for the appearance of the structure on the evolutionary scene.

  Although we have found structures with a reason for being, the only control structures described have been of a stimulus-response variety, granting that no limits have been set on the mediation between stimuli and responses, and a centralist account of learning has replaced the peripheralists’ external rewards and punishments. Critics of stimulus-response theories have often contrasted response behaviour with goal-directed behaviour. Can these structures that have themselves a reason for being provide for goals, for an animal’s having a goal and hence having a reason for doing something? To answer this question we must first establish criteria for goal-directed behaviour.

  ‘Goal’ is a popular word among computer programmers. Every programme has a goal, and some have hierarchies of goals. A great deal of this ‘goal’ talk is exaggeration. When a programme is designed merely to terminate in the solution of a single problem for which a linear enumeration of steps is provided by the programmer, it is stretching a point to say that the computer – or the programme – has a goal; the programmer himself has the goal, and the programme is merely his means of achieving it. Other programmes, however, are designed to accept a variety of quite different problems, and once a ‘goal’ has been set for it (e.g., solving Problem P) it proceeds to ‘search’ for a solution with no further direction from the programmer. Here it looks much more plausible to say that the computer has been given a task to complete, so that it is the computer itself that (at least momentarily) has and is directed by a goal.

  For example, Newell and Simon, in describing their GPS (General Problem Solver) computer programme,9 claim that GPS operates in a goal-directed way. GPS consists of a number of ‘sub-routines’ which perform mathematical transformations on input expressions, together with an ordering system that allows the computer to move from sub-routine to sub-routine. Giving GPS a problem to solve involves specifying an input expression and an end-state; typically the input expression is a set of premises and the end-state is the last line of a proof to be constructed, and thus the goal of the computer’s activity is said to be embodied in the specification of end-state, and the means to the end are the series of sub-routines, which the computer tries one after another. The result of a completed sub-routine is checked by the computer against the specified end-state. Partial similarity is treated as apparent progress towards the goal and the result is saved for further transformation. Does this system of subroutines and specified end-state provide an adequate model of goal-direction?

  It is particularly interesting to note that the authors of GPS are prepared to call such activity goal-directed in spite of their acknowledgement that not much in the way of heuristics is built into their programme for changing the order in which subroutines are tried or for ruling out obviously inapplicable subroutines. The computer rather inelegantly grinds away until the end-state is achieved or it runs out of sub-routines. GPS is not very insightful about recognizing progress; any similarity, however unpromising to human observers, is treated a
s apparent progress towards the end-state, deserving further work. At first this may strike us as a serious shortcoming. One is reminded of the extremely vague but compelling notion we have of a goal, lighting the way, informing our choices, hovering and helping us decide as we pick our way towards it. There seems to be little or no direction in GPS, and we might decide to call its activity an example of goal-terminated behaviour, but not goal-directed behaviour. If we are to make this distinction, however, we must be prepared to provide a more exact description of genuine goal-direction.

  We speak of goal-directed behaviour in animals, but when we do our standards are set quite low. Prima-facie evidence for goal-directedness in animals is the production of a repertoire of alternative motor patterns until one ‘attempt’ pays off by bringing about a certain sensed environmental end-state. The more ‘random’ the successive attempts appear – especially if dead-ends are repeated – the less we are inclined to call the behaviour goal-directed, but there are no minimum standards of elegance in motor choice. We cannot even rule out as goal-directed the repetition of dead-ends, for even fully rational human beings often make repeated attempts at unlikely means to their ends. Are we to say that the prisoner who tries repeatedly to scale the unscalable prison wall is not engaged in a goal-directed activity? Assiduity and the ability to recognize that the end-state has been achieved count more heavily than insight or brilliance in execution. We do ‘expect’ intelligent animals to improve and prune their repertoire, but that is because we expect intelligent animals to be learners as well as goal-havers.

  It is possible then to make a clear distinction between genuine goal-directed behaviour on the one hand and goal-terminated behaviour coupled with the capacity to learn on the other? The vague notion of goal-direction suggests that having a goal is also having rationales for the means one attempts, having the ability to distinguish appropriate from inappropriate courses of action, but this is a misleading suggestion, for one need not have a particular goal in order to be able to decide what would be appropriate courses of action if one did have the goal. The ratiocinative capacity is separable from the having of goals, at least in man, and surely we want to maintain a distinction between well-reasoned goal-directed activity and ill-reasoned or unreasoned goal-directed activity.

  But at least, it may be argued, one must know what one’s goal is before one can even begin to bring into play the separate capacity of ratiocination, while in goal-terminated activity one can be entirely in the dark about what the end-state is until it is reached. This seems like a promising mark of difference until it is asked what the criteria are to be for knowing one’s goal. For people we generally want to say that being able to state that p is a necessary condition for knowing that p (the one trivial counter-example being the paralysed aphasic who has no means of communication). If this is held to be a necessary condition for knowing, then only human beings can know, and hence only human beings can be goal-directed. (Since there is a great difference between being incapacitated by aphasia and being, like a dumb animal, a constitutional non-speaker, animal knowledge cannot be brought in under the counterexample mentioned above.) Now perhaps this is what we want. Perhaps we want to say that only human beings exhibit true goal-directed behaviour, and the capacity for goal-direction, like the capacity for erudition or eloquence, is reserved for language users only. This is unconvincing. Nothing in our vague notion of goal-direction suggests that the use of language is a prerequisite, and we do not, I should think, want to rule out as goal-directed the more remarkable activities of higher animals and very young children. If so, then we must find a different sense of knowledge, one that does not require that what is known be statable by the knower. We must find some other behavioural criteria for knowing, and, more particularly, behavioural criteria for knowing one’s goal.

  What other behavioural cues for knowing one’s goal could there be but taking steps towards achieving the goal and stopping when the goal is achieved? If these are the only relevant cues we are back to intelligent goal-terminated activity, for the behavioural evidence will be the same for both it and putatively ‘genuine’ goal-directed activity. Deciding whether a particular animal is exhibiting goal-directed behaviour will hinge on how we interpret its motions: are they sufficiently directed towards achieving the goal? How we answer this question depends on our evaluation of the appropriateness of the attempts, e.g., does the animal succeed in recognizing partial progress towards its goal, does it learn, does it abandon too early what to us seem like promising avenues? Answers to these questions admit of degrees and disagreement. If we set the standards high, only well-reasoned goal-directed activities will count as goal-directed at all, and once again the distinction between well-reasoned and ill-reasoned goal-directed activities is lost. If we do not set standards but allow the notion to admit of degrees, then we are left saying that some activities are more truly goal-directed than others, an unhappy way of looking at things; either one has a goal or not. An alternative and better way of describing such activities would be to describe some goal-terminated activities as more appropriately marshalling their sub-routines than others.

  It is tentatively proposed, then, that the GPS model of goal-direction can do justice to the observed behaviour of animals. Nothing an animal could do short of giving us a disquisition on its goals and methods would give us evidence pointing to more marvellous control systems than those sketched for GPS. I make the proposal tentative not because I intend to replace it later, but because it is difficult to see what, if any, limits can be set on the heuristics and learning potentials of GPS-styled systems, and so it is difficult to say with any certainty that such systems would or would not be adequate to model any animal behaviour yet to be discovered. The behavioural evidence so far culled by the psychologists does not suggest that animals are capable of behaviour superior in shrewdness or different in style from the behaviour of such a control system. The question of whether human beings, with their greater sophistication, require different sorts of control systems is a very difficult question, but some progress will be made on it indirectly in Chapters 7, 9 and 10.

  Supposing, then, that we accept GPS as an adequate simple model of goal-direction, the next question is whether a control system having the general characteristics of GPS could be produced by the evolutionary system proposed. The first problem is the specification of an end-state. If animals were in the habit of scrambling about until they were presented by a unique retinal projection or olfactory stimulus, or any other simple and easily described peripheral afferent, the problem would be simple. Such a stimulus could serve as the terminating stimulus for whatever efferents had been operating, and its occurrence could be designated the end-state. But animals do not do this. Their goals are not often the experiencing of particular peripheral sensations, with the possible exception of hunger satiation signals. Animals are more apt to have the goal of finding their way home, or to the feedbox, or of reaching safe ground, or of breaking the clam-shell, and no unique sensory presentation or even finite disjunction of such presentations would serve to signal the achieving of these goals. For these relatively sophisticated goals there is no hope of finding a peripheral neural state that would serve as a suitable end-state. So the peripheral stimulation must be processed into something more sophisticated.

  This is the problem of pattern recognition and stimulus generalization, and considerable progress has been made in exhibiting the power of stimulated neural nets to perform these tasks. The details need not concern us, for all that is important to us is that the output activity of the afferent nervous system (i.e., the central activity that becomes the input to the efferent nervous system) should be capable of being determined not by specific peripheral patterns but by external conditions described more generally. The successes of relatively small pattern analysing devices built of modules analogous to, but much simpler than, neurons indicate that the immense and highly complex neural net that makes up the afferent nervous system is fully equal to this task
. Supposing the afferent portion of the brain of a higher animal to be such a neural net, it should be capable of producing output states sufficiently interpreted relative to peripheral stimulation to serve as specifications of goal-states. For example, while particular neural signals near the retina might fire normally if and only if vertical lines predominated on the retinal projection, signals at a higher level in the neural net would fire normally if and only if the animal were surrounded by large vertical objects, and at a higher level still there would be signals – that is, as we have seen, concatenations of severally ambiguous signals – firing normally if and only if the animal were safely hidden among the white pines. Such high-level activity or some state resulting from it could serve as an end-state for goal-directed behaviour, e.g., running until safe ground is reached.

  The question that remains is whether such systems of subroutines and end-states could evolve within the individual brain under the principles we have already established. A detailed account of mechanisms and structures that would be required for such ‘learning’ would take us much farther into the area of detailed empirical speculation than I wish to go. Our ‘model’ of the evolving brain would have to be made much more detailed and even fitted out with variables to which numerical weights could be assigned in formulae, and so forth. Since it is very much in our interests to keep all such hypotheses as general as possible, such specifications would work to defeat our purposes. In a very general way, however, we can see what direction such an evolution might take.

  It is a common belief among psychologists that the normal behaviour of animals and perhaps even of man is divisible into hierarchies of patterns generated from the animal’s basic needs – essentially food, defence and procreation. This belief can be found in various forms in philosophical and psychological theories dating back at least to Aristotle, and is, of course, the offspring of a more homespun belief that the purposes of man and beast are nested in a few or perhaps just one basic purpose. One saws the plank to build the door to put on the house to keep it secure to protect one’s health to stay alive. Many models have been proposed to account for this characteristic of behaviour, and typically they have been hydraulic in inspiration, with pressures being channelled this way and that. Such hydraulic models, with their mysterious fluids – humours, libido, élan vital – being shunted about, are, of course, very much out of date today, but they do suggest a general principle of generation that might find many different embodiments within the more powerful and versatile framework of information theory.