I. Strategies of simplification
The field of Artificial Intelligence has produced so many new concepts--or at least vivid and more structured versions of old concepts--that it would be surprising if none of them turned out to be of value to students of animal behavior. Which will be most valuable? I will resist the temptation to engage in either prophecy or salesmanship; instead of attempting to answer the question: "How might Artificial Intelligence inform the study of animal behavior?" I will concentrate on the obverse: "How might the study of animal behavior inform research in Artificial Intelligence?"
I take it we all agree that in the end we want to be able to describe and explain the design and operation of animal nervous systems at many different levels, from the neurochemical to the psychological and even the phenomenological (where appropriate!), and we want to understand how and why these designs have evolved, and how and why they are modulated in the individual organisms. AI research, like all other varieties of research on this huge topic, must make drastic oversimplifications in order to make even apparent progress. There are many strategies of simplification, of which these five, while ubiquitous in all areas of mind/brain research, are particularly popular in AI:
(1) Ignore both learning and development; attempt to model the "mature competence" first, postponing questions about how it could arise.
(2) Isolate a particular subcomponent or sub-sub-component, ignoring almost all problems about how it might be attached to the larger system.
(3) Limit the domain of operation of the modeled system or subsystem to a tiny corner of the real domain--try solve a "toy problem", hoping that subsequent scaling-up will be a straightforward extrapolation.
(4) Bridge various gaps in one's model with frankly unrealistic or even deliberately "miraculous" stopgaps-- "oracles", or what I have called "cognitive wheels" (Dennett, 1984). (In the neurosciences, one posits what I have called "wonder tissue" to bridge these gaps.)
(5) Avoid the complexities of real-time, real-world coordination by ignoring robotics and specializing in what I call "bed-ridden" systems: systems that address the sorts of problems that can be presented via a narrow "verbal" channel, and whose solutions can be similarly narrowly conveyed to the world. (Dennett, 1980)
Many of the best-known achievements of AI have availed themselves of all five of these strategies of simplification: chess-playing programs, and natural language parsers and speech recognition systems, for instance. Some critics are hostile to any efforts in cognitive science enabled by these strategies, but there is no point in attempting to "refute" them a priori. Since they are strategies, not doctrines or laws or principles, their tribunal is "handsome is as handsome does." The results to date are an inconclusive mixture. One theorist's deep but narrow insight is another's falsely lit detour; just which points of verisimilitude between model and modeled should count as telling partial confirmation and which as tricked up and misleading is often a matter of free-form dispute.
Instead of spending yet more time debating the wisdom of these strategies of simplification, one might just adopt some rival strategy or strategies, and let posterity decide which are the most fruitful. An obvious candidate, especially on this occasion, is to turn from the simulation of human cognition to the simulation of animal cognition. If human minds (or brains) are too complex, why not start with simpler minds--insect minds or bird brains?
'Why not try to do a whole starfish, for instance? It has no eyes or ears, only rudimentary pattern-discrimination capacities, few modes of action, few needs or intellectual accomplishments. That could be a warm-up exercise for something a bit more challenging: a turtle, perhaps, or a mole. A turtle must organize its world knowledge, such as it is, so that it can keep life and limb together by making real time decisions based on that knowledge, so while a turtle- simulation would not need a natural language parser, for instance, it would need just the sorts of efficient organization and flexibility of control distribution you have to provide in the representation of world knowledge behind a natural language parsing system of a simulated human agent such as SHRDLU' (Winograd, 1972) --(Dennett, 1978)
So one reasonable motive for attempting AI modeling of animals is that it permits simplicity without unnatural truncation of systems--and you can get as much simplicity as you want by just descending the phylogenetic scale. If starfish are too hard, try paramecia. simplifying side-step in this descent is to opt for the modeling of imaginary simple animals, living in simulated simple environments. Such thought experiments can be brought to half-life, so to speak, and halfway put to the test, thanks to the computer's capacity to keep track of, and resolutely follow up the implications of, the loose ends among one's assumptions. The three-wheeled Martian iguana I fantasized in 1978 has not, to my knowledge, been born in any computer, but several of its brethren have been created. Braitenberg (1984), coming from the neuroscientific end of the spectrum, has described a considerable menagerie of ever more complex "vehicles" exhibiting increasingly "psychological" competences, and coming from the opposite, AI corner, we haveRod Brooks' artificial insects, real robots that perform uncannily biological-seeming feats with extremely simple control circuits (Brooks, 1987).
Of course the farther you get from human beings the less likely your successes are to shed light on the puzzles of our cognitive economies. but by training our attention on the differences that emerge, as well as the invariances that persist, as one moves along the actual phylogenetic scale, we may harvest insights about fundamental design principles of nervous systems, and about the traditions and precedents of design that are the raw materials of subsequent design innovations.
There is nothing new about this strategy, except for the relative ease with which very intricate models can now be "built" and "flight-tested", thanks to computer modeling. Some will object that much of the best computer modeling of simple nervous systems has had nothing in common with AI--and is indeed often the work of people quite hostile to the methods, assumptions, and pretensions of the ideologues of AI. That is true, but I think it is a fact of diminishing interest. The gulf between neuroscientists trying to build realistic models of simple neural systems from the bottom up (e.g., Hawkins and Kandel 1984) and "pure" AI modelers who frankly ignore all biological constraints (e.g., Doyle 1979) is being filled in with just about every intermediate variety of modeler. The antagonisms that remain say more about the economics and sociology of science than about the issues.
One reason people in AI have been dissuaded from simulating animal cognition is that they would have to give up one of their favorite sources of inspiration: introspection and reflection about how they perform the cognitive operations they choose to model. I say "inspiration" and not "data" since only under the most structured and well-studied conditions do people in AI count a match between model and "introspection" as corroboration of their model (Ericsson and Simon, 1984) But without the luxury of such self-modeling, or even the wealth of everyday lore we all accrue about human habits of thought, AI modelers are short on materials for the task of modeling animal cognition. It is here, of course, where the study of animal behavior might come to the rescue.
II. The intentional stance as a designer's strategy
All the AI efforts to simulate cognition, variously enriched by data from experimental studies of real creatures (and people) and by casual observation and introspection, are essentially engineering projects: attempts to design "machinery" with particular "psychological" talents. As engineering projects, their success depends heavily on the imaginative and organizational powers of the designers, who must juggle an ever increasing number of somewhat vague, somewhat conflicting set of "specs"--specifications--of the phenomena being modeled.
One way of imposing order--at least a temporary, tentative, order--on the interdependent tasks of clarifying (and revising) the specs and designing a system that meets the specs is to adopt the intentional stance. One adopts of strategy of treating the systems in question as intentional systems, approximations of rational agents, to whom one attributes beliefs, desires, and enough rationality to choose the actions that are likely to fulfill their desires given the truth of their beliefs. We all adopt the intentional stance towards our friends and relatives and other human beings, but one can also get results by adopting the stance when designing or diagnosing certain artifacts-- typically computer systems--and when studying the behavior of non-human animals.
My analysis and defense of adopting the intentional stance in the study of animal behavior (Dennett, 1983) has been greeted by workers in the field with about equal measures of enthusiasm, dismissal, utter disinterest and skeptical curiosity. A particularly insightful curious skeptic is C. M Heyes (forthcoming), who slyly wonders whether I have managed to drum up the enthusiasm by merely "providing a concert party for the troops"--making ethologists feel better about their lonely and ill-understood campaigns in the bush--while failing utterly to make good on my promise to show how "disciplined application of the intentional stance in cognitive ethology will yield descriptions of animal behaviour that are especially useful to the information processing theorist."
This would seem to be the ideal forum for me to respond to Heyes' challenge, for while I am delighted to be found entertaining (by the better sort of audience), I do aspire to something more. Heyes quotes my central claim:
The intentional stance, however, provides just the right interface between specialties: a "black box" characterization of behavioral and cognitive competences observable in the field, but couched in language that (ideally) heavily constrains the design of machinery to put in the black box. (Dennett, 1983, p.350)
and then goes on to ask "but how might intentional accounts 'constrain' information processing theories?" In particular, since the ultimate destination of theory on my view is an utterly mechanistic account of the brain's activities, and since I insist that the intentional stance yields an idealized and instrumentalistic account at best, it seems to Heyes that the intentional stance is at best a digression and distraction from the task at hand. How can a frankly idealizing model--which unrealistically describes (or prescribes) presumably optimal performance--actually constrain the development (from what I call the design stance) of a mechanistic and realistic model? To put it even more bluntly, how could instrumentalistic fictions help us figure out the mechanistic facts?
One can view the intentional stance as a limiting case of the design stance: one predicts by taking on just one assumption about the design of the system in question: whatever the design is, it is optimal. This assumption can be seen at work whenever, in the midst of the design stance proper, a designer or design investigator inserts a frank homunculus (an intentional system as subsystem) in order to bridge a gap of ignorance. The theorist says, in effect, "I don't know how to design this subsystem yet, but I know what it's supposed to do, so let's just pretend there is a demon there who wants nothing more than to do that task and knows just how to do it." One can then go on to design the surrounding system with the simplifying assumption that this component is "perfect." One asks oneself how the rest of the system must work, given that this component will not let down the side.
Occasionally such a design effort in AI proceeds by literally installing a human module pro tempore in order to explore design alternatives in the rest of the system. When the HWIM speech recognition system was being developed at Bolt Baranek and Newman (Woods and Makhoul, 1974), the role of the phonological analysis module, which was supposed to generate hypotheses about the likely phonemic analysis of segments of the acoustic input, was temporarily played by human phonologists looking at segments of spectrograms of utterances. Another human being, playing the role of the control module, could communicate with the phonology demon and the rest of the system, asking questions, and posing hypotheses for evaluation.
Once it was determined what the rest of the system had to "know" in order to give the phonologist module the help it needed, that part of the system was designed (discharging, inter alia, the control demon) and then the phonologists themselves could be replaced by a machine: a subsystem that used the same input (spectrograms--but not visually encoded, of course) to generate the same sorts of queries and hypotheses. During the design testing phase the phonologists tried hard not to use all the extra knowledge they had--about likely words, grammatical constraints, etc.--since they were mimicking stupid homunculi, specialists who only knew and cared about acoustics and phonemes.
Until such time as an effort is made to replace the phonologist subsystem with a machine, one is committed to virtually none of the sorts of design assumptions about the working of that subsystem that are genuinely explanatory. But in the meantime one may make great progress on the design of the other subsystems it must interact with, and the design of the supersystem composed of all the subsystems.
The first purported chess playing automaton was a late 18th century hoax: Baron Wolfgang von Kempelen's wooden mannikin which did indeed pick up and move the chess pieces, thereby playing a decent game. It was years before the secret of its operation was revealed: a human midget chess master was hidden in the clockwork under the chess table, and could see the moves through the translucent squares--a literal homunculus (Raphael, 1976). Notice that the success or failure of the intentional stance as a predictor is so neutral with regard to design that it does not distinguish von Kempelen's midget-in-the-works design from, say, Berliner's Hitech, a current chess program of considerable power (Berliner and Ebeling, 1986). Both work; both work well; both must have a design that is a fair approximation of optimality, so long as what we mean by optimality at this point focuses narrowly on the task of playing chess and ignores all other design considerations (e.g., the care and feeding of the midget vs the cost of electricity--but try to find a usable electrical outlet in the eighteenth century!). Whatever their internal differences, both systems are intentional systems in good standing, though one of them has a subsystem, a homunculus, that is itself as unproblematic an intentional system as one could find. Intentional system theory is almost literally a black box theory--but that hardly makes it behavioristic in Skinner's sense (see Dennett, 1981).
On the contrary, intentional system theory is an attempt to provide what Chomsky, no behaviorist, calls a competence model, in contrast to a performance model. Before we ask ourselves how mechanisms are designed, we must get clear about what the mechanisms are supposed to (be able to) do. This strategic vision has been developed further by Marr (1982) in his methodological reflections on his work on vision. He distinguishes three levels of analysis. The highest level, which he misleadingly calls computational, is in fact not at all concerned with computational processes, but strictly (and more abstractly) with the question of what function the system in question is serving--or, more formally, with what function in the mathematical sense it must (somehow or other) "compute". At this computational level one attempts to specify formally and rigorously the system's proper competence (Millikan, 1984, would call it the system's proper function). For instance, one fills in the details in the formula:
"given an element in the set of x's as input, it yields an element in the set of y's as output according to the following formal rules . . . "
--while remaining silent or neutral about the implementation or performance details of whatever resides in the competent black box. Marr's second level down is the algorithmic level, which does specify the computational processes but remains as neutral as possible about the hardware, which is described at the bottom level.
Marr claims that until we get a clear and precise understanding of the activity of a system at its highest, "computational" level, we cannot properly address detailed questions at the lower levels, or interpret such data as we may already have about processes implementing those lower levels. This echoes Chomsky's long insistence that the diachronic process of language-learning cannot be insightfully investigated until one is clear about the end-state mature competence towards which it is moving. Like Chomsky's point, it is better viewed as a strategic maxim than as an epistemological principle. After all, it is not impossible to stumble upon an insight into a larger picture while attempting to ask yourself what turn out to be subsidiary and somewhat myopically posed questions.
Marr's more telling strategic point is that if you have a seriously mistaken view about what the computational-level description of your system is (as all earlier theories of vision did, in his view), your attempts to theorize at lower levels will be confounded by spurious artifactual puzzles. What Marr underestimates, however, is the extent to which computational level (or intentional stance) descriptions can also mislead the theorist who forgets just how idealized they are (Ramachandran, 1985).
The intentional stance postpones consideration of several types of cost. It assumes that in the black box are whatever cognitive resources are required to perform the task or subtask intentionally described, without regard (for the time being) of how much these resources might cost, either in terms of current space, material and energy allocations, or in terms of "research and development" --the costs to Mother Nature of getting to such a design from a pre-existing design. And so long as cost is no object, there is no reason not to overdesign the system, endowing it with a richer intentional competence than it usually needs, or can afford.
But it is precisely these costs that loom large in biology, and that justify the strategic recommendation that we should be bargain hunters when trying to uncover the design rationales of living systems: always look for a system that provides a mere approximation of the competence described from the intentional stance, a cheap substitute that works well enough most of the time.
When great tits do a surprisingly good job of approximating the optimal foraging strategy in Krebs' "two-armed bandit" apparatus (Krebs, Kacelnik and Taylor, 1978), we do not make the mistake of installing a mathematician-homunculus in their brains to work out the strategy via a dynamic programming algorithm. As Krebs notes, we cast about for cheaper, more realistic machinery that would obtain similar results. When the honey bees' oleic acid trigger was uncovered, this deposed the public-health- officer-homunculus whose task it was to recognize bee corpses as health hazards and order the right measures.
But if such intentional systems are always destined to be replaced by cheap substitutes, what constraining power do the interim intentional stance descriptions actually have? Only this: they describe the ideal against which to recognize the bargain. They remind the theorist of the point of the bargain device, and why it may be such a good deal. For instance, the vertical symmetry detectors that are ubiquitous machinery in animal vision are baffling until we consider them, as Braitenberg (1984) recommends, as quick-and-dirty discriminators of the ecologically important datum that some other organism is looking at me. The intentional stance provides a tentative background against which the researcher can contrast the observed behavior as a competence --in this case the competence to detect something in the environment as another organism facing head on, about which one might want to ask certain questions, such as: What is the prowess and cost-effectiveness of this machinery?--a question that cannot even be posed until one makes an assumption about what the machinery is for. If you consider it merely as a symmetry detector, you miss the rationale for its speedy triggering of orientation and flight-preparation subsystems, for instance. It is the intentional characterization that can vividly capture the larger role of the machinery in the whole system.
To return to Heyes' question, with which we began, in what way does the intentional stance constrain the development of design hypotheses in information processing theories? It contrains in the same way arithmetic constrains the design of hand calculators. Arithmetic can also be viewed as an abstract, ideal, normative system (how one ought to add, subtract, multiply and divide), and then we can see that although individual hand calculators all "strive" to meet this ideal, they all fall short in ways that can be explained by citing cost-effectiveness considerations. For instance, arithemtic tells us that 10 divided by 3 multiplied by 3 is 10, but hand calculators will tell you that it is 9.9999999, due to round-off or truncation error, a shortcoming the designers have decided to live with, even though such errors are extremely destructive under many conditions in largers systems that do not have the benefit of human observer/users (or very smart homunculi!) to notice and correct them.
Just as there are many different designs for hand calculators, all of which implement--with various different approximations and shortcomings--the Arithmetical System, so many different designs of the neural hardware might implement any particular intentional system--with different attendant fallings short. So an intentional stance chacterization does constrain design, but only partially. It is one constraint among others, but a particularly fruitful and central one: the one that reminds the designer or design-interpreter of what the system is supposed to do.
III. Why vervet monkeys don't perform speech acts
Another way of looking at the intentional stance as a tactic to adopt in the field is to consider the likely fruits of taking it as seriously as possible. One says to oneself, in effect: "Now if these animals really believed such-and-such and really desired such-and-such, they would have to believe (desire, intend, expect, fear) such-and-such as well. Do they?" It is the intentional stance's rationality assumption that generates ("a priori" as it were) the consequent to be tested. Such an excercise can help uncover particular aspects of falling-short, particular hidden cheap shortcuts in the design, and help explain otherwise baffling anomalies in an animal's behavior. One uses the intentional stance to ask the question: What is it about the world in which this animal lives that makes this cheap substitute a good bargain? I will sketch an example of this drawn from my own very limited experience as an amateur ethologist (Dennett, 1988).
In June of 1984, I had a brief introduction to ethological field work, observing Seyfarth and Cheney observing the vervet monkeys in Kenya. In (Dennett, 1983) I had discussed the vervets and their fascinating proto-language, and had speculated on the likely fruits of using the intentional stance to get a better fix on the "translation" of their utterance-types. In particular, I had proposed attempting to use what I called the Sherlock Holmes method: setting cognitive ruses and traps for the vervets, to get them to betray their knowledge and understanding in one-shot experiments.
Once I got in the field and saw first hand the obstacles to performing such experiments, I found some good news and some bad news. The bad news was that the Sherlock Holmes method, in its classical guise, has very limited appplicability to the vervet monkeys--and by extrapolation, to other "lower" animals. The good news was that by adopting the intentional stance one can generate some plausible and indirectly testable hypotheses about why this should be so, and about the otherwise perplexing limitations of the cheap substitutes discovered in the vervets.
A vocalization that Seyfarth and Cheney were studying during my visit had been dubbed the Moving Into the Open (or MIO) grunt. Shortly before a monkey in a bush moves out into the open, it often gives a MIO grunt. Other monkeys in the bush will often repeat it--spectrographic analysis has not (yet) revealed a clear mark of difference between the initial grunt and this response. If no such echo is made, the original grunter will often stay in the bush for five or ten minutes and then repeat the MIO. Often, when the MIO is echoed by one or more other monkeys, the original grunter will thereupon move cautiously into the open.
But what does the MIO grunt mean? We listed the possible translations to see which we could eliminate or support on the basis of evidence already at hand. I started with what seemed to be the most straightforward and obvious possibility:
"I'm going"
"I read you. You're going."
But what would be the use of saying this? Vervets are in fact a taciturn lot, who keep silent most of the time, and are not given to anything that looks like passing the time of day by making obvious remarks. Then could it be a request for permission to leave?
"May I go, please?"
"Yes, you have my permission to go."
This hypothesis could be knocked out if higher ranking vervets ever originated the MIO in the presence of their subordinates. In fact, higher ranking vervets do tend to move into the open first, so it doesn't seem that MIO is a request for permission. Could it be a command, then?
"Follow me!"
"Aye, Aye, Cap'n."
Not very plausible, Cheney thought. "Why waste words with such an order when it would seem to go without saying in vervet society that low-ranking animals follow the lead of their superiors? For instance, you would think that there would be a vocalization meaning 'May I?' to be said by a monkey when approaching a dominant in hopes of grooming it. And you'd expect there to be two responses: 'You may' and 'You may not,' but there is no sign of any such vocalization. Apparently such interchanges would not be useful enough to be worth the effort. There are gestures and facial expressions which may serve this purpose, but no audible signals." Perhaps, Cheney mused, the MIO grunt served simply to acknowledge and share the fear:
"I'm really scared."
"Yes. Me too."
Another interesting possibility was that the grunt helped with coordination of the group's movements:
"Ready for me to go?"
"Ready whenever you are."
A monkey that gives the echo is apt to be the next to leave. Or perhaps even better:
"Coast clear?"
"Coast is clear. We're covering you."
The behavior so far observed is compatible with this reading, which would give the MIO grunt a robust purpose, orienting the monkeys to a task of cooperative vigilance. The responding monkeys do watch the leave-taker and look in the right directions to be keeping an eye out. "Suppose then, that this is our best candidate hypothesis," I said. "Can we think of anything to look for that would particularly shed light on it?" Among males, competition overshadows cooperation more than among females. Would a male bother giving the MIO if its only company in a bush was another male? Seyfarth had a better idea: suppose a male originated the MIO grunt; would a rival male be devious enough to give a dangerously misleading MIO response when he saw that the originator was about to step into trouble? The likelihood of ever getting any good evidence of this is minuscule, for you would have to observe a case in which Originator didn't see and Responder did see a nearby predator and Responder saw that Originator didn't see the predator. (Otherwise Responder would just waste his credibility and incur the wrath and mistrust of Originator for no gain.) Such a coincidence of conditions must be extremely rare. This was an ideal opportunity, it seemed, for a Sherlock Holmes ploy.
Seyfarth suggested that perhaps we could spring a trap with something like a stuffed python that we could very slyly and surreptitiously reveal to just one of two males who seemed about to venture out of a bush. The technical problems would clearly be nasty, and at best it would be a long shot, but with luck we might just manage to lure a liar into our trap. But on further reflection, the technical problems looked virtually insurmountable. How would we establish that the "liar" had actually seen (and been taken in by) the "predator", and wasn't just innocently and sincerely reporting that the coast was clear? I found myself tempted (as often before in our discussions) to indulge in a fantasy: "If only I were small enough to dress up in a vervet suit, or if only we could introduce a trained vervet, or a robot or puppet vervet who could . . ." and slowly it dawned on me that this recurring escape from reality had a point: there is really no substitute, in the radical translation business, for going in and talking with the natives. You can test more hypotheses in half an hour of attempted chitchat than you can in a month of observation and unobtrusive manipulation. But to take advantage of this you have to become obtrusive; you--or your puppet--have to enter into communicative encounters with the natives, if only in order to go around pointing to things and asking "Gavagai?" in an attempt to figure out what "Gavagai" means. Similarly, in your typical mystery story caper, some crucial part of the setting up of the "Sherlock Holmes method" trap is--must be--accomplished by imparting some (mis)information verbally. Maneuvering your subjects into the right frame of mind--and knowing you've succeeded--without the luxurious efficiency of words can prove to be arduous at best, and often next to impossible.
In particular, it is often next to impossible in the field to establish that particular monkeys have been shielded from a particular bit of information. And since many of the theoretically most interesting hypotheses depend on just such circumstances, it is often very tempting to think of moving the monkeys into a lab, where a monkey can be physically removed from the group and given opportunities to acquire information that the others don't have and that the test monkey knows they don't have. Just such experiments are being done, by Seyfarth and Cheney with a group of captive vervets in California, and by other researchers with chimpanzees. The early results are tantalizing but equivocal (of course), and perhaps the lab environment, with its isolation booths, will be just the tool we need to open up the monkeys' minds, but my hunch is that being isolated in that way is such an unusual predicament for vervet monkeys that they will prove to be unprepared by evolution to take advantage of it.
The most important thing I think I learned from actually watching the vervets is that they live in a world in which secrets are virtually impossible. Unlike orangutans, who are solitary and get together only to mate and when mothers are rearing offspring, and unlike chimps, who have a fluid social organization in which individuals come and go, seeing each other fairly often but also venturing out on their own a large proportion of the time, vervets live in the open in close proximity to the other members of their groups, and have no solitary projects of any scope. So it is a rare occasion indeed when one vervet is in a position to learn something that it alone knows and knows that it alone knows. (The knowledge of the others' ignorance, and of the possibility of maintaining it, is critical. Even when one monkey is the first to see a predator or a rival group, and knows it, it is almost never in a position to be sure the others won't very soon make the same discovery.) But without such occasions in abundance, there is little to impart to others. Moreover, without frequent opportunities to recognize that one knows something that the others don't know, devious reasons for or against imparting information cannot even exist-- let alone be recognized and acted upon. I can think of no way of describing this critical simplicity in the Umwelt of the vervets, this missing ingredient, that does not avail itself explicitly or implicitly of higher-order intentional idioms.
In sum, the vervets couldn't really make use of most of the features of a human language, for their world--or you might even say their lifestyle--is too simple. Their communicative needs are few but intense, and their communicative opportunities are limited. Like honeymooners who have not been out of each other's sight for days, they find themselves with not much to say to each other (or to decide to withhold). But if they couldn't make use of a fancy, human-like language, we can be quite sure that evolution hasn't provided them with one. Of course if evolution provided them with an elaborate language in which to communicate, the language itself would radically change their world, and permit them to create and pass secrets as profusely as we do. And then they could go on to use their language, as we use ours, in hundreds of diverting and marginally "useful" ways. But without the original information-gradients needed to prime the evolutionary pump, such a language couldn't get established.
So we can be quite sure that the MIO grunt, for instance, is not crisply and properly translated by any familiar human interchange. It can't be a (pure, perfect) command or request or question or exclamation because it isn't part of a system that is elaborate enough to make room for such sophisticated distinctions. When you say "Wanna go for a walk?" to your dog and he jumps up with a lively bark and expectant wag of the tail, this is not really a question and answer. There are only a few ways of "replying" that are available to the dog. It can't do anything tantamount to saying "I'd rather wait till sundown," or "Not if you're going to cross the highway," or even "No thanks." Your utterance is a question in English but a sort of melted- together mixture of question, command, exclamation and mere harbinger (you've made some of those going-out-noises again) to your dog (Bennett, 1976, 1983). The vervets' MIO grunt is no doubt a similar mixture, but while that means we shouldn't get our hopes too high about learning Vervetese and finding out all about monkey life by having conversations with the vervets, it doesn't at all rule out the utility of these somewhat fanciful translation hypotheses as ways of interpreting--and uncovering-- the actual informational roles or functions of these vocalizations. When you think of the MIO as "Coast clear?" your attention is directed to a variety of testable hypotheses about further relationships and dependencies that ought to be discoverable if that is what MIO means--or even just "sort of" means.
But is that all there is? Perhaps this is the "concert pary for the troops" Heyes supposes I am offering: I seem to end up saying that vervet monkeys don't really mean anything at all by their vocalizations. Am I also saying that vervet monkeys don't really believe anything What literally can the intentional stance show us about animal belief--about what is going on in the minds of the animals being studied?
That question, I am saying, is misguided. The intentional stance attributions of belief, for all their caveats, are as literal as any attributions of belief--including self-attributions--can ever get. There are no deeper facts about the beliefs of animals--or about our own. (In disagreement with Griffin (1981, 1984), I think the more particular hope that cognitive ethology will shed light on animal consciousness is a red herring. The only concepts of consciousness that yield genuinely explanatory attributions are applicable only to creatures with a full-fledged natural language--human beings). If you want to know the deep, objective truth about the contents of animal minds, then either you are curious about the actual design of their brains (at many levels of analysis), and the rationale of that design, or you just want to know the most predictive intentional stance characterization of the animal, with all its idealizations. If you think there is another, deeper sort of fact about animal minds, then the intentional stance won't help you find it--but then nothing will, since if that is your curiousity, you are no longer doing the cognitive science of wild animals; you are on a wild goose chase.
That is all I had in mind when I suggested that the intentional stance could be of use in characterizing the "behavioral and cognitive competences observable in the field." There is nothing revolutionary about my suggestion. I didn't invent the intentional stance, after all; I just proposed a way of thinking about the implications of what one is doing when one does what comes naturally, and suggested that it would help clarify the competences that any realistic design of an animal nervous system must provide.
Bennett, J. 1976, Linguistic Behaviour, Cambridge Univ. Press.
Bennett, J. 1983, "Cognitive Ethology: Theory or Poetry?" (commentary on Dennett, 1983) Behavioral and Brain Sciences, pp.356-58.
Berliner, H., and C. Ebeling, 1986, "The SUPREM Architecture: a new Intelligent Paradigm," Artificial Intelligence, 28, pp.3-8.
Braitenberg, V., 1984, Vehicles: Essays in Synthetic Psychology, MIT Press
Dennett, D., 1978, "Why not the whole iguana?" (commentary on Pylyshyn) , Vol 1, #1 pp.103-4
Dennett, D., 1980, "The Milk of Human Intentionality," (commentary on Searle, "Minds, Brains and Programs," Behavioral and Brain Sciences, pp. 428-30.
Dennett, D., 1981, "Three Kinds of Intentional Psychology," in R. Healey, ed., Reduction, Time and Reality Cambridge Univ. Press.
Dennett, D., 1983, "Intentional Systems in Cognitive Ethology: the 'Panglossian Paradigm' Defended," Behavioral and Brain Sciences
Dennett, D., forthcoming, "Out of the Armchair and into the Field," Poetics Today (special issue on Interpretation and Context), Israel.
Doyle, J., 1979, "A Truth Maintenance System," Artificial Intelligence, 12, pp. 231-72.
Ericsson K. A., and Simon, H. A., 1984, Protocol Analysis: Verbal Reports as Data, MIT Press.
Hawkins, R. D. and E. Kandel, 1984, "Steps toward a cell- biological alphabet for elementary forms of learning," in G. Lynch, J. L. McGaugh, and N. M. Weinberger, eds., Neurobiology of Learning and Memory, New York: Guilford, pp.385-404.
Heyes, C. M. (forthcoming) "Cognisance of Consciousness in the Study of Animal Knowledge," in Callebaut, W. & Pinxten, R., eds., Evolutionary Epistemology: a Multiparadigm Approach, Dordrecht/Boston: Reidel
Marr, D., 1982, Vision, MIT Press
Millikan, R., 1984, Language, Thought and Other Biological Categories, MIT Press.
Raphael, B., 1976, The Thinking Computer: Mind Inside Matter, San Francisco: Freeman.
Ramachandran, H. 1985, [Harvard Cognitive Science Symposium]
Rosenshein, S. and L. Pack Kaelbling, 1986, "The Synthesis of Digital Machines with Provable Epistemic Properties," Proc. of the 1986 Conference, AAAI, Monterey, Calif. March, 1986. ed., Joseph Y. Halpern, Los Altos: Morgan Kaufmann Publ.
Winograd, T., 1972, Understanding Natural Language New York: Academic Press
Woods, W., and Makhoul, J.,1974, "Mechanical Inference Problems in Continuous Speech Understanding," Artificial Intelligence, 5, pp. 73-91.