A case for the Cognitive principle of relevance

I have a cautious, even grudging, appreciation for the ideas of Sigmund Freud.  Though I am dispositionally inclined to dismiss Freud's ideas wholesale, I have come to credit him with some genuine insights, and to be curious about the automatic hostility that arose in me when first I encountered his ideas.  In contrast, I was smitten from the start with Dan Sperber & Deirdre Wilson's (1995) Relevance theory, and especially the Cognitive principle of relevance (hence, CPR).  The CPR states (I paraphrase) that ceteris paribus inferences are processed in an order determined by the size of their cognitive effects, from greatest to least.  From the start I found this principle elegant and an enormously powerful analytic tool.  But just as, with time, I came to some appreciation of Freud's ideas, so I came also to doubt the CPR.

My doubts focused on two issues in particular: the mechanism by which the relevance of an inference was assessed and the empirical tractability of Sperber & Wilson's theory.  The problem of the mechanism could be summed up thus: how can the mind, when prioritizing two inferences, know which of them will lead to greater cognitive effects without actually processing them?  The problem of empirical tractability was simply whether there could be any real empirical test of such a high-level theory.  Was there any real reason to believe it?

Both questions were answered for me when I connected the CPR with John Holland's discussion of adaptation in classifier systems.  Classifier systems are characterized by a set of inference rules (more strictly, IF-THEN statements defining state transitions) that define the behavior of the system in response to inputs from outside.  In Holland's work (1992; but see also 1995), he envisions classifier systems embedded in adaptive agents that interact with their environment, which includes other similar agents.  Classifier systems become adaptive when the rules receive feedback as to whether they contribute, however indirectly, to successful behavior by the agent.  Holland assigns each rule a strength property to summarize its history of effectiveness, and competing rules are processed partly in order of decreasing strength.  (The other consideration that determines their processing order is specificity of match between the rule's IF clause and environmental inputs.)  Holland goes on to discuss other sorts of adaptive mechanisms, but strength is the critical one for the present discussion.

Any system characterized by competing rules faces the credit assignment problem, that is, the problem of passing credit for an outcome back to all the rules that contributed to the outcome in a way that is proportionate to that contribution.  (The generalized delta rule handles this in feed-forward connectionist systems, but at the cost of requiring that credit assignment be handled by a distinct backward-looking process.)  Holland solves this problem with what he calls the "bucket brigade algorithm," wherein each rule passes some of its strength back to its antecedents in a kind of bid for processing priority.  In this way, credit assignment is handled as part of the regular sequential workings of the agent.

In this kind of system, a rule's processing priority grows in proportion to its history of contributions to successful behavioral outcomes.  And this is, I think, precisely the sort of process required to implement the CPR: a rule's strength, in effect, sums its (history of) contributions to cognitive/behavioral effects.  Thus Holland's classifier systems involve a mechanism that approximates the CPR in its operation.  Not only is there a mechanism that approximates the CPR, but it does so simply as a consequence of its fundamental, adaptive architecture.

Two limitations to the approximation must be noted.  First, the CPR requires that inferences be prioritized according to their actual relevance, not according to their history of relevance.  Classifier systems use a rule's history of contributions to successful outcomes as a proxy for their future contribution to successful outcomes.  This is a difference, but one that does not, so far as I can see, run afoul of anything important in the CPR.  Second, a rule's strength (and thus its processing priority) is determined partly by the specificity of its contribution to successful outcomes, and there is no analogy to this in the CPR, though I believe one could and probably should be developed, as otherwise a processor heavily favors very general inferences over specific ones, with the result that it incurs all the limitations of a general problem solving mechanism rather than a collection of specific problem-solving strategies.

If this approximation is admitted, then there still remains the question whether there is ground for believing the CPR.  Although I am aware of some of the empirical work on it, I never feel quite sure that this work really shows the CPR, as opposed to something else, at work.  I feel about this research like I feel about test of psychoanalytic theory: it doesn't disconfirm the hypothesis, but neither does it seem like strong support for such general claims.  This is a general problem for all high level theories.

I think there are two arguments for accepting the CPR, one weak and one strong.  The weak reason is the argument that human mental processes seem to lend themselves to description in terms of classifier systems, and that at least some cognitive dynamics could be understood as the result of a classifier-system architecture.  We might call this the resemblance argument, and although I find it persuasive on the whole, certainly there is room for doubt.  The stronger reason has to do with a property of classifier systems, computational completeness.  Computational completeness means that classifier systems can simulate the behavior of any other computational system.  If the human mind involves computational systems-which it certainly does, among other things-then the behavior of these systems can be mimicked by a classifier system.  If its behavior can be mimicked by a system involving an approximation of the CPR, then it is as if human behavior were guided by the CPR.  This latter argument may be rephrased that the CPR might as well be correct.

These considerations have persuaded me of the CPR enough to go ahead confidently using it as an analytic tool.  Its application to particular situations is still fraught with difficulty, of course, but I feel that at least the tool is sound.  This argument has been rambling around in my head since early 1997, and it seems sound to me, but I invite your comments and corrections.





Holland, John H. (1992). Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial intelligence (1st MIT Press ed.). Cambridge, Mass.: MIT Press.

Holland, John H. (1995). Hidden order: How adaptation builds complexity. Reading, Mass.: Addison-Wesley.

Sperber, Dan, & Wilson, Deirdre (1995). Relevance: Communication and cognition. Oxford, UK ; Cambridge, Mass. :: Blackwell.


  • Olivier Morin
    Olivier Morin 5 January 2009 (11:54)

    Thank you Brian, for this extremely stimulating post! The problem you uncover is extremely interesting: if the mind uses a criterion of relevance, i.e. cost/benefit ratio, to allocate resources to various thought processes, then it should somehow know beforehand what the payoff of investing in these thought processes will be.And, at least in some cases, we don’t seem capable to predict what we’ll gain from thinking a thought without actually thinking that thought. I am not entirely satisfied with your solution, though – it seems to me that allocating resources solely on the basis of past benefits would lead to a bootstrapping problem: if you give resources exclusively to a small number of cognitive processes that have proven their worth in the past, on the basis of their past record, then you cannot unlock resources for new inferential mechanisms that do not have a record at all. Thus it would seem that a cognitive relevance mechanism, so construed, would keep us from solving new problems, and would lock us in strategies that have proven efficient in the past, but might now be outperformed by other strategies that we cannot even try. At the extreme, one might argue, such a principle would prevent us from thinking anything at all – since everything we do we must do, once, for the first time, and since your relevance mechanism denies resources to thought processes lacking a track record, most inferences could never take place for the first time. [b] Relevance and Dithering[/b] One obvious solution would be dithering – introducing noise in the system, random leaps that would allow it to stimulate new inferential systems randomly. Put another way, we may imagine a system that allocates resources at random, but tends to favor inferential mechanisms with an interesting output. Think of it like the semi-random browsing of web users who explore new pages erratically, then redirect all their friends, through Twitter or Facebook, to pages that are more likely to interest them. Noise is a good way to avoid lock-in phenomena, and to allow bootstrapping. It doesn’t even have to be purposefully generated by the system: brains are complicated organs, they are likely to produce endogenous noise quite naturally. Such dithering, however, would be a direct violation of the principle of relevance: the system would allocate resources to inferential systems that are overwhelmingly likely to produce no interesting result at all, in order to explore new possibilities. At the end of the day, your mechanism would merely be relevance-biased, but cognitive relevance would fail to explain some of its properties among the most important. So, what is cognitive relevance worth, if the mechanism that implements it is likely to violate it regularly and on purpose ? I’d like to argue, in the second part of this comment, that the principle of cognitive relevance (and relevance principles in general) need not describe the way any specific mechanism works: relevance, as efficiency in general, may just be a beneficial by-product of several, more or less specific causal factors. [b] Cognitive relevance is best thought of as a principle, not a mechanism[/b] To me, Relevance is just efficiency with another name. A principle of efficiency states that we more or less try to maximize the effect of whatever it is that we do, relative to the cost that we pay for doing it – walking, breathing, swimming, digesting, building machines, etc. While we might not always maximize utility relative to our efforts, we do seem to try to avoid those moves that are entirely unlikely to yield any interesting result. Although not trivial, principles of efficiency are routinely held by doctors when they try to understand digestion or breathing, biomechanists when they try to explain walk, by everyone when we try to make sense of the way a machine works. Sperber and Wilson’s Relevance principles are just principles of efficency applied to thought and communication. I cannot see any good reason why most of our behaviours would be guided by efficiency, but somehow, thought and communication would escape that rule. Note that efficiency principles need not specify anything concerning how or why we are efficient at what we do: merely assuming that a behavior is efficient is sufficient to predict many interesting things about it. If a walker is efficient, we may expect her not to twist her ankles purposefully, insofar as her behavior is guided by her intention to walk, and the context is, walking-wise, more or less normal. Why is she efficient? How is she capable of it? Actually we don’t need to know – my guess would be, a uninteresting blend of genetic predispositions, many kinds of learning, with a conscious decision or two thrown in. She might be endowed with a specific mechanism geared to make her walking efficient. But I don’t see the need to assume the existence of such a mechanism, and most importantly, its absence would not compel us to abandon the principle of walking-efficiency. To wrap it up: in my view, the principle of cognitive relevance is not committed to any general assumption concerning the way efficiency might be obtained in treating information in a given brain. Efficiency, and the means of obtaining it, is likely to be as context-dependent and variable as the various costs and benefits of processing various pieces of information. That is: so variable and context-dependent as to make a general mechanism for ensuring cognitive efficiency both costly and underperforming – in other words, inefficient. But the absence of such a general mechanism does not in the least mean we should abandon the principle of cogntivie relevance. Is relevance so construed trivial, untestable or overly vague? Not at all! Many models of the mind (for example many Freudian models) allow for the possibility of pointless thought processes that the Cognitive Principle of Relevance excludes. In Freud’s models, the mind is not guided by a search for efficiency, as it is geared towards a sort of hydraulic equilibrium between pulsions. Those models make strikingly different predictions on specific matters – for example, language processing and the explanation of lapsus linguae. So cognitive relevance is not trivial – which, or course, makes it open to criticism. One thing it does not – and should not do? – is give a complete history and description of a Relevance mechanism.

  • Jean-Baptiste André
    Jean-Baptiste André 6 January 2009 (15:41)

    I feel that one of the key problem raised by Brian in the case of the CPR amounts to the classic problem of reverse causality. “My brain made computation X rather than computation Y [b]because[/b] X led to more inferences than Y” is a typical example of reverse causality. A first phenomenon is explained by a second one that came after it in time. This is of course inconsistent. The basic principle of causes and effects states that effects must follow their cause, not precede them. The functionalist analysis of living systems is full of reverse causality, e.g., “functional eyes develop in animal embryos [b]because[/b] they will allow the animal to see, once it is born”… For all living systems, including human cognition, the solution to reverse causality comes from Darwin. The eyes developing in an embryo are by-products of the success that similar eyes gave to the ancestors of this embryo. The past success of the eyes in other embryos, not their future success in this very embryo, cause their development. The same goes on in the case of the CPR. My brain computes in priority the most relevant inferences [b]because[/b] brains that did so survived and reproduced best in the past than brains that didn’t. It needs not involve any foreseeing ability by my brain. Of course, adaptive principles at the level of the individual life (of the kind mentioned by Brian) play the same role. They also apparently reverse causalities. But, for the “poverty of the stimulus” problem mentioned by Olivier, adaptive principles at the scale of the species are more likely to explain generally that brains (human or not) follow a principle of relevance.

  • Brian Malley
    Brian Malley 6 January 2009 (17:56)

    In reviewing my post, I see that I contradicted myself: in paragraph 3 I said that a rule’s specificity is not a matter of its strength, but a distinct consideration; in paragraph 6 I said that strength was partly a function of a rule’s specificity. Obviously there is a problem here, and it is this: in paragraph 6 and following I identify a rule’s strength with its bid to be processed. This is not correct. The bid is a function of strength and specificity, not strength alone. I don’t believe that this undermines any of my general points, but it certainly makes things muddy in my discussion, and I’m sorry about that.

  • Brian Malley
    Brian Malley 6 January 2009 (18:22)

    Although any contributor is grateful if anyone thinks his or her ideas are worth discussing, I’d like to quibble respectfully with Jean-Baptiste’s analysis of the problem and solution. My concern with his analysis is perhaps a minor one, but one about which I should like to be clear. It has always been my understanding of Sperber & Wilson’s theory that there was some proxy for an inference’s relevance that serves to determine its processing priority. I don’t recall if I ever asked Dan about it, but I asked Pascal Boyer, who knows Dan’s ideas well, and Pascal said he thought Dan had in mind some yet-to-be-discovered biological process. So, in defense of Sperber & Wilson’s proposal, I don’t think they ever claimed (or at least, I never interpreted them as claiming) that an idea’s processing priority was determined by its future relevance. I have been troubled by the lack of a specified proxy, and so I wrote to share my solution to the question of mechanism, but I do not think Sperber & Wilson’s theory is teleological. Even if it was, I do not think a Darwinian solution would be an answer. Even if, as we may suppose (without any actual evidence), that selection favored brains that processed in accordance with the CPR, this would at most favor the kinds of genes that built CPR-oriented minds, and we would still need an explanation of how CPR-oriented minds work. It’s very tricky to talk about the relation between the information encoded by genes and phenotypical behavior, but I think that if what genes specify is not the actual inferences to be prioritized but rather a process for prioritizing inferences, the selectionist account does not address the problem I raised. Jean-Baptiste, if I have misunderstood some part of your post, please forgive me and post a clarification.

  • Jean-Baptiste André
    Jean-Baptiste André 6 January 2009 (22:23)

    I agree with you more than 100% in saying that the darwinian view does not tell us anything about the way all this works. Saying that the eye develops because it was selected by natural selection does not tell us anything about the actual way eyes develop and function. Saying that the brain’s efficiency to cope with the world is a product of natural selection does not tell us anything either about its actual functioning! And it is certainly very exciting to speculate about potential mechanisms, such as the adaptive mechanism you discuss, that could allow the brain to approach optimal efficiency. What I wanted to point out in my comment is more basic (and almost trivial in fact). I wanted to stress that the adaptive/efficient functioning of cognition, which is the essence of the cognitive principle of relevance as Olivier puts it, is no more paradoxical than any other adaptive/efficient aspect of living systems. Said differently, the apparent paradox that computations are prioritized according to their future effects is not fundamentally different from the apparent paradox that functional designs develop in life in general. It is thus misleading, I believe, to present this paradox as a potential weakness of the relevance principle. The way relevance is approached by the brain may be highly complex, and poorly understood, it entails no more paradox than the way an eye comes to develop. Is that clearer? Do you still disagree?

  • Brian Malley
    Brian Malley 7 January 2009 (00:25)

    Thanks, Jean-Baptiste, for your clarification. I think we are in agreement on all points.

  • Brian Malley
    Brian Malley 7 January 2009 (03:59)

    Thanks, Dan, for weighing in here. I see the distinction you are driving at in your first point, and I must admit I have thought far less about the relevance-driven allocation of memory than about the prioritizing of inferences, so this is an important limitation of my discussion. Thanks also for the reference to Pezzulo’s work.

  • Dan Sperber
    Dan Sperber 7 January 2009 (13:26)

    I am grateful to Brian for raising these serious and difficult issues with the Cognitive Principle of Relevance (hoping that he or some other ICCI blogger will in a later post highlight the relevance of this principle to the cultural interests of the members and users of this Institute). I too find these issue quite difficult and the best answer still quite speculative. Still, I find that the discussion so far has been useful, and I hope I too can usefully contribute to it. Here are five points: 1) I find Brian’s reformulation of the CPR (“ceteris paribus inferences are processed in an order determined by the size of their cognitive effects, from greatest to least”) somewhat problematic. The standard version goes something like this: “Human cognition tends to be geared to the maximisation of relevance” where relevance is defined as a property of input to cognitive processes. The greater the cognitive effects of processing a given input, the greater its relevance (everything else being equal), and also the greater the processing effort involved in achieving these effects, the lesser its relevance (everything else being equal). Important points here are a) that the claim is about a tendency; b) realising the tendency involves gearing not only mechanisms involved in inference but also those involved in attention and in memory towards maximizing relevance. 2) I agree with Olivier that the CPR is just a principle of efficiency applied to cognition. What is not trivial – and could well be challenged – is that cognitive efficiency is, first and foremost, a matter of tending to maximise relevance (rather than, say, solve problems). This I would strongly claim for the human case, based on the fact that humans are massive information hoarders (and here, by the way, is a link with culture begging to be developed). For other species, especially those with much simpler cognitive systems, efficiency may well be solving a fixed set of problems at the lowest cost. 3) I agree with Jean-Baptiste that if Darwinian selection can favour cognitive architectures and mechanisms that tend to increase relevance, it will have done so, and that — but the argument would need to be properly spelled — something similar is true also of cognitive development. In “Modularity and Relevance” in particular, I have tried to speculate on how this might be so and to flesh out the counter-arguments Deirdre and I have given against the objection (first raised by Gazdar) that in order to process the most relevant inputs, the mind would have to process all inputs to assess their relevance, defeating the whole idea of efficiency. 4) Brian suggestion’s that something like Holland’s ideas on classifiers systems might be highly useful in thinking on how a cognitive system could tend toward maximal relevance is one well worth digging into (I say sheepishly, because Brian made the suggestion to me years ago and I read Holland sufficiently to appreciate Brian’s suggestion but not sufficiently to exploit it). 5) I agree with Olivier and Jean-Baptiste that the maximisation, or even the evaluation of relevance is not achieved by a dedicated mechanism – I would argue that it is a kind of ‘invisible hand’ by-product of massive modularity evolving under pressure for efficiency –, but I also agree with Brian that we want to understand how cognitive mechanisms and the way in which they are articulated contribute to relevance. So far only rather vague suggestions have been offered. Still, for interesting ideas about modelling the issue, see [url=http://www.istc.cnr.it/doc/1a_20060307141224t_pezzulo.pdf] “How Can a Massively Modular Mind Be Context-Sensitive? A Computational Approach” [/url] by [url=http://giovannipezzulo.blogspot.com/]Giovanni Pezzulo[/url] , and [url=http://www.istc.cnr.it/createhtml.php?nbr=1]other papers of his[/url]

  • Giovanni Pezzulo 9 January 2009 (14:47)

    First of all, thanks to Dan Sperber for inviting me to join this stimulating discussion, and for citing my paper. Admittedly, I did not elaborate too much this theme after the mentioned paper (at least not explicitly). However, the idea that cognitive agents should prioritize their processes depending on their (expected) relevance is a tacit assumption in my models. I agree with Dan’s “5 points” and in particular that an explicit evaluation of relevance is -obvioulsy- non efficient (“in order to process the most relevant inputs, the mind would have to process all inputs to assess their relevance, defeating the whole idea of efficiency.”) Nevertheless, I think that relevance is realized by a (computational/neural) *mechanism* and is not only a *principle*, in the sense that it is computationally/neurally realized and not only a regularity in the eye of the observer. The “trick” is performing an *implicit* (not *explicit*) evaluation of relevance; this can be done for example with “resources allocation” or “competition for limited resources”, like in my computational architecture. Of course there are other possible realizations, but I like the idea of “implicitness” (or “invisible hand”), and I believe that establishing relevance is a byproduct of some “optimization” process (selected by evolution) rather than a dedicated brain mechanism. (Maybe the ambiguity is in the term “mechanism”, which is often confused with “explicit mechanism”.) So, overall, I agree with Dan totally on these topics and I think that a computational specification of these ideas can be extremely useful (at least for making it clearer). Even if we assume so, as I would do, it remains the question of what are the relevant dimensions of “relevance” (and how are they measured/compared). I have two considerations. First, according to the CPR relevance depends on the “cognitive effects of processing a given input”. This is quite general, but seems to me that CPR focuses mostly on beliefs and belief revision (that is, an input is very relevant if it has great impact on my beliefs). In addition, I would stress the importance of *goals* and *actions* for determining relevance. An input is more or less relevant for me (1) depending on my current goals (or simpler motivations such as drives), and (2) if it helps me selecting/taking an action instead of another. In some sense these dimensions are already present in CPR (since beliefs can serve to select goals and actions), but: – I have the impression that CPR de-emphasizes goals and goal contexts, that are of primary importance. Btw: Barsalou has provided a good cognitive model of this process in his paper “Deriving categories to achieve goals”: http://www.psychology.emory.edu/cognition/barsalou/papers/Barsalou_TLM_1991_goal_derived_categories.pdf – I think that CPR applies more to full-fledged cognitive agents (having complex cognitive minds), while I would say that determining relevance is important for simpler agents, too, since all agents have to face the problem of action selection—and, as a consequence, of input selection. This fact can be interesting in evolutionary/developmental terms. For example, I would expect that evolution has favored the selection of the most relevant stimuli for action selection first, and only successively for cognition or belief revision (is there some adaptaion/exaptation?). Anyway, these are largely empirical issues. Second: whatever the relevant dimensions of relevance are, is “actual relevance” evaluated, or also “past relevance” or “expected relevance”? Personally I believe that “expected relevance” plays a fundamental role… recent empirical research is showing that the brain is highly proactive and future-oriented, and it is always busy doing predictions and evaluating them… anyway, these are largely empirical issues, too. [to conclude: I apologize if I misunderstood parts of the CPR theory… I have read it a long time ago.]

  • Stewart Guthrie 22 January 2009 (22:14)

    Hi all, great thread! Advance apologies for jumping in late (just discovered it last night) & then jumping right out again (as I’m prepping for a trip & will be mostly off line until at least Feb. 2). I’ve a few ideas that may be relevant, since I’ve been advocating the CPR since 1980 (briefly) and 1993 (in some detail). It’s important to my cognitive theory of religion and, more generally, that of anthropomorphism and animism. These ideas may help with the “specificity” of contribution to success, and some mechanisms, that Brian (Jan 5 & 6) wants in CPR. They’re also consistent with those of Olivier, Brian & Dan that CPR itself is a principle not a mechanism; with Olivier & Dan that relevance & efficiency are close correlates; with Jean-Baptiste that Darwin supplies the solution to reverse causality; with Dan that massive modularity is at work, and with Giovanni that mechanistic “tricks” and specific goals are (at least at a lower level) involved. They may also provide a few, limited & highly specific, answers to Dan’s question, how cognitive mechanisms contribute to relevance. My argument (1980, 1993 & elsewhere) begins with the pervasive ambiguity and uncertainty of perception & cognition, & describes the unconscious, evolved strategy with which we meet them, namely guessing first at what matters most. What matters most usually is (though the hierarchy is context dependent) what is most highly organized and thus inferentially most powerful. This is complex organisms, especially complex animals and, most of all, humans. Hence our inferences tend disproportionately to feature animate and humanlike beings and qualities. Permit me some abbreviated, illustrative quotations on why they do, first from “A Cognitive Theory of Religion,” 1980:188: “A model may be ‘important’ for many reasons, but its level of organization surely is a crucial one. This reason is both intellectual and pragmatic. Intellectually, the more phenomena we can bring under a single organization, the more experience we ‘understand.’ Pragmatically, more highly organized things normally have a greater impact on us . . . . Snakes are more highly organized than twigs . . . . Nothing, on the whole, is more relevant to humans than humans [in part] because they are complex and multifaceted [and hence] generate a wide variety of phenomena.” The CPR, that is, favors inferences that tap into natural organization, especially complex life forms, partly because that gets more cognitive bang per cognitive buck and partly for narrower pragmatic reasons. However, cognition aims not only at life forms but more broadly at the highest organization possible, living or not. “Animism, Perception, & the Effort after Meaning” (Ch 2 of Faces in the Clouds, 1993) makes “three linked observations: perception is interpretation, interpretation aims at significance, and significance generally corresponds to the degree of organization perceived . . . . Interpretation [is necessary to see even] colors, lines, and edges . . . Seeing a . . . line means guessing the best configuration [of dots]: that which is most coherent and most significant . . . . [i.e.,] which generates the most information . . . hence we scan what registers at lower levels of complexity and integration (for example, dots and lines) with models from higher levels (for example, edges and objects) . . . . [Moreover,] compared to inorganic matter, organisms are hot spots of order, humming with information, with process serving process. . . . organisms including ourselves are significant to each other not only as threats and resources . . . but as information about their environments as well. Fishermen know that where gulls gather over the ocean, small fish likely are at the surface, chased by larger ones below . . . . The strategy for discovering these patterns is, again, that of Pascal’s wager, namely, guessing high” (1993:41-45). This unconscious strategy is built both on the CPR and on the principle, better safe than sorry. Various modular mechanisms have evolved in the service of this perceptual and cognitive strategy. Any of them may key us to infer the presence of particular complex animals, including humans and human communications, artifacts, and traces. Subsequently and cumulatively, they serve as checks on inferences. Inferences of animacy especially produce bursts of information, and our mechanisms for initiating such inferences are multiple, well developed, and on hair triggers. They include sensitivities to eyes; faces; bilateral symmetry; motion that is self initiated, contingent, goal oriented, or articulated in certain ways (e.g., those of arthropod or vertebrate motion); smells (e.g., of conspecifics, predators or prey); and sounds (e.g., those of locomotion, communication, and other activities). Other mechanisms—more complex and perhaps partly learned—include sensitivities to design (an aspect of goal orientation not contingent on motion), symbolism, and immanent justice (cf Dan’s inquiry in another thread). Though diverse and modular, all these mechanisms concern inferences about complex animals, so they may be idiosyncratic; but perhaps they’ll suggest analogous mechanisms for other kinds of inferences as well. Perhaps the common thread is that they are (as Nietzsche noted perception is) built into the ecological relation between cognizer and cognized. As Olivier points out, this seems to lock us into strategies of the past. To a degree, it must; but that’s evolution for you. Best to all, Stewart

  • Dan Sperber
    Dan Sperber 24 January 2009 (19:27)

    I am grateful to Giovanni for his contribution to this discussion. We are in agreement on most points, in particular on his concluding remark that what matters to directions of our cognitive processes is [i]expected [/i]relevance rather than past or actual relevance. Of course, expectation of relevance will be satisfied or frustrated to the extent that they match actual relevance. And of course, past relevance, or traces of it may inform expectations of relevance. To give an illustration, I attend to the ease with which I can break the stem of a melon because I expect this to be relevant to determining its ripeness. I do so because such a test has been relevant to me in this way in the past. If this time the test fails to be actually relevant, I will be disappointed and may not use this test in the future. Giovanni writes: “- I have the impression that CPR de-emphasizes goals and goal contexts.” It is true that we have been working mostly on relevance achieved through belief revision, but this is linked to the fact that our focus has been on comprehension, and it implies no theoretical bias or commitment to considering epistemic inference more important than practical inference. In fact, when some input leads to action, the cognitive effect is likely to be greater than if did not (everything else being equal). Compare Joan telling you that she takes sugar in her coffee (1) when you are talking about sweetening coffee in general or (2) when you are preparing coffee for her. In both cases, Joan’s statement may trigger the same range of epistemic inferences (about her taste, her life style, her health…) and, on top of that, in the second case it allows practical inferences (if the sugar is in the cupboard, open the cupboard, and so on). Where we might disagree—but I am not sure we do—is as to whether this needs a revision of or an addition to our definition of relevance. I believe it does not. Action is cognitively driven, and practical cognition is fully fledged cognition, so that there is no practical relevance without relevance as we have defined it. There is no way I can see in which something could be practically relevant without being relevant tout court, with variations in degree of relevance going together. In particular, the achievement of a goal modifies an agent’s behaviour only to the extent that it has cognitive consequences. From a personal point of view, when we act we want to bring about a change in reality. At a sub-personal level, the organism behaves so as to modify its representation of reality (which in normal conditions can be done only by actual changing reality). Wholly uncognized achievements don’t matter. Giovanni writes: “ I think that CPR applies more to full-fledged cognitive agents (having complex cognitive minds), while I would say that determining relevance is important for simpler agents, too, since all agents have to face the problem of action selection—and, as a consequence, of input selection.” One may imagine a simple agent for which the priorities are predetermined. It has, say, a danger detector and a food detector each with its own behavioural responses. Whenever some input matches the input condition of the danger detector, then the agent’s attention and behaviour is determined by this detector, and otherwise it is determined by its food detector. Well, yes, in a sense, its cognitive system can be said to be geared to the maximisation of relevance, but is it relevant to put it this way? We would get the same explanatory power if we just talked of fitness and efficiency, without ever introducing the notion of relevance: more bangs for the buck. In the human case however (and more generally in the case of agents where selecting input cannot be done by a simple prioritization of inputs according to their type, maximizing relevance becomes a complex problem for which specific solutions must have evolved.

  • Giovanni Pezzulo 4 February 2009 (20:07)

    Thanks again to Dan Sperber for his stimulating comments. I agree with him on almost everything, but I would like to add a couple of comments. One more point on ‘relevance and goals’. I see that in some sense, since beliefs are ‘in the service of’ actions and goals, the notion of relevance in CPR implicitly covers the fact that relevance is relative to them. Anyway, I wonder if it accommodates the fact that the same piece of information can be relevant or irrelevant depending on the goal context. You write: “Compare Joan telling you that she takes sugar in her coffee (1) when you are talking about sweetening coffee in general or (2) when you are preparing coffee for her. ” In the former case knowing that “I have enough sugar in my cupboard” is not relevant, while it is in the latter. [A more ‘dramatic’ case is when I want sweet coffee, too, and I know that I have not enough sugar for both us.] Overall, does the CPR discriminate the two situations (the same info being irrelevant or relevant depending on your/my goals)? I must confess I am unsure. [I would say that CPR discriminates the two senses ‘indirectly’, since in the latter case (not in the former) the same info triggers a lot of pragmatically-directed inferences.] My second point is about ‘relevance in simple (non cognitive) organisms’. Well, seems to me to be a terminological discussion, and I am ready to accept that it is futile to say ‘relevance of stimuli’, since ‘fitness’ or ‘efficiency’ are sufficient. Anyway, it is interesting to remark that (evolutionary speaking), the mechanisms for gathering relevant information (for the sake of belief revision) could derive from earlier (attentional) mechanisms which arose for gathering stimuli in a timely way. This is not sufficient anyway, and I fully agree with Dan’s idea that at some point “maximizing relevance becomes a complex problem for which specific solutions must have evolved.”