Wednesday, September 22, 2010

Runeson, Nomic Constraints, and Coincidences

Runeson's appeal to physical constraints seems to me not entirely foreign to cognitivist approaches.  Perhaps that is another reason I find it so congenial, even if I am not convinced about its robustness.

In his discussion of the Ames Room, Runeson writes "Geometrically, the chances that an equivalent configuration would occur by random is therefore only 1 in a 100 million and that is for a very simple, barren case. For a room without the size restriction or with furniture and structured surfaces, the chances are many orders of magnitude smaller yet"  (Runeson, 1988, p. 299) and "One must then ask, do prevailing physical and ecological constraints suffice to exclude the kind of room shapes that would be projectively equivalent with normal rooms?  (ibid.)


At least some cognitivist accounts of amodal completion appeal to things like image statistics to account for the completion of the thing on the left and the non-completion of the thing on the right.

27 comments:

  1. Runeson’s not suggesting (as the cognitivist people do with image statistics type approaches) that the observer sees the rectangular room because it is the a priori most likely solution to an ambiguous proximal stimulus. This is the problem cognitive people have because they assume ambiguity and thus ‘most likely’ is all you ever get (just with varying levels of mathematical sophistication). He’s noting that, due to the non-accidental physical constraints operating in the world, there’s every reason to expect a smart perceptual mechanism* to develop that takes advantage of that constraint and thus not cope with the outstandingly unlikely circumstance in which the smart solution fails. Smartness becomes an increasingly likely solution once you understand the real odds of the whacky distal configuration actually are; note how bad the odds are just for a simple, uncluttered layout.

    This approach puts the contribution of likelihood into the process of building your smart mechanism, rather than the online act of perception. These are related but importantly different.

    *Runeson, S. (1977). On the possibility of "smart" perceptual mechanisms. Scandinavian Journal of Psychology, 18, 172-179.

    ReplyDelete
  2. All of this seems to me to be a priori reasoning about the most likely solution to a geometrically ambiguous.

    The variability of six-panel enclosures. For an example, consider a cubical enclosure made from six flat panels, with the near one having a peephole in its center. By changing the
    orientation of the panels to various oblique positions, and changing their shapes accordingly, a set of closed irregular
    hexahedrons can be generated. To restrict overall size variations, imagine that each panel can pivot only around its fixed
    center point. If the near panel with the peephole is kept in the same orientation, there are five panels to reorient, each one on two axes. Thus, we can say that the shape of the room has 10 geometrical degrees of freedom because, within limits, each can be varied independently of the others. With a resolution of 10 steps on each axis, we would get 1010 different possible rooms. For a given enclosure, how much of all this variation is
    permitted if its projected shape is to remain unchanged? It turns out that only 2 degrees of freedom remain because any reorientation of the far panel, for instance, forces specific
    reorientations of the other four panels. Projective equivalence requires that each corner travel on a fixed sight-line that
    extends from the peephole through the original location of the corner; hence, the reorienting of one panel creates new intersections between that panel and the sight-lines. These, in
    turn, define new locations for two corners of each adjoining panel. With center points already fixed, new orientations and shapes of all panels are defined. Thus, the static optic array at the peephole cancels 8 of the 10 degrees of freedom, which means that out of the 10 billion possible shapes, only 100 (really, 99) are equivalent configurations to a given enclosure. Geometrically, the chances that an equivalent configuration would occur by random is therefore only 1 in a 100 million—and that is for a very simple, barren case. For a room without the size restriction or with furniture and structured surfaces, the chances are many orders of magnitude smaller yet.

    ReplyDelete
  3. (Sorry about the formatting on that last one.)

    But, I don't think the physical or geometrical ambiguity is a mere assumption. It's often a simple demonstrable geometrical fact that multiple objects can have the same projection. That's what Ittelson types mean by the ambiguity.

    Now, Runeson seems to me to (at least at times) admit that there is ambiguity in this sense. He seems to admit that there are equivalent configurations (hence ambiguity in that sense), but that some configurations are "irrelevant". So, for example, there is the section entitled "
    The Irrelevance of Equivalent Configurations". It's not "The Non-existence of Equivalent Configurations". Moreover, he later writes,

    The actual existence of nonrectangular rooms may seem to provide evidence for ambiguity. Indeed, the purpose of constructing rooms that are projectively equivalent to rectangular rooms (e.g., the Ames' "L room" and "T room"; Kilpatrick, 196 Ib) was to drive home this point. However, there are two ways that distorted rooms, and equivalent configurations in general, could fail to be relevant: (p. 298).

    Now, he is not, however, entirely unequivocal on this score. He does seem to deny ambiguity as well here:

    As recognized also by some of his opponents (Epstein, 1977), the most important constituent of Gibson's approach is the rejection of necessary proximal ambiguity: the "doctrineof intractable nonspecificity" (p. 298)

    ReplyDelete
  4. Thanks for the paper reference. The paper is conveniently available via Google scholar.

    ReplyDelete
  5. But, I don't think the physical or geometrical ambiguity is a mere assumption. It's often a simple demonstrable geometrical fact that multiple objects can have the same projection. That's what Ittelson types mean by the ambiguity.
    Runeson is talking about how incredibly unlikely equivalent configurations are for a very simple example - turn that into a slightly more cluttered environment and the odds rocket. So sure, you can build an Ames Room, but it's not the kind of thing a smart mechanism need worry about coping with.

    ReplyDelete
  6. I think this is getting off track. You wrote, "This is the problem cognitive people have because they assume ambiguity" So, you are proposing that cognitive people *assume* ambiguity. But, this is not an assumption. It's an easily demonstrated bit of projective geometry that Runeson accepts.

    ReplyDelete
  7. But it's not something that an organism existing any time prior to the last 60 or so years would have to cope with. This post has already laid out the improbability of equivalent projections. This estimate also relies on objects with properties that are unlikely to obtain prior to human technology (e.g., flat panelled enclosures). Such configurations would be even less likely in the environment in which our perceptual abilities evolved.

    This means that there is no reason that a smart perceptual mechanism would have evolved to deal with these types of situations. You keep focusing on the possibility of such ambiguity, but do you really think that these situations would have been a driving force in evolution? Given the rarity of any ambiguity, why build a system based on inference rather than direct perception?

    Runeson was making the same point when he argues that typical definitions of ambiguity ignore nomic constraints on objects and enclosures in the environment.

    Being able to demonstrate one type of ambiguity in highly constrained conditions that are completely at odds with the way we typically experience the world doesn't warrant an assumption that perception must involve assumptions or inference. You must demonstrate the this type of ambiguity is actually problematic for organisms in the environment, that it cause enough trouble often enough to drive the evolution of a fundamentally different type of perceptual device than what Gibson proposes.

    ReplyDelete
  8. No, it really is an assumption. Cognitive psychology assumes poverty of stimulus and cognitive theories of perception are no different; it's the only reason you need to posit internal, supplementary representations. The Gibsons referred to these theories as 'enrichment', vs their own 'differentiation' approach; adding value vs learning to discriminate.

    Runeson accepts that there can be equivalent configurations because, well, you can build an Ames Room. But he's not accepting that this potential for ambiguity is interesting and the entire thrust of the article is laying out why (ie the sheer improbability of it means why evolve to cope with it?).

    But, this is not an assumption. It's an easily demonstrated bit of projective geometry that Runeson accepts.
    Another way to attack this point: Runeson's noting that it's a readily demonstrated technical possibility of one particular flavour of geometry (which can be ruled out as anything a smart mechanism need worry about because physical & ecological constraints make this technical possibility a staggering improbability.

    ReplyDelete
  9. Gerry, how does the Runeson story go for the amodal completion of the pac-man?

    ReplyDelete
  10. or, for that matter, why, according to Runeson, do we get amodal completion in the left figure above, but not the right one?

    I'm not objecting the his handling of the Ames room at this point. My challenge is to the robustness of the solution.

    ReplyDelete
  11. Poverty of the stimulus is not an assumption. There have been plenty of studies concerning what stimulus children get for language learning. There is an extensive discussion of Gold's theorem. You may not agree with what the evidence shows, but it's not as though cognitivists have spent 40+ years with an important empirical hypothesis and just assumed that it is true.

    So everyone agrees that there is this geometrical ambiguity and that there are equivalent configurations. Their existence is not an assumption. Then the argument is that the ambiguity does not matter.

    ReplyDelete
  12. "But it's not something that an organism existing any time prior to the last 60 or so years would have to cope with. This post has already laid out the improbability of equivalent projections. This estimate also relies on objects with properties that are unlikely to obtain prior to human technology (e.g., flat panelled enclosures). Such configurations would be even less likely in the environment in which our perceptual abilities evolved.

    This means that there is no reason that a smart perceptual mechanism would have evolved to deal with these types of situations. "

    But, this kind of argument seems very dicey to me. No organism existing any time prior to the last 60 or so years would have to cope with iPhones, but we see them perfectly well.

    ReplyDelete
  13. But, let me summarize a bit.
    1) I am, for the most part, willing to accept Runeson's account of the Ames Room.

    2) I am not so optimistic about his approach as a general strategy for dealing with geometric ambiguities or equivalent configurations. I think cases like the pac-man and the one above will not be covered by his theory of evolutionarily unnecessary interpretations.

    3) In this post, I am (mostly) taking exception to Andrew's claim that geometrical ambiguities are merely assumed by cognitivists. I don't see that at all. Now, it could be that there are lots of papers in which cognitivists do make this assumption, because they take it to be well-established that they exist. But, not every paper is going to defend first principles. At some point, one has to build new stuff.

    ReplyDelete
  14. Or, how about this? I think Andrew can agree that the *physical/geometrical information* is ambiguous, but might try to hold the Gibsonian line by maintaining that this physical/geometical information is irrelevant. Instead, the *Gibsonian information* is unambigous.

    Is it really all that much to come to the conclusion that a physical counterpart consisting of tiles could consist of either one spotted tile behind the homogeneous tile or two spotted tiles adjacent to the homogeneous tile? Those are two physical/geometrical possibilities, right? What's problematic in that?

    ReplyDelete
  15. But, this kind of argument seems very dicey to me. No organism existing any time prior to the last 60 or so years would have to cope with iPhones, but we see them perfectly well.
    Spherical cow alert: iPhones are perfectly well behaved objects which project their properties into optics in a manner that is a) appropriate to their shape and b) concordant with the rules the smart mechanism would have developed under.

    So everyone agrees that there is this geometrical ambiguity and that there are equivalent configurations. Their existence is not an assumption. Then the argument is that the ambiguity does not matter.
    You are way over-stating the degree to which ecological psych 'agrees' about these ambiguities. That they are physically possible to build but that they are ruled out by nomic constraints is an entirely different claim than the cognitivist "dogma of universal equivocality" )I'm keeping this phrase, it's great :).

    ReplyDelete
  16. But, you seem to me to be equivocating here. On the one hand, it is *physics* and *geometry* that makes it the case that the counterpart consisting of tiles could consist of either one spotted tile behind the homogeneous tile or two spotted tiles adjacent to the homogeneous tile? Those are two physical/geometrical possibilities, right?

    On the other hand, you suggest that there is some dogma to this. But, look, Runeson agrees that it is physically possible that there could be a hole in the table beneath the books. Similarly, there could be a hole in the back of an apple. Etc, etc. This kind of physical ambiguity is universal. That's no dogma. That's just physics.

    But, the Gibsonians are in some sense not supposed to care about physics. What matters to the perceiver is supposed to be affordances, Gibsonian information, ... I don't see how you are helping win over folks to the Gibson approach by needlessly waving about this misnomer of a "dogma of universal equivocality".

    ReplyDelete
  17. It's physically possible for all the atoms in your body to suddenly move 3ft to the left, but they don't because it's so improbable. Just because something is possible doesn't make it anything worth worrying about, and the 'smart' part of systems is that they only worry about what is reliably behaviour.

    Cognitive theories assume visual perception begins at the retina and thus that the signal is fundamentally ambiguous. This is indeed at the heart of cognitive theories (it's the only reason you need to invoke representations) and Gibson showed it's just an unwarranted assumption by being the first person to step back and actually ask if the signal was, indeed, ambiguous. He then found reason to think it wasn't.

    Please give up this weird 'Gibson hates physics' thing, though: the laws of physics are a happy part of how the world can project into and be specified by information. I've explained three or four times now what the actual issue is; physics is simply one of several sensible abstractions of the world and organisms organise their behaviour according to one of the alternatives (so not size, or eight, but perceived heaviness, for example).

    ReplyDelete
  18. "It's physically possible for all the atoms in your body to suddenly move 3ft to the left, but they don't because it's so improbable. Just because something is possible doesn't make it anything worth worrying about,:

    Ok. So, let's move on. How does this bear on the occluded pac-man? Is this so improbable that it is not worth worrying about?

    ReplyDelete
  19. The tile and pac man cases are interesting and, to my mind, reveal a subtle difference in what cognitivists and Gibsonians expect is going on. The tiles on the left are consistent with two obvious states of affairs, as you point out. For cognitivists, perceivers infer a connection between the flanked tiles because statistical regularities in the environment make this the best bet. Meaning is not in the environment, it is created by us as cognitive agents. For Gibsonians, we directly perceive a connection between the tiles because that is what this configuration means.

    Probabilistically, it is much more likely that two unconnected tiles will be offset, even if slightly. Looking back to the types of objects that were around during our evolution drives this point home (e.g., an animal passing behind a tree). And, as with the Ames room, the problems that arise from these tiles only obtain in highly constrained viewing conditions. Any instantaneous strangeness disappears when we are allowed to explore. This is the case for the pac man figure as well.

    The *design* of smart perceptual mechanisms reflect these regularities in the environment, obviating the need to use *inference* to explain why we perceive the Ames room as rectangular or the flanked tiles as connected. For Gibsonians, meaning is in the environment.

    As for Andrew's claim about cognitivists assuming ambiguities, I agree that it is quite prevalent in the literature, specifically on perceptual symbols systems and categorisation (Barsalou is a good starting point). Barsalou is quite influential, but he assumes sensation-based perception, where inference is necessary to cope with even simple transformations of the same object. This assumption really does dominate a large portion of the literature and I'm happy to supply references when they are more to hand. Of course at some point you have to stop defending first principles, but in this case their first principles are wrong (e.g., the belief that perceptual similarity can't cohere natural categories like animals). Even more problematically, because categorsation underlies so many other behaviours (e.g., memory, decision-making, language learning, stereotyping, etc.) getting it wrong here affects much of cognitive psychology.

    Poverty of stimulus for language learning (which is still equivocal) is not evidence for poverty of stimulus in perception. Language is weird and there is not an ecological account of how language is acquired. At the very least, there are critical differences between language in the environment and objects in the environment. While the acoustic properties of spoken language obey physical laws, the use of language, while regular, is not lawful. A noun doesn't specify its referent in the same way an information variable specifies a state of affairs in the world.

    The argument about evolutionary relevance doesn't rule out novel objects. Most things we've created conform perfectly well to major physical constraints in pre-modern environments. It shows when things don't conform to these constraints - consider how much trouble is caused by a large handle on a door (which clearly affords pulling) under a sign that says "Push". iPhones are actually ecologically very well designed - compare the ease of using an iPhone to the difficulty in changing the time on a first generation digital watch.

    ReplyDelete
  20. Gerry,
    Going back to your first comment. I do think you have the right idea, namely, provide a simple indisputable fact, e.g. that there were no Ames rooms 60+ years ago, in an attempt to confound the cognitivist or support ecological psychology (EP).

    But, the core of my reply is that this simple indisputable fact is not, by itself, enough to cause the cognitivist problems or support EP. That's the point about iPhones. There are other assumptions that have to be added. Andrew brings forward some additional assumptions that might help, but this immediately takes us beyond the immediately obvious facts.

    ReplyDelete
  21. Gerry,

    I'm 100% with you with your analysis of the situation in the first paragraph of your second post. But, what I'm working on is how one gets to an account of the meaning of the pacman that isn't ad hoc. I take it that we don't want the meaning to be whatever it has to be to make the experimental results come out right. Andrew tells me that I need to read the Fodor-Pylyshyn critique of Gibson, then the Turvey et al reply. That's on my reading list for early next month.

    Regarding ambiguities, part of what is rubbing me the wrong way is that claim that cognitivists are assuming that there are ambiguities versus cognitivists have a false belief about ambiguities. I am much more open to the latter claim than the former. The difference to me is that merely assuming some important empirical position is a kind of bad scientific practice, whereas having false beliefs happens all the time. Maybe this seems like a slight difference, but I am a philosopher after all.

    And, I don't doubt that Barsalou makes begins with stimulus ambiguity. And, as you anticipate, I think this is probably an instance of some sort of extrapolation. There's ambiguity over there, so there's likely to be ambiguity over here. But, that's not merely making an assumption. And I'm open to hearing how he's got a false belief regarding ambiguity.

    I also agree with you that poverty of the stimulus claim vis a vis language is the same as the poverty of the stimulus claim vis a vis perception. There is a lot to be teased out empirically.

    ReplyDelete
  22. Gerry,
    "And, as with the Ames room, the problems that arise from these tiles only obtain in highly constrained viewing conditions. Any instantaneous strangeness disappears when we are allowed to explore. This is the case for the pac man figure as well."

    Unlike what Runeson's "opponents" found with the Ames, I don't think the pac-man illusion goes away with free exploration. I have a bit of video here: http://theboundsofcognition.blogspot.com/2010/09/whats-trick-here.html

    ReplyDelete
  23. The difference to me is that merely assuming some important empirical position is a kind of bad scientific practice, whereas having false beliefs happens all the time.
    I agree there's a difference, I just think cognitive science is (on average) guilty of the worse one.

    Unlike what Runeson's "opponents" found with the Ames, I don't think the pac-man illusion goes away with free exploration.
    The exploration Gerry is referring to would include removing the occlusion blocking the pac-man part; the occlusion removes the information about that part so naturally a system resonating to information will lose access to that fact after the information is removed.

    I am thinking about this example, though. I think I have an angle but I want to mull it over a bit.

    ReplyDelete
  24. Ah, but even with exploration the Ames illusion returns from that one particular perspective, just as the pac man always looks whole when the chip is occluded. Now, if these illusions were based on inference, then we should be able to teach people to not see the illusions. This isn't possible because the illusions are based on information rather than assumptions.

    The iPhone is a slightly different case because it obviously requires additional information (e.g., that exploring essentially an unadorned rectangular solid is likely to produce interesting results if there is an apple logo on it). Lots of our behaviour requires additional conditioned learning and inference certainly has a place as well. The Gibsonian position is just that is isn't necessary to invoke inference to explain perception.

    As for the false belief / assumption, I think your distinction is accurate. The problem is that cognitivists believe that the old, sensation-based model of perception is correct.

    ReplyDelete
  25. Gerry, I thought that the experiment Runeson was criticizing indicates that the Ames illusion (mostly) disappears after non-static viewing.

    But, when you write "Now, if these illusions were based on inference, then we should be able to teach people to not see the illusions," there is a *very* familiar answer to this in the cognitivist literature. It is that perceptual inference is not based on everything a person knows; it is instead based on a set of "informationally encapsulated" processes. This is what is at the core of Jerry Fodor's Modularity of Mind. So, you can have an "informationally encapsulated inferential approach" and an "informationally unencapsulated inferential approach". The refractory character of visual illusions argues for the encapsulated approach. (I can provide a ref to Fodor, if necessary.)

    Gerry, regarding "sensation-based models of perception" I can tell that from comments by Andrew that there is likely to be a long story about what's wrong with these. But, I've been putting off hearing it. Do you have a (preferably short) thing to read on that?

    Thanks,
    Ken

    ReplyDelete
  26. Two comments to back up my new fellow traveller up there:

    1. The paper showed that the magnitude of the illusion is almost entirely reduced by free viewing: Gerry's pointing out that when you return to the 'correct' location for the Ames Room illusion (the point in space the illusion is designed for) the effect is perfectly compelling again. To the best of my knowledge you can't learn to not see it from that one place.

    2. I always wondered how you make an informationally encapsulated Fodorian module. Given that at some point it had to be 'leaky' in order to develop the appropriate structure (either in learning or evolutionary time) I always wondered why you'd ever encapsulate it again. But that's mostly just a standard ecological bitch about Fodor :)

    ReplyDelete
  27. 1. Ok. My mistake on the return to static viewing of the Ames room.

    2. Informationally encapsulation is conceptually no more mysterious than is a well-programmed sub-routine in a structured computer program. In such a sub-routine, the sub-routine is set up so that it only takes certain information as input and provides only certain information as output. Information flow is restricted, since one does not want what is going on 'inside' the sub-routine to influence what is going on 'outside' the sub-routine. The well-programmed interface controls the leakage via "gates".

    Further, informationally encapsulation is physically no more mysterious than are such facts as that some neurons don't connect to others.

    ReplyDelete