Tuesday, September 7, 2010

What's the "trick" here?

In some comments over at "Notes from Two Scientific Psychologists" here and here, I have been wondering about the Gibsonian claim that "Only the eye considered as a fixed camera can be deceived. The actual binocular visual  system cannot. (Gibson, 1979, p. 281)"  and that "all illusions are tricks".  I am curious about the snake illusion and certain auditory illusions.  What's the trick in these cases?

But certain cases of amodal completion raise the same problem.  The illusion is that some occluded object looks circular, when it is not.  What's the "trick" in these cases?

Now, I know that Gibsonians often worry about 2D drawn illusions, but at least some of the problematic amodal completion cases seem not to depend on the 2D character of the illusion.  Readers can verify this for themselves by creating circular "tiles" with parts cut out and see for themselves.  

Seeing, however, as I learned how to use Windows Moviemaker for my appearance on Philosophy TV, I thought I would venture to add some video showing the "tile" I made.  The video is not meant to replace doing the demo for one's self, but it was fun for me to make. How many philosophers add video to their blogs?  (It's a good thing I'm not dreaming of going to film school, but that's another matter.)


  1. Step 1 in any ecological analysis: what's in the information? There are two cases:

    1. When your hand covers the bit missing, there is no visual information about that bit. The information about what remains specifies a partially occluded circular object. You are currently in error, but you haven't been fooled; you're just missing some information.

    2. When I watch you occlude and reveal the gap, there is now visual information about an object with a notch in it being progressively occluded and revealed (occlusion was a big thing for Gibson, because it is extended in time and thus complicated for a non-information based account). It certainly appears fully circular once occluded, but that's because the information now available specifies such an object. This appearance, I think, becomes stronger over time as it becomes longer and longer since you saw anything to the contrary.

    There is, however, persistence; perceiving progressive occlusion involves information about a specific type of event, and for Gibson, we continue to perceive the object as it becomes occluded (think about a picket fence: as I pass behind it you don't see me pop in and out of existence, even if the pickets are quite wide; nor do you alternate perception-memory-perception). So you persist in perceiving the object you saw progressively occluded.

    This persistence isn't permanent, however: there is a dynamic at work here, a time course. Over time the object will appear more and more like a circular object, although you may still 'know' that it actually isn't (this last, I will admit, is a hazy idea for me and I'm not sure how to build it out of information, unless the persistence is longer than even I think it is).

    So there isn't a trick. There is visual experience, rooted in experience and information.

    One of the things that has occurred to me that's still missing from our discussions is time; you can be initially in error here but uncover your error with time and exploration, for instance. Until the error is uncovered, you are wrong about the actual shape of the object, but only because you had yet to detect any information about the non-circular bit. You then cease to be incorrect given time and exploration.

    I think Gibson wants to say you were never fooled, it's just that the information you have might simply be incomplete. But what you do know is accurate, within those bounds.

  2. Hi, Andrew, I won't respond to all of this now (since I need to think about it more), but on my walk this morning it occurred to me that Gibson would give a different analysis, one that I should have foreseen. It is this.

    The information available over the course of the whole video is not ambiguous. The information over the course of some smaller time intervals, say, from about 0:11 to 0:14, is ambiguous. The "trick" is temporally restricted viewing conditions.

    This seems to be the account to extrapolate from Gibson, (1979), 166-8.

  3. That's about the size of it. Moving from retinal images to flow fields gave Gibson a way to think about how perception can extend over time without a representational system stitching images together. Once time gets into the system, all kinds of things becomes possible, specifically the kinds of things an ecological optics needs to get off the ground.

    wrt to the eye's constant motion - it's always struck me that this is a hint. For retinal image based theories of vision this constant motion is a disaster that must be endlessly compensated for somehow. For Gibson, it's precisely what you'd expect to see because you need it so badly. Given how evolution is not known to favour genuinely stupid solutions, I've always felt this was a hint about who was thinking in the right terms.

    There are numerous recent examples from Geoff Bingham's lab (my PhD advisor) about the necessity of a temporal extent for things to work; calibration of reaches, plus some new stuff on shape perception. All very cool, and likely to come up on my blog at some point (note to self: write these posts :).

  4. "It certainly appears fully circular once occluded, but that's because the information now available specifies such an object. "

    How does on measure or determine the information the occluded object specifies? I take it that it can't be simply that we perceive a circle so the environment must be specifying a circle. Right?

  5. What I can see is a circular object with another object in front of it; that's what the information is at time t. I have no basis to perceive anything else until the occlusion is removed (time t+1) at which point there is now further information.

    Occlusion was one of Gibson's favourite examples; he talks a lot about it in the 66 and 79 books and will reference Kaplan displays a lot (I can't find a good example online, sorry). But he talks a lot about the information for the event 'one surface passing behind another'.

  6. "What I can see is a circular object with another object in front of it"

    Agreed, but my question is how you check the next part:

    "that's what the information is at time t."

    It sounds like Gibson's prediction is "If I perceive X at t, then that's because I get the information X at t". But, I'm trying to think of how to test this prediction. How can I get a handle on whether I get information X at t, apart from the fact that I see x at t.

  7. Via careful experimentation. It's hard.

    First, to be precise: if you perceive X at t, that's because you have information Y (which specifies X) at t. X and Y are not identical, but they must be lawfully related.

    Then, you have to define X carefully, and establish what it means to perceive X. This usually means 'exhibit the behaviour for which X is necessary', eg 'catch the fly ball' because you perceived 'where and when it was going to land'.

    Make your hypothesis about what Y is on the basis of what X is and how that could conceivably be projected into, for example, the optic array. For instance, you only have kinematic (motion) variables in vision, so the information can only have, say, position or its derivatives in it, not mass.

    Experimentally manipulate Y (however many candidates you have). In the presence of Y, the behaviour will be possible; in the absence of it, the behaviour will fail. When behaviour fails, you've found your information.

    The JoV paper (Fink et al) on the outfielder problem does this, as does a paper of mine, by perturbation method. I spent time in my paper laying out the methodological logic; it's online here.

    This stuff really is a huge pain in the ass to do right (VR is handy here, although of course it has it's own limitations). But when you break the real information, the effects are not subtle. None of this 'explaining 40% of the variance' rubbish; behaviour the requires the information you've broken fails. It's deeply satisfying :)

  8. But, can you tell me how to do this for the example I have? You've already said that one has the information that "It certainly appears fully circular once occluded, but that's because the information now available specifies such an object. " It's the amodal completion case where there seems to me to be some ambiguity. I don't think there is in the fly ball case.

  9. This is shape perception: a relevant action task might be prehension. Get people to perform reach-to-grasp actions under various conditions and see what happens as you mess with candidate information.

    Also there's no ambiguity; you either perceive an occluded circular object or an occluded non-circular object depending on what information you have had access to and when. There's no ambiguity about this, just varying levels of accuracy. Perception might not get tricked but it's not all-knowing.