Thursday, November 11, 2010

Runeson on Perceptual Tools

In analogy with the planimeter and its user, our perceptual systems will be considered as a set of smart instruments which are (more or less actively) used by our intellect to get information about the environment.
     The study of perception would then be the study of the perceptual instruments. This may be subdivided into the search for the principles behind the function of the instruments, and the discovery of the physical realizations of these principles, i.e. how these instruments are actually built. The former would be the psychological part of the enterprise and the latter would be the physiological part.
     The relation between perception and cognition is modelled by the relation between the planimeter and its user. However, it is only the non-perceptual
functions of the user which are relevant to the model. Thus, our model does not contain a complete homunculus--only a cognitive, emotional, etc., homunculus. This should be a proper procedure when the focus of interest is on perception.

Sensory psychophysics
The study of relations between simple physical variables and experience is based on the implicit or explicit assumption that such relations are fundamental for the apprehension of "secondary" properties like causality and depth. Even when one finds the latter properties more interesting, one feels obliged to study the "primary" ones first. Mostly, such studies indicate that we are very bad at judging simple variables. This seems paradoxical when confronted with the delicate perceptual tasks we repeatedly perform in normal life.

I find the tool plus user story striking for its similarity to Fodor's later hypothesis of modules plus central system.  This parallelism then invites the comparison between Runeson's tools and Fodor's modules.  Right away, the principal difference appears to be that Fodor's tools begin with simple transductions, then performs computations to yield outputs that are "useful" to the organism.  So, the vision module will provide something like an accounting of the 3D layout of objects in the world.  It would be a rote mechanism, I assume.  By contrast, Runeson tools will (likely) be ones that provide direct pick up of things like the 3D layout of objects in the world.  These would be "smart" mechanisms.

If we think about Fodor's account of modules, Runeson's "paradox" goes away.  What the visual module does is take physical stimuli plus "assumptions" to generate hypotheses about the 3D layout of objects in the world.  It is because of the nature of module output that we are not so good at describing the inputs.  The modules don't simply regurgitate the inputs; they give us, instead, things like the 3D layout of objects in the world.

Note as well that Fodor's modules can respect some version of the Gibsonian idea that vision is not for detecting points, lines, etc.; instead it is for perceiving "meaningful" things in the environment.  It's just that Fodor's module yields the meaningful by indirect perception.

And, indeed, this seems to me to be a common way for tools to operate.  Often tools measure things we do not care about in order to give us information about things we do care about.  So, in fMRI, we do not really care about changes in blood oxygenation levels, we care about changes in brain activity and changes in blood oxygenation levels (at least putatively) just gives us part of the means to infer brain activity.  IR spectroscopy gives us a measure of the absorption of infrared light at different frequencies.  We don't really care about that absorption spectrum; instead we care about what it enables us to tell about the chemical structure of the tested compound.

Runeson, S. (1977).  "On the Possibility of 'Smart' Mechanisms" Scandinavian Journal of Psychology, 18, 172-9.


  1. Let me pursue the fMRI analogy, because I think it has an interesting aspect.

    Think of all the (external) resources required to make sense of the indirect measures from fMRI and IRS. These all involve a complex set of assumptions and transformations, all of which impact the validity of your interpretations of your measures. The process is delicate, fragile: if you aren't careful you end up detecting brain activity in dead fish, for instance.

    So yes you can run in theory things this way; but it's troublesome and fraught with risks of error and it requires a lot of checking by people. These are the kinds of flaws that get you eaten by a lion, to be crudely evolutionary for a moment, and so perhaps are unlikely to be favoured.

  2. But, it's not just that you *can* run things this way. Very often it is how one *must* run things. You cannot directly perceive brain activity, so you have to rely on indirect methods such as those by way of BOLD signals.

    You cannot directly detect fruit, which is why you have to rely on color. You cannot directly detect quasars, that's why you have to rely on tools.

  3. Well, you think you can't directly detect fruit. The whole point of Gibson 66 and 79 is to demonstrate that direct perception is a legitimate option.

    This is the assumption of poverty of stimulus I go on about sometimes. You only need indirect methods if the stimulus doesn't allow direct detection; if you can establish that the stimulus does allow for it, then you don't need the (fragile) indirect methods.

  4. My point in the last post is to indicate how the tool analogy works against Runeson. Tools typically work indirectly, so why not think the perceptual tools work indirectly?

    But, the real reason to think that perception is not direct is that it fails for the illusion cases, amodal completion, etc. It's basic empirical evidence, rather than this speculative stuff, that does in EP.

  5. Direct perception != errorless perception, and we've done the illusion stuff to death.

    And tools only work indirectly if that's how they're built. There's nothing all that indirect about a hammer.

  6. I've been reading what TSRM say about the bent stick illusion. Thin.

    Last I recall about the Mach bands, you were not giving an account of how the inks patterns of panel generate the kind of information you hypothesize in a panel 2 or a story about the pac-man, and telling me the TSRM will save the day.

    As I said, "Tools typically work indirectly." That nicely allows for exceptions.

  7. But indirect tools are rubbish (the point of extending the fMRI analogy). That's fairly important.