Sunday, 18 April 2010

"Smart" perceptual mechanisms

I mentioned yesterday that James' theory about pure experience being all there is, and relations being legitimate objects for perception does at one level sound crazy. Surely physics is where it's at - if you aren't talking about a physical variable, how can you be talking about anything real?

 Ecological theorists have one good example to hand, courtesy of a 1977 paper by Sverker Runeson, that might help: the polar planimeter (pictured). This is a device that measures area directly, rather than measuring the 'simpler' physical unit length and then performing the necessary computation. Runeson uses this device as an example of a 'smart' mechanism, and proposes that perception might entail such mechanisms.

Relations and higher order properties seem like the kind of thing that cannot, by definition, be the primary objects of perception. After all, a relation only exists between at least two things: there are therefore a minimum of two things to be detected (the two objects) and the relation between them must then be inferred or added in. Surely simple things must (logically) be measured before any composite?

Area is such a higher-order variable (requiring two lengths) and it is a specific relation between these lengths (implemented mathematically with the multiplication operation). Planimeters do not measure any lengths, however: they measure area (of any closed shape, even ones with irregular boundaries) directly. They are comprised of two arms (the pole arm and tracer arm) hinged next to a calibrated roller. The user simply fixes the free end of the pole arm outside the area to be measured and traces the shape with the free end of the tracer arm. Once you have traced the entire shape, the roller has rotated by an angle directly proportional to the area. Suitably calibrated (with sensible lengths) this angle can be identical to the area.*

There are several important lessons:
  1. Simplicity is relative to the measurement device: length may seem intuitively to be the simpler unit, but to get lengths from a planimeter requires an absurd amount of computation (using mathematics that possibly were not invented until after the device was). Bernstein (1967) talks a lot about this too.
  2. The planimeter is "smart": it's not a general purpose "rote" device the way a ruler is. It can only do one thing (measure area) - however, it does this fast, efficiently and reliably by taking advantage of certain stable characteristics of the task at hand. These allow shortcuts that only work for a given task, but work very very well for that task.
Runeson then simply suggests: what if perceptual systems are smart? What if they evolve, not according to universal laws of physics but according to the local, biological, ecological conditions, and evolve to take advantage of local stability to take clever short-cuts? This becomes a possibility once you consider what organisms had available to them: the information described by Gibson (1979).

This is an excellent article and well worth the quick read. For things you have to do repeatedly, under similar conditions, often-times the best design solution is a custom solution. Gibson shows us a way in which perception and action can show the required stability; Runeson then simply suggests evolution is likely to have taken advantage of this. And why compute, when all you need is a correctly built and calibrated device?

RUNESON, S. (1977). On the possibility of "smart" perceptual mechanisms Scandinavian Journal of Psychology, 18 (1), 172-179 DOI: 10.1111/j.1467-9450.1977.tb00274.x
*This kind of thing is a standard trick in geometry: you can simplify a lot of trigonometry by rescaling your data to a unit circle (with radius = 1). The sine of an angle is the length of the opposite line over the radius (ie it is directly proportional to the opposite length); if the radius is 1, this reduces to sin(angle) = the length. In trigonometry this trick is implemented via a computation; the planimeter can simply implement it by it's construction.

1 comment: