Sunday 21 May 2017

Ecological Information Is a Perceptual Mapping That Tracks Evolutionary Fitness

Interface theory in cartoon form. Thanks to Louise
Barrett for reminding me this exists :)
In my last post I was thinking out loud about some ecological lines of attack on interface theory (Hoffman et al, 2015). The first line of attack (Hoffman et al mischaracterise Gibson) fell over eventually; they don't quite go at it right, but using ecological information does fit their definition of a naive realist perceptual strategy ( 'a perceptual strategy for which X [perceptual experience] ⊂ W [the world] and P [the perceptual mapping] is an isomorphism on this subset that preserves all structures on W'; pg 1483). The second line of attack (everything they say about veridicality vs fitness applies only to inferential, constructivist theories of perception and Gibson's not playing that game) is true but not that interesting or convincing to anyone with established views.

Thanks to chats on Twitter (thanks Greg!) and applying the basic move of 'those aren't working but IT is still weird, what's next?', my new line of attack relates to a result from their simulations.

Hoffman et al pit various perceptual strategies against one another in evolutionary simulations, and find that
strict interface strategies that are tuned to fitness routinely drive naïve realist and critical realist strategies to extinction
...
The only situation in which realists have a chance against interface strategies is when payoff varies monotonically with resource quantity, i.e., when truths and payoffs are roughly the same thing.
pg 1486
This happens because of the nature of the simulations; the payoff function literally rewards fitness and not veridicality, so when the two are not the same veridicality cannot win. Hoffman et al defend this by saying
But we cannot expect, in general, that payoff functions vary monotonically with truth, because (1) monotonic functions are a (unbiased) measure zero subset of the possible payoff functions, and (2) even if they weren’t, the ubiquitous biological need for homeostasis militates against them. Thus, we cannot expect, in general, that natural selection has tuned our perceptions to truth, i.e., we cannot expect our perceptions to be veridical.
pg 1486-7
Their homeostasis example is about 'amount of water'; the payoff for water is not a monotonic linear function of amount of water; a little and a lot are both bad for you.

As I noted last time, Hoffman et al do get one thing wrong about Gibson; they claim he wanted veridical perception and tried to make it adaptive. Actually he wanted adaptive perception and he then identified some of the real parts and processes that form a mechanism that can support this and it just so happened to produce a perceptual experience with a form that was veridical to a subset of the world, specifically the form of the information. Gibson's approach (and the critical formal defence of this approach by Turvey, Shaw, Reed and Mace, 1981) is in fact a hypothesis that there is at least one payoff function that varies monotonically with truth (kinematic information that specifies the dynamical world) and that this is the mechanism that shapes our perceptual experience of the world. The ecological approach since then has been the empirical investigation of this hypothesis, with many successes.

One example of how this works is coordinated rhythmic movement. The classic movement phenomena are that movements at 0° are more stable than those at 180°, while those elsewhere (especially 90°) are unstable. At the level of relative phase, there's nothing to drive this asymmetry. The asymmetry is present at the level of the information, relative direction of motion. The argument embodied in Bingham's model (reframed for the current context) is that the form of perceptual experience, as measured by the gold standard of the form of action, is a naive realist mapping onto the perceptual information supporting the perception of relative phase. Learners do not learn this variable because of it's veridicality, however; they learn to use it because it works and it happens to be veridical. Work on the use of non-specifying variables shows that 'truth' is not the driving force but 'can it drive adaptive behaviour?'.

This, I think, is the best line of attack on Hoffman; reframe the ecological hypothesis to illustrate exactly how it opposes this latest iteration of constructivist theories of perception to show that the one case where realist interfaces work is an option and showing empirical promise.

Comments on this idea welcome!

No comments:

Post a Comment