Friday 6 July 2012

Evolution vs.Specification (Specification V)

One of Gibson's key contributions was to reveal that it was possible for the optic array to specify a meaningful property of the world. Gibson insisted that specification existed between the world and optics (each property produced one unambiguous pattern, and thus the mapping is 1:1). Specification, said Gibson, meant direct perception was possible, because picking up that one variable meant perceiving the one property that caused it. 

Turvey, Shaw, Reed & Mace (1981) formalised this idea by describing how ecological laws governed which properties of the world could be specified and identifying that these laws allowed affordances into this set. Turvey et al (hence TSM, because Reed changed his mind later on) then insisted that, in order for perception to be direct, specification also had to exist between the optics and the perceiver; an organism should only use one variable per property, and thus the mapping from world to perceiver is 1:1:1. This is a very high bar, and was put in place to defend ecological psychology from the Establishment attack (Fodor & Pylyshyn, 1981).

Withagen & Chemero (2009) think that the 1:1:1 account is incompatible with evolutionary thinking, and they aren't hot on the 1:1 account either. Specifically, they think that any given species will show individual variation in it's members ability to use information, and that in many cases species will end up using sub-optimal solutions (two important elements of evolutionary thinking). The1:1:1 bar, they say, is implausibly high and a naturalised theory of perception (one that is compatible with evolution) will instead predict the common use of non-specifying information. They also claim that this does not stop perception from being direct, so long as you allow 'directness' to live along a continuum.

I think there are some important issues here, but I think this paper's presentation is problematic. It contains no analysis of any particular information or task, and instead is full of sentences such as 'it seems more plausible to us that' and 'it is possible that'. This comes off as the kind of woolly evolutionary thinking psychology is rightly scolded for. Gibson and TSM spent a lot of time trying to make us pay less attention to what might be and more to what is.My concerns are mostly along these lines, and once I get them off my chest I want to turn in future posts to some ideas for a research programme to pursue this all in more detail.

Problem for TSM #1: Individual variation in information use
Evolution implies variability within populations. For perception, W&C claim that this implies that different members of a population will probably use different information variables to solve the same task. W&C note that the TSM account insists that all members use the same information: the specifying information. These ideas are incompatible, evolution is bigger, therefore TSM lose;
...given the myriad functions of perception and their different degrees of importance to the survival and reproduction of the animal, it is quite improbable that there is minimal variation in what information is exploited...
Withagen & Chemero, 2009, p. 373.
They then go on to cite the various studies that demonstrate individual variation.

This, for me, is where I discover this paper has thrown the baby out with the bath water. W&C want to ground ecological psychology in biology, not physics (as TSM do in the 'laws' paper). But they've missed what TSM gained for the field; a method for actually answering this question for specific tasks. One of their key points was that not just any old property ended up structuring light, only properties that can be directly involved in the process of ecological optics (so 'shoeness' no, 'walk-on-ability' yes). Not all information is equal, either: some is available for extended periods of time, some only briefly. The amount of individual variation will therefore actually depend mostly on what information is available. The first place to go looking for the causes of variation, therefore, is in the ecological optics analysis of the invariants created by a given event, and this is where TSM shine. In addition, while individual variation is interesting, you can only interpret it in the context of an ecological optics analysis.

This means that the TSM research programme of identifying the local physical environment/event and the properties of that environment/event that are being projected into light is still worth pursuing. It also means considering the spatial and temporal stability of the resulting invariants as well as their relationship to the property in question in order to predict and understand individual variation and actual information use.Take collisions: not everyone finds the variable that specifies the mass ratio, relative velocity change. First, mass ratio is an odd property to want to know about and is not clearly related to the control of action. Second, relative velocity change is really only available around the time of the collision. Exit speed and scatter angle,however, are available for extended periods after the collision (and sometimes correlate to the odd property mass ratio, making the whole thing quite complicated from the first person perspective of the organism). The fact that organisms can fail to find the short lived specifying invariant in the collision task is interesting, but a key part of the explanation is still likely to be in the dynamics of the information itself (and, perhaps, variation in thresholds on the part of the organism). So there is a lot of work left to do here before these experiments live up to their hype from W&C.

(One side note here, though: I do think individual variation will show up in interesting places. Conditions such as schizophrenia, autism, developmental coordination disorder, and others are very commonly associated with perceptual deficits or difficulties. Perhaps this is the place to go looking for variation and for seeing what the consequences of genuinely poor perceptual performance is? Perhaps the effects are unlikely to be subtle?)

Problem for TSM #2: Sub-optimal solutions
W&C talk about the fact the biological systems do not always evolve the optimal solution to a task. Evolution is full of hacks and weird features that reflect the fact that it is a process that can only build on what it has available at the time. A given species cannot simply acquire a required feature unless the potential for that feature is within the range of the variability in the population. Given that this is such a ubiquitous feature of biology, W&C suggest we should also expect to see it in perception.

This may be partly true, but I think there is a key difference here. Humans will not suddenly acquire wings because of the limited variation in our body plans doesn't allow it. Our anatomy is too stable. Of course, this stability is not true throughout the body; our central nervous systems, for example, are extraordinarily flexible and only hold their 'shape' while the information flowing through them remains the same. Change the information, change the shape (the way tools are swiftly integrated into the brain's information about the body). One thing our brains therefore provide us with is the ability to very swiftly alter the capabilities of the 'back end' of all perceptual systems; this, in fact, is what most researchers think of when they talking about perceptual learning. So, in perception, we might actually have access to the kind of flexibility we need to dodge the evolutionary trap of the local minimum and wire up smart measurement systems for the necessary information. Finally, a place to go looking to connect ecological psychology with neuroscience?

Would we still expect individual variation in the information variables different brains can learn, though? Maybe, but not the way W&C suggest. All information is essentially the same kind of thing (kinematic patterns in perceptual arrays) and if you can detect one such thing, I would expect you to be able to detect any of them, in principle. Individual variation may sneak in when the required variables have limited stability (such as relative velocity change); small changes in sensitivity could lead to either being able to get or not get that particular variable. Another place this kind of variation might reveal itself is in elite sports: the difference between making the pro leagues or not may be partly due to your ability to learn to perceive the relevant information.

If this is true, what would we expect such a system to learn about over time? The specifying invariant, if there is one. This is because only specifying variables are always present and usefully informative during all examples of an event. Think about collisions again; exit speed does specify mass ratio within a limited sub set of all possible collisions, while only relative change in velocity does so for the entire set. Which variable you learn will be a function of your learning experience (as well as the stability of the information, etc); Jacobs, Michaels and Runeson did find some evidence for this, too. EJ Gibson took advantage of this fact for her ecological theory of perceptual learning; invariants are the only thing that can serve as targets for perceptual learning because learning takes time and only invariants are around long enough. Then, within the set of invariants from a given event, only the specifying variable works all the time. Given enough time and enough pressure to learn the thing that works, you would expect our nervous systems to become attuned to specifying invariants.

Now, all that said; there are some important caveats here. People may end up using something that's good enough if they don't have enough time (which is typical in most learning studies) or they don't have sufficient pressure to find the specifying invariant (as I suspect is the case in judgment studies like collisions and my phase perception studies). This certainly raises some complicated questions for the ecological approach - are these people directly perceiving the world, or not? If not, what are they doing? If they suddenly learn the invariant, have they switched from indirect to direct perception? From the point of view of the organism, what's the difference?

(A side note on this last point: W&C suggest that using non-specifying information doesn't remove direct access to meaning, it simply places it on a continuum where you are more or less in tune with the environment. They give the example of the information birds use to locomote along a clear path. This information presumably specified the clear path until humans invented windows. W&C say that while the information no longer specifies (i.e. sometimes it's wrong) the birds didn't suddenly stop having direct access to the path because that's a bit of a weird claim. I know Withagen still thinks this is quite a good example - but it's not. The issue is that birds hit windows because windows don't structure light in a way that specifies obstacle. The problem is a lack of information, not information which has changed meaning.)

W&C think evolution predicts we should use non-specifying variables. I think that because their analysis does not include any task-specific informational analyses, and because the empirical case they're resting on is much weaker than they think it is, they have not made their case. I do think that individual variation is interesting. I do think that people using sub-optimal solutions is interesting. I don't think that we have yet pushed anyone in these learning studies in a way that really tests people's commitment to using the non-specifying variables they might latch onto for some reason, so it is as yet unclear whether the typical pressures of day-to-day perception and action are enough to drive learning to find the specifying information. I also think that without detailed analyses of the dynamics of the information taken one task at a time, this argument has nowhere to go. The collision stuff is a good model of what I have in mind, but the judgement task is a problem.

TSM and Gibson gave us ecological optics so that we could identify the actual invariant structures created by an event or affordance in the world, and we have methods for probing what variables people actually use. This evolutionary argument is not based on a sufficiently rigourous analysis of what is, and while the general points about information use are valid, the conclusion - throwing out specification - is simply not yet warranted. Work to be done!

Next time
There are a couple more papers I could read, but I think I've done what I wanted to, namely familiarised myself with the arguments and the data and started to think about ways to take this question forward, empirically. My next post will sketch out some ideas, and I'll see if I can talk Rob, Tony and anyone else interested into working on a grant to fund them :) Looks like we'll need a neuroscientist, too, so if anyone interested in network theory approaches wants in, let me know!

Fodor, J. A., & Pylyshyn, Z. W. (1981). How direct is visual perception?: Some reflections on Gibson’s “Ecological Approach”. Cognition, 9,139-196. DOI 
Turvey, M. T., Shaw, R. E., Reed, E. S., Mace W. M. (1981). Ecological laws of perceiving and acting: In reply to Fodor and Pylyshyn (1981) Cognition, 9 (3), 237-304 DOI: 10.1016/0010-0277(81)90002-0 Download 

Withagen, R., & Chemero, A. (2009). Naturalizing Perception: Developing the Gibsonian Approach to Perception along Evolutionary Lines Theory & Psychology, 19 (3), 363-389 DOI: 10.1177/0959354309104159 Download


  1. I suggest that you don't dismiss W&C too quickly.

    If you are talking about specification, then you are taking a platonic view of information as you are talking about what, in principle, is out there. But perception is a pragmatic solution to a practical problem. It cannot deal in platonic ideals. A pragmatic gathering of information is going to be based on somewhat arbitrary choices that are tested for their usefulness. Even a pair of identical twins are likely to lock onto different invariants and come up with different solutions for how to perceive their world.

    1. Lurking in amongst this analysis (and in W&C) seems to be the idea that specifying variables are somehow different in kind; that they are harder to find, or harder to perceive, or just different in some way that makes them off in a category of their own. This simply isn't the case! Specifying variables are still just spatio-temporal patterns in perceptual arrays, and they are, by definition, more stable than any other information source. They are the thing that is always present in the presence of the dynamic property in the world and the right perceptual medium. This is as far from being some kind of platonic ideal as you can get.

    2. In many circumstances, one might expect the specifying variables to be even easier to detect (i.e., of lesser complexity) than some of the non-specifying variables. Some of the known examples, such as accelerating symmetric expansion, are quite elegant.

  2. Now, all that said; there are some important caveats here. People may end up using something that's good enough if they don't have enough time or they don't have sufficient pressure to find the specifying invariant.

    Having gotten my start studying animal behavior, I think these situations are more common than you are giving them credit for. I agree with Gibson's assertion that the normal state of perception is one in which animals have the time and ability to pick up the invariants, but more restrained situations are not uncommon. Animals make inconsequential perceptual mistakes fairly often in situations that are not particularly dangerous. Further, life and death situations often require responses to be quicker than would allow for thorough exploratory motions.

    Don't forget superfecundity. The vast majority of all organisms die before reproduction, and most die from situations that would have been avoidable given better perceptual-motor abilities. Life-history theory tells us that organisms will be specially tuned to the demands of their relationship with their environment at that point in their lives -- which necessarily means they are not tuned to other things. That isn't to say most organisms are not good at perceiving the things they perceive, but only that most organisms are sensitive to a quite limited range of phenomenon, and have a limited range of response.

    Even if full TSM criterion are the norm, exceptions still abound... and not just in the laboratory.

    If you want a fun example (because it involves, humans, dogs, and social interaction), what do you make of the situation in which a dog chases a ball its owner never released? Surely, nothing could be more obviously specified in the ambient optics, surely the dog is capable of picking up said information, and just as surely it does not... because it has already turned and started to run before the information indicating non-throw existed.


    P.S. I can't quite put my finger on it, but somewhere in this discussion there is a muddle between what it means for perception to be 'direct' and the particular solutions that Gibson and TSM offered to explain how that criterion could be met. This seems to create confusion (in the published literature) over where, and for what, specificity is crucial.