Non-Specifying Variables in the Perception of Collisions (Specification III)

Part of Withagen's critique of specification and whether it's necessary to underpin direct perception is a brief review of some empirical literature that shows people using non-specifying variables. I want to spend a few posts reviewing these, because all good potentially sensible ideas need data to confirm whether they're right or not.

First up, the perception of relative mass after a collision. Events in the world are dynamic, that is, they involve motion caused by a pattern of underlying forces. Perceptual systems want access to the underlying dynamics of events, because this is the level at which the event is defined (Wilson & Bingham, 2001). However, perceptual systems can only detect kinematics, that is, motion - this is the perceptual bottleneck  (Bingham, 1988 and this note on dynamics and kinematics). We can only perceive the underlying dynamics of an event, according to the ecological approach, if we can detect motion that is specific to that dynamic. Runeson coined the phrase kinematic specification of dynamics (Runeson & Frykholm, 1983) and investigated whether there were such kinematic patterns and whether we can detect them. Working with Claire Michaels and David Jacobs, he has also investigated the use of non-specifying variables.

Runeson uses judgments of relative mass as an exemplar task. He simulates a collision between two moving balls of varying mass; their behaviour after the collision reflects which one is heavier. There are several kinematic variables available after the collision, including exit speed and scatter angle (the angle between the ball's original and new trajectory). These do not specify relative mass, and how they correlate with it depends on details of the collision. They often feature in cognitive models of judgements in this task, acting as cues refined by heuristics (eg Gilden & Proffitt, 1989).

Runeson & Vedeler (1993; Runeson, 1995) identified a kinematic variable that specifies the mass ratio: the relative amount of velocity change. They then demonstrated that participants used this invariant rather than any of the cue-heuristic approaches. One issue with this study was the fact that all the observers had at least one experiment's worth of experience with the task, and Jacobs, Michaels & Runeson (2000) therefore investigated whether there was any perceptual learning in the task that might be causing different labs to find different information being used.

8 naive observers performed baseline and post-training judgements of mass ratio, with extensive training in between (288 trials split over 4 blocks and two days). Participants improved with training (their judgements became more accurate and less variable); the question was then, what had changed to support this learning?

Jacobs et al correlated the judgements people generated for each collision with the answers each of 5 possible variables would predict; the invariant 'relative amount of velocity change', (INV), exit speed difference (ESD), scatter angle difference (SAD), exit speed and angle differences combined (ESA) and a heuristic model from Gilden & Proffitt (1989; HM). They looked at these correlations in each block of trials (1=baseline, 2-5 = training, 6 = post training), and the results are in Figure 1.

Figure 1. Correlations between judgments and various sources of information about the mass ratio, for each observer across all sessions (Figure 5 from Jacobs et al, 2000)

The first thing to note are the individual differences; each observer starts out most likely relying of different information (e.g. Observers 2 & 4 both begin using exit speed). Most of the other observers, are using either the invariant or the linear combination of exit speed and angle (ESA); part of the problem here is that these variables are themselves highly correlated (r=.93) and thus it's not possible to disentangle them from these data (but see below). The second thing to note is that with practice, most of the observers were producing judgements with high correlations to the specifying invariant. Finally, the heuristic model rarely outperformed the other variables. 

We therefore have evidence of different people using different variables to judge the same event, and most of these variables did not specify the dynamic property being judged. With training, however, people tended to switch and begin to rely on the invariant (or a non-specifying variable that happened, for these collisions, to correlate highly with the invariant).

De-correlating the candidate variables

Jacobs, Runeson & Michaels (2001) ran a follow up study to cope with the problem that ESA and the invariant were highly (.93) correlated. There simply may not have been enough instability in performance controlled by ESA to drive further exploration of the space for better information, so Jacobs et al created some.

Experiment 1 tested baseline performance using the collisions from Jacobs et al (2000), then trained people with collisions from one of two sets. The global constraint set were collisions where the correlations corresponded to the set of all possible collisions, and the local constraint set were collisions from a limited set where all the variables specified the mass ratio. Participants in the global constraint group replicated Jacobs et al, and all changed variable use to either the invariant or ESA (which correlated at .89 here). Participants in the local constraint group never changed variable - whatever they began using, they stuck with. They therefore ended up using different (but in their experience, equally effective) information variables and didn't find the actual invariant.

Experiment 2 then trained people on sets of collisions in which the non-specifying information variables either did not vary over collisions between balls of different masses (no variation) or had no correlation to the actual mass ratios (zero correlation). A random condition randomised the precollision velocities and the various non-specifying kinematic variables had intermediate correlations with the actual mass ratio. In the no variation group, most observers came to find the invariant during training, but interestingly often switched back to, say, exit speed in the post-test when there was useful variation and correlation for this variable again. They therefore haven't come to rely on the invariant, although they had come to detect it. The zero correlation training group stopped using  non-specifying variables during training where they were not informative about mass ratio but didn't always succeed in finding the invariant; if not, their performance simply remained poor throughout. As in Experiment 1 and Jacobs et al, the random group often didn't find the invariant because the non-specifying variables correlated fairly well with mass ratio.

Finally, Experiment 3a trained people in a zero correlation condition where only one alternative (either exit speed or scatter angle) had zero correlation with mass ratio. This meant each collision set had the invariant and one other fairly good (.75) information source. The speed zero correlation group all stopped using exit speed and mostly found the invariant; the angle zero correlation group weren't using scatter-angle to begin with and never tried to; in addition they mostly didn't switch from what they started with because it was either the invariant or exit speed (.75 correlation) and both worked well enough. Experiment 3b took people from the angle zero correlation group who were still using exit speed and trained them again with a speed zero correlation training set; here, 3 of the 4 observers switched variables.

The lessons here: different observers began using different variables which correlated to varying degrees with the dynamic property being judged, relative mass. People were very sensitive to the collision ecology in which they were operating - after training, people typically used an informative variable, but this was not always the invariant. When other information was available and sufficiently correlated to relative mass, people often settled on the non-specifying variable. When prompted to continue their search by a change in collision ecology, however (Jacobs et al, 2001, Expt 3b) people were able to continue their search and typically found the invariant.

Summary

The take home message: people learn according to the local constraints of the event ecology and will therefore not always find the specifying invariant. How we perceive our environment depends on our personal developmental histories, and we use whatever works well enough to succeed under the constraints at the time. Change the constraints and the system will search for new information, but without that drive, the perceptual system will settle into locally optimal but globally non-specifying solutions. The scope of our experience is critical.

While the perception of relative mass is a slightly odd task, there is a clearly defined and available specifying variable so the task has the right kind of ecological validity (the definable kind!). The fact that people probably don't tend to do this task if left to their own devices makes the necessity of some training unsurprising, however. Of course, this is true of everything, even the things we do get up to on a regular basis; we had to learn how to do all of it. This task is therefore an okay model task for the question at hand.

That said, this is just a judgement task, and while it's a legitimate place to start it's not an action measure of perception and is thus an indirect way to evaluate perception. I've used judgments of relative phase before to measure visual perception of phase information; the results basically work out, but are very noisy and the training without action takes a very, very long time to work (up to 1500 trials, compared to about 40 or 50 for the action learning task). So while these data are relevant to the case Withagen is making, it's far from the kind of data required to topple something like specification and are, at best, a hint of something worth looking at further.

References

Gilden, D. L., & Proffitt, D. R. (1989). Understanding collision dynamics. Journal of Experimental Psychology: Human Perception & Performance, 15, 372-383.

Jacobs, D., Michaels, C., & Runeson, S. (2000). Learning to perceive the relative mass of colliding balls: The effects of ratio scaling and feedback Perception & Psychophysics, 62 (7), 1332-1340 DOI: 10.3758/BF03212135

Jacobs, D., Runeson, S., & Michaels, C. (2001). Learning to visually perceive the relative mass of colliding balls in globally and locally constrained task ecologies. Journal of Experimental Psychology: Human Perception and Performance, 27 (5), 1019-1038 DOI: 10.1037//0096-1523.27.5.1019

Runeson, S., & Frykholm, G. (1983). Kinematic specification of dynamics as an informational basis for person and action perception: Expectation, gender recognition, and deceptive intention. Journal of Experimental Psychology: General, 112, 617-632.

Runeson, S. (1995). Support for the cue-heuristic model is based on suboptimal observer performance: Response to Gilden & Proffitt (1994). Perception & Psychophysics, 57, 1262-1273.

Runeson, S., & Vedeler, D. (1993). The indispensability of precollision kinematics in the visual perception of relative mass. Perception & Psychophysics, 53, 617-632. 

Wilson, A. D., & Bingham, G. P. (2001). Dynamics, not kinematics, is an adequate basis for perception – Commentary on Shepard (2001). Behavioral and Brain Sciences, 24(4), 709-710. Download