[Martin Taylor 2010.07.05.00.05]
[From Rick Marken (2010.07.04.0840)]
Martin Taylor (2010.07.04.10.15)_-
As a passing comment, very little of psychophysics is concerned with
magnitude estimation. Most psychophysics has to do with the capabilities of
the processing channels, such as timing effects, noise limits,
cross-interference, cross-support, and the like.
Yes, but since all psychophysical research is based on a an open-loop
causal model of the systems under study, those who study psychophysics
(which used to include me;-) can't really be learning much about the
"capabilities of the processing channels" of these systems using
conventional methods. Agreed?
As I said earlier [Martin Taylor 2010.07.04.17.02], definitely not agreed.
I've been thinking about this more and more over the course of this thread, and I'm beginning to see that a good part of what is taught as good experimental practice in psychophysics actually is performing the test for the controlled variable, whether the experimenters know it or not (and most don't). "Good practice" is intended to ensure that the subject is trying to do what the experimenter wants, and that the subject's output mechanism doesn't interfere with measuring the properties under study (unless, of course, the output pathways themselves are of interest, as sometimes is the case; in that case, the perceptual input is usually made very simple).
Of course, most practitioners will not realize that they are doing "The Test", any more than most people realize they are always controlling their inputs, but it seems to me that good practice nevertheless usually, if not always, results in the experimenter performing the Test for the controlled variable. Experimental technique has developed by normal evolutionary methods -- keep what works, change what doesn't -- and in my opinion that evolutionary development has approached what would have been prescribed from first principles by PCT analysis.
In any experiment, both the experimenter and the subject are controlling many variables, but from the experimenter's point of view only two of the subject's controlled variables matter very much, namely (1) that the subject is controlling for perceiving that the experimenter perceives that the subject is performing as the experimenter wishes, and (2) that the subject is in fact acting to control against disturbances in the way the experimenter intended. The controlled variable and reference value in (2) are part -- but only a part -- of the output structure of (1), but could take the experimenter's desired values in the absence of (1). The experimenter's perception of both of these variables is controlled, meaning that if the experimenter does not perceive their state to be as wished, the experimenter acts to bring their state toward their reference values. (The experimenter also has quite a few other controlled perceptions, mostly relating to the proper functioning of the equipment, but we don't need to be concerned much with those).
Even if (1) is satisfied, (2) may not be, and good experimental technique ordinarily implements "the test for the controlled variable" to determine whether the subject is actually doing what the experimenter hoped for (controlling the intended variable at the intended reference value). Poor experimental technique does not. Poor experimental setup may make it impossible (e.g. a setup in which the subject finds it hard to provide an environmental output precisely related to the output of the control unit that controls the experimental variable).
I'll give an example of a real case in which (1) was satisfied, but (2) was not, because the subject misinterpreted what variable the experimenter wanted the subject to control. A colleague of mine, when he was an undergraduate, was a subject in an experiment that was intended to assess tracking ability, as measured by the proportion of time the subject could keep a stylus on a metal spot on a rotating disk (maybe there was mirror viewing or some such). My colleague interpreted the instruction "Keep the stylus on the small metal spot as long as you can" as an intelligence test, and instead of tracking the spot while the disk rotated, he dismounted the disk and laid the stylus on the metal spot. He was controlling a variable he thought the experimenter had intended, but it was not the actual variable the experimenter had intended. The experimenter could test for this by observing first the record of the tracking error, and when that was obviously not a track any but a superhuman could attain, he then directly observed the placement of the disk and stylus. The experimenter might not have been able to know what variable my colleague was controlling for, but "The Test" showed clearly that it was not the variable the experimenter had intended. (I would say that this experimenter's technique was not very good, but at least it was adequate to ensure that my colleague's data were not included in the experimenter's interpretation of the results).
"The Test for the controlled variable" has several components, which can be stated in the form of questions.
1. Is there a possibility for a control system to perceive the effect of a disturbance to the putative environmental correspondent of the controlled variable? (Let's call that the ECV -- "environmental controlled variable" -- for short, even though the environmental variable is NOT the controlled variable).?
2. Is there a possibility for the output of a control system to influence the ECV?
3. When the disturbance is applied, is its effect on the ECV less than would be expected?
4. Can the lessening of the effect of the disturbance on the ECV be attributed to the output of the putative control system? In other words, can the output of the control system be correlated with the disturbance?
[Parenthetical note: In the past, we have used "CEV" or "Complex Environmental Variable" to represent the function of environmental variables that truly corresponds to the perceptual input function of a control system. I use "ECV" here to distinguish a guess as to the CEV from the actual CEV.]
In that form, "The Test" only determines that the perception corresponding to the ECV is correlated with the controlled perception. It does not determine that the corresponding perception IS the controlled perception (as Rick showed in his comparisons of different possible controlled variables in when the length and width of a rectangle were independently disturbed and the subject was trying to control (I think) "size" -- it's not on Mindreadings.com).
One of the well known issues of experimental technique in psychophysics is to determine whether the subject is actually doing the task the experimenter intends -- controlling the "right" variable. The "clever Hans" phenomenon illustrates an early example of where no such test was done until fairly late in the game, when the horse Hans was shown to be controlling for a relationship with the handler's movements rather than to the intended "stimulus". In psychophysical detection studies, experimenters go to considerable lengths to ensure that nothing correlated with the "stimulus" is perceptible to the subject -- no clicks when a signal goes on, no visible signs of experimenter movement (and the experimenter should not know what stimulus is presented on a particular trial, anyway). Good experimental technique is, in part, aimed at ensuring that the disturbance provided by the stimulus in only to the variable intended by the experimenter.
In psychophysical experiments, the controlled variable is usually quite simply described, and it is usually a relationship variable, such as to match a selection of a response with a particular value of the stimulus presentation -- e.g. "did the left or the right light come on?". Since the question of interest is some property of the perceptual channel that leads to one input of the relationship control perceptual input function, good technique requires that the pathway from the output function of the relationship control system to the outer world be as simple as possible. In particular, since most perceptual pathways have much higher bandwidths than the muscular part of the output pathway, if the perceptual bandwidths are under study, it is often impractical to have feedback through the external environment (which means that tracking methods cannot be used). The feedback pathway must, in those cases, be internal (in imagination), and not visible to the experimenter.
So now we have to ask: If the Test requires that the output of the control system be observed and checked for its relationship with the disturbance, but the experimental question cannot allow for externally observable tracking, how can the Test be performed in practice? The answer must be that the externally observed output has to be very nearly perfectly correlated with some phase (ordinarily the final phase) of the output of the tested control system. If that can be guaranteed -- and good experimental technique demands that this be tested initially by making the disturbance-creating "stimuli" such as to make the control very easy -- then the output visible to the experiment can reasonably be taken as a surrogate for the output of the control system whose input pathway is the object of study.
As an example, imagine a test of the ability to detect a 500 Hz tone burst in the presence of some masking noise. Good technique suggests that the experimenter should first present to the subject similar tone bursts that are loud enough for the subject to hear them clearly, and to ask "Do you hear the tone bursts?". The subject may answer that she is not be able to hear them, or that she does not know what it is that should be heard. That would generate error in a control system in the experimenter, to be corrected, perhaps, by checking the wiring or by instructing the subject more carefully, until the subject says: "Yes, I hear them clearly now".
Next, the experimenter checks whether the subject knows what is to be controlled, and how to report the output of the control system: "You will hear four noise bursts, and in one of them you will hear the kind of tone I just showed you. Your job is to determine which of the four bursts had the tone in it, and push the button with the corresponding number." So then the experiment tries giving four bursts, with a clearly audible tone in one of them. If the subject presses the correct button, the experimenter's relevant control system experiences no error, but otherwise, the experimenter does whatever is needed to get the subject to press the correct button and have the result recorded. Usually the experimenter would provide several stimuli to be sure that the output pathway is accurately representing what the subject should have matched to the interval containing the perceived tone.
From the subject's viewpoint, the loop is closed by the experimenter's telling him that the response was correct or that it was wrong. When the subject has learned to press correctly and consistently the button that the experimenter wants to have associated with each presentation interval, neither the subject nor the experimenter will have error in the relevant control systems.
Next, the experimenter will decrease the loudness of the tone until it is quite difficult to hear, and make sure that the subject presses a button on every trial, by seeing whether this happens and letting the subject know when a trial occurs but no button response was observed. Maybe the subject says: "But I did push the button every time" or "I didn't push a button because I didn't hear a tone". In that case, perhaps the subject needs instruction about pushing every time, or about pushing hard enough, or maybe there is a faulty connection in the wiring. Either way, there's an error that needs to be corrected in one of the experimenter's control systems.
Finally, when the tight relationship between observable button pushes and the output of the subject's relationship control system has been established, the experiment proper can begin, and the experimenter can start investigating whatever property of the input perceptual pathway is of interest.
I don't think there would be any difference to this technique if the experimenter realized that what he was doing was making sure of the controlled variable, and of the relationship between the output of that control system and the visible environmental output by the subject. It's possible, however, that in some cases a knowledge of PCT might alter the experimenter's interpretation of the results.
···
=========
The next question is whether this kind of experiment can provide any information about the perceptual pathway between the physical sensor and the corresponding input to the control system for the variable being controlled. For some properties, it can. For example, in the kind of experiment used as an example above, if the noise burst has a spectral gap around the 500Hz tone, changing the upper and lower bounds of the gap will influence the correlation between the stimulus and the response for some intensities of the tone. If the gap is narrow, to be heard the 500Hz signal tone will need to be louder than if the gap is wide. Varying the bounds of the gap tells the experimenter something about the acceptance bandwidth of whatever internal mechanism separates the tone from other acoustic input.
As another example in the same vein, if the noise spectrum is flat, the actual level at which the tone is correctly reported x% of the time allows a determination of the efficiency of the internal signal pathway as compared to the mathematical ideal observer. As a third example, the inverse of the first, if a tone at 500+-n Hz is present in all four intervals, how does the value of n affect the intensity of the tone at 500Hz required to allow a correct report x% of the time? This latter would seem to be studying the same property as the first example -- the acceptance bandwidth of the signal detection mechanism, but in fact it gives a different estimate for that bandwidth; this difference tells something about how the detection mechanism works.
Yes, a lot can be determined about perceptual pathways from conventional psychophysical experiments done with good technique.
Martin