[Bulk] Re: [Bulk] Re: NYTimes.com: Cells That Read Minds

CSGnet Archive CSG2006
1

[Martin Taylor 2006.01.18.09.51]

[From Rick Marken (2006.01.17.2200)]

I tried to post this earlier from work but it hasn't shown up yet. I may eventually but just for the sake of keeping up the conversation I'm sending it off again. I apologize for the likely repetition.

It happened!
I'm glad you responded to this message, because I was going to suggest to you a (hopefully easy) extra condition for your "divination of intent" demo that set off Erling.

----
[From Rick Marken (2006.01.17.1350)]

Martin Taylor 2006.01.17.13.41]

Rick Marken (2006.01.17.0800)--

... As far as zero value of the perception, there is presumably zero
value of the cursor position perception when you are not looking at
the cursor (or covering it up with something)

Having no value is not the same as having a value that happens for
the moment to be zero.

I don't see that. I was thinking of the perceptual signal, which is presumably represented
as rate of neural firing. So when the rate of neural firing is virtually zero impulses
per second (there is always _some_ firing) there is zero perception.

Don't think of intensity perception, think of position, and think of your demo. Where on the screen is zero position? If I want to control for the cursor being at position zero, and it is in the middle of the screen, what do I do? If I want to control for the cursor being at the middle of the screen and I can't see it, what do I do?

Those two conditions are different, aren't they? In the first case, I may have defined "zero position" as the middle of the screen and there is no error, or I may have defined "zero position" as the left edge, and I act to move the cursor leftward, reducing the error until it gets to the edge. On the other hand, if I can't see the cursor, I can't act reliably to move to a position of zero error. I have no perception of it, not a perception that is equal to zero.

Which brings me to my proposal for an amnedment to your demo.

Instead of having the mouse control the screen display directly, have it control through a hidden (2-D) variable I'll call {P,Q}, and let the screen {x,y} be {P+a,Q+b}, where a and b are variables. Give the experimenter/subject two sliders for each of a and b, one of which determines the standard deviation of the magnitude of a and b, the other of which controls the bandwidth of the variation (which must be slower than variation in P and Q. The subject's task now is to control {P,Q}, not the screen position. Let a and b be zero for some time at the start of a run before adding in the variation.

The idea of this is to simulate Erling's covering the screen with an envelope, but also having intermediate conditions rather like having a translucent envelope that gives you a fuzzy idea of where the one you are tracking (and the cursor) are at any moment. The subject won't lose control immediately, and the demo should be able to track how long it takes before it happens.

An alternative form, which is probably rather harder to program, is to fuzz the display of the cursor and/or the potential targets.

Yet another alternative, which might be even better for the purpose of demonstration, is to define a precise area on the screen that acts like a small envelope. In this area, the screen contrast of the cursor and targets is diminished (the area could be grey, ranging from white to black according to a slider setting). At zero contrast, it's just like the envelope, and at low but non-zero contrast, it's like a translucent envelope. As the dot being tracked (and the cursor) passes in and then out of the obscured area, one could determine the time course of both the loss and the regaining of control.

I think some such variant of your demo might both illustrate the difference between a zero valued perception and a non-existent perception, and at the same time allow for some interesting observations about losing and regaining control when the data are obscured and the tracker must rely on control through imagination during an interim period (see the thread of a few weeks ago).

Martin

2

[Martin Taylor 2006.01.24.14.30]

[From Rick Marken (2006.01.24.0915)]

I still don't understand what you mean by "seeing something to be precisely zero". We were talking about a specific example: the tracking task that is the first of my demos ("The Nature of Control"). What is being perceived is the position of a cursor and target. What is the state of that display that is seen as "precisely zero"?

That comes from Erling's introduction to this thread. But see below.

In your mind, the model trumps the observations, and I don't know how to go about explaining it differently than I have heretofore.

I don't think so. I'm trying to think of how the model would differentially represent the two different situations that I experience: 1) the situation where I am controlling the distance between cursor and target (a perceptual variable that is presumably represented by the rate of firing in an afferent neuron) and 2) the situation where I can't see the screen at all (which I would represent in the model as zero firing in the afferent neuron that represents the perceptual variable that had been controlled).

OK, but I think there's an internal contradiction in your paragraph. Let's try to follow it through.

For the sake of argument (and as the mathematicians say) without limiting its generality, let us assume that the controlled perception is a location left-right on a screen. And let us further say that a perception of increasing "rightwardness" occurs when the firing rate of the left-right perceptual input function ("afferent neuron" in your paragraph) increases. Correspondingly, the perception is more leftward as the firing rate decreases.

In the absence of some special functional organization not specified in any model I know, this arrangement would mean that zero firing rate corresponds to a perception of the most leftward imaginable position. However, you take this firing condition to mean that you don't perceive the position to be anywhere at all. To me, that's a contradiction.

I grant that your didn't specify the model -- I did. But I think I only used an assumption I believe you were using -- that the neuron in question is one that signals change in left-right position by changing its firing rate, and that we are talking about the standard HPCT arrangement of the perceptual input functions.

I can't see any way out of the problem that you are using the same firing rate to result in two quite different perceptions at the same time: that the position is maximally leftward, and that the position is unknown.

I do see a way out if there are two separate dimensions of the perception: the most probable value of the thing perceived simultaneously with a separate perception of the uncertainty of that perception.

Along this line, a few years ago I was inroduced to a company who was making multi-level perceiving systems that were like multi-layer perceptrons except that the value of a signal and its uncertainty were represented together as a complex number A*e^(i*B), where B (a phase angle) represented the value of the thing perceived and A represented its precision. Using this representation, the network learned quite fast to make very good decisions in complex spaces. I've often wondered whether this kind approach might be physiologically valid, and if it might have some use in developing PCT theory.

But let that be...it was just an aside. The important point is that perception of value is a perception of a different kind than perception of the precision of a value considered as a representation. They are as different as are the perceptions of a sensation and an event. Perception of precision is something we do (the envelope covering the screen is just an extreme example), but it has no obvious place in the canonical HPCT configuration.

Maybe if you were to make the modifications to your demo that I suggested, you could see what I'm trying to get across.

But the problem is that I don't understand the modifications you are suggesting.

I'll try again, without reference to what I wrote before, so as to make it more likely that one wording will illuminate the other.

You have a bunch of objects that move around the screen with random relative motion. Call them cursors. If you move the mouse, that affects all the cursors equally. The objective of my suggestion is to reduce the accuracy with which the subject can perceive the location of one or all cursors, at a time, or over a region of the screen, known to the computer that will analyse the results. I made a variety of different suggestions about how to do this, some perhaps more practical than others. On thinking about it subsequently, I thought that fading the contrast might be the best, because the loss of accuracy would be directly perceptible by the subject.

So, what I now suggest is that you allow the subject to control normally for the first part of the run, but at some moment you reduce the contrast between the cursors and the background, either by changing the background or by chaging the cursors. Changing the background might be best, because it allows you to generalize to a range of different experiments based on the spatial statistical character of the background masking (changing brightness overall, or putting random dots or lines or structured patterns into the background, for example).

What you should be able to do is generate a graph of the time course of loss of control as a function of the degree of contrast reduction or masking, using the same statistics you use to "mind-read" which cursor is being controlled.

I know this isn't exactly what I proposed earlier, but I thought a bit about it since then, and this is what I think now. I hope it makes more sense.

Martin

3

[Martin Taylor 2006.01.24.23.26]

The crux of the misunderstanding seems quite simply put. Whether it is simply resolved is another matter.

[From Rick Marken (2006.01.24.1350)]

Martin Taylor (2006.01.24.14.30]

Rick Marken (2006.01.24.0915) --

I still don't understand what you mean by "seeing something to be
precisely zero".

>In the absence of some special functional organization not specified

in any model I know, this arrangement would mean that zero firing
rate corresponds to a perception of the most leftwards imaginable

>position. However, you take this firing condition to mean that you

don't perceive the position to be anywhere at all. To me, that's a
contradiction.

I don't believe that cursor position is the variable controlled.

Isn't the object of your demo that the subject chooses one of the cursors and controls its position according to any pattern (s)he sees fit?

And I don't think the left right position of the cursor is represented in terms of firing rate, with 0 being farthest one way and max firing being the farthest the other way.

I don't really care whether firing rate is the physical manifestation of the output of a perceptual function. You said it was, so I went along. The argument doesn't depend on that. The point is: for a perception of "left=right-ness" greater output means more one way and less output means more the other way. At least that's the case in any reasonable interpretation of HPCT. You are the one that talked about firing rate.

I think zero firing rate has to be reserved for zero of whatever perception is computed by the perceptual function; that is, zero firing has to mean no perception.

Here's the crux, very simply. You say, to paraphrase (I hope correctly): Firing rate value non-zero indicates some value of the perception; firing rate value zero indicates that the perception doesn't have a value.

I said, in the bit you quoted up above: "In the absence of some special functional organization not specified in any model I know, this arrangement would mean that zero firing rate corresponds to a perception of the most leftwards imaginable position."

Now you are asserting that there IS some such special fucntional organization that changes the meaning of the perceptual signal when its value is exactly zero. If its value is 10, 1, .1, .01, .001, the thing perceived (in this case a position) is precisely that value, but let that .001 be reduced by .001 and suddenly the thing perceived is not that value. There's a new mechanism.

If the value is represented by a firing rate, as you like, when the firing rate gets very low, like one impulse per second, or one per minute, ... just when does this new mechanism decide that the rate is actually zero so that there is, in effect, no firing rate (note: not "no firing" but "no firing rate")?

I would guess that what is controlled in the tracking task is the distance of the cursor from the target and that this distance -- left or right of the cursor -- is represented by rate of firing, with zero distance being some intermediate firing rate, left distance being rates higher and right distance being rates lower than that intermediate value.

That's adifferent perceptual situation, but I'm happy to use it in the argument. As the rightward distance increases, the firing rate gets lower and lower. At one impulse per minute it's an awful long way right. At one per hour, it's even further. At what near-zero rate does the rate stop being an indicator of position and become an indicator of ignorance about position?

I can't see any way out of the problem that you are using the same
firing rate to result in two quite different perceptions at the same
time: that the position is maximally leftward, and that the position
is unknown.

Right. We can't let that happen. I makes most sense to me to have zero firing rate represent none of the perception and other firing rates represent other values of the perception. So if the perception is of zero deviation from the cursor the IS a perception -- there not NO perception -- so that zero deviation has to be represented by a positive rate of neural firing.

I do see a way out if there are two separate dimensions of the
perception: the most probable value of the thing perceived
simultaneously with a separate perception of the uncertainty of that
perception.

I like my way out better;-)

But your exit door turns out to be a trompe l'oeil painting on a concrete wall.

I'll go along with the conventional notion in HPCT that if a signal line has some significance in a control loop at one value of the signal on it, it has the same significance for other values of the signal -- if a line is the value of the perception of "red" when the firing rate is 200 Hz, it does not become the value of the perception of "democracy" when the firing rate is 20 Hz. If a signal line represents the value of the perception of position at 100 Hz, it does not become the value of the perception of the precision of the perception of position at 0 Hz.

Martin