If I am driving down the street with the sun in my field of view, clearly everything doesn't turn black. You are going to answer that by saying that you are going to discount objects that are luminous, that are light sources...

Gilchrist

I am not going to say that

Wolfe

Everything DID turn black?

No I think that you are going to discount things that are in a totally different plane

Wolfe

Oh, I don't know. You drive down that Mass pike from my house at this time of the year in the evening and it's smack there in your fovea..

Gilchrist

Yeah! I know you can have an accident (laughter) but look.. to describe the phenomenology of that correctly, I don't think it makes things look black, I think it just makes it hard to see things.

Yes, I'll grant you that!

That's not black... What's needed here and I'm proposing to do some experiments, we need to have some things, some objects like that which have bright things in the distance and then have people judging their shade of grey. And you might get someone to tell you "well, I can't make the judgement, I just can't tell", but I don't think you're going to find people calling those things black. I know the glare makes it very very difficult to see...

Well it doesn't have to be the sun. I was just wondering, if you have luminous sources in the field you've got this notion that what you're doing in this sort of example is an integration from a global neighbourhood and a local neighbourhood....

OK I got the question. That's a very good point. For example, in this room, the highest luminances in this room are the light sources up here. And if the highest luminance rule was the whole story, then the ceiling should look black or dark grey next to those light sources. OK so we have to account for the question of self luminosity, things appearing luminous, OK. And I think the answer there basically has to do with the role of relative area. I didn't talk about this but if you also do those experiments in a dome and again you have two regions a lighter and a darker region in the dome and you vary the relative area of the two regions, we've been doing these experiments where you have these sectors. You start out like this where this is light, this is dark (draws on board) and you make it bigger and bigger. You get a very systematic result. So if the dark thing starts out small then its lightness will be determined as a ratio to this, the lighter region which is seen as light... as that dark thing gets larger and larger, it gets lighter and lighter and when it finally takes over the whole visual field it appears totally white. So it's the area function that explains these kinds of things. These kind of light sources are relatively small in the visual field, so they don't tend to anchor things, so things don't only have to be the highest luminance to anchor things, they have to be relatively large, too.

That's not very big!

But these are not really changing the global anchor. They're very small, you're absolutely right, that's the reason why these things really DON'T affect anything out here, right? because this is a very small part of the visual field ...(gap in tape)

....the scattering goes down as the inverse square of the angular distance; so as you increase the size, if the black area's very small, the lightness would be perceived as higher, because the lightness on the retina would be higher because of scatter, as you make it bigger it can't get so far into it. So that would have an effect in the same direction, but how big do think these effects would be?

Well, let me remind you again though about the white border condition. When we put a white border around this then the errors go away. And with a white border around here you would expect even more of that scattering, if I understand your point

Yes you would.... I was talking about the dome experiment

The dome... so you're saying, aha, well, I'd have to think about that. I don't know, I don't think you can account for much of this with scattering of light. I don't know. I am just giving you my intuition.

Does it work backwards? When you have a small dark wedge it looks darker. You should get more light scatter onto a dark region you open up the thing and it gets lighter.

Thank you Jerry, you saved me! The rule is the larger, the lighter .. so that takes care of that... (laughter)

I'd like to postpone beginning of the session after lunch so that we can have a few minutes to see if we can get anywhere here.

I've noticed that there is a distinct lack of data in these last couple of talks. Very phenomenological, great experiments, obviously something's going on there. But there's no quantitative numbers and I'm wondering whether there's anything methodological problems that could act as a resolution to some of the controversies arising? What do you think about that? Are you using a nulling technique generally for measuring... or are you just saying "it looks a bit bigger" and that's the end of it?

Well, speaking for myself... (laughter) Everything I was showing pretty much is backed up by a lot of data. Now its true that I didn't necessarily present a lot of graphs here and stuff. I have publications if anyone wants...

It's just that you're emphasising it in a different way and I'm wondering whether the methodological problem...

I just thought that you wouldn't want be burdened down by a lot of data. Maybe I'm wrong...

I think that it's quite correct that methodology is very important. And you'll recall that I was instructing people to look at things globally not to focus on individual points. That is how you get the effects of the sort that I'm showing. And it does generally seem to the case that if people scrutinise individual points and compare them to each other that the effects tend to go in a different way. In fact they go more in the way predicted by a local low-level mechanism. That might be because your actually picking up on different mechanisms or that the window of attention has got so narrow in that case that there's nothing but local effects...

But I think that is a very important issue that someone should explore more quantitatively. There's no question that it's there and that it's big.

The effect of instruction set?

Yes.

No, I agree with Ted actually. A very good example is the checkerboard illusion, in which  you have one grey square surrounded by white checks and another surrounded by black checks. Which one looks brighter depends on your mode of viewing: if you focus on the grey squares, the grey square surrounded by black checks looks brighter;  if on the other hand, you let your eyes wander over the checkeboard, the grey square on the white checks looks brighter.

So its hard to get quantitative data that repeatable because very small differences like the attention set of the subject can switch things around totally

With some of these, that's true.

You can get stable data with stable instructions, if you have subjects that are reasonably experienced they'll give you stable data. But if you tell them "scrutinise" vs "look globally" you'll get different results.

Don't exaggerate the tension, there are effects but I don't think you're going to show really big effects, on lightness, Jerry, on lightness (laughter)

For those of us in the attention trade, you cannot exaggerate the effect of attention (laughter) I'm not sure its been done in lightness per se but it is the case that if you give subjects instructions to look for a red x or something like that among things that are red and green, they'll say, "well, you damn fool, you made the red things all brighter". If you ask a different set of subjects to look for a green x they'll say "you damn fool, you made the green things all brighter". So that at least there's anecdotal data (and there's one published paper by whom, I can't remember) that there are actually quite substantial effects of attention on brightness ... wasn't that what doesn't exist?

It exists but we don't care about it.

I'm intrigued by this notion of grouping and what I seem to be hearing is a lot of stuff about variable grouping but also some notion that grouping sometimes is a substitute or surrogate for something called "illumination" which in a previous model had a kind of a formal analysis. The thing I'm wondering about is what is the characteristic of something that becomes not just a blueprint but a surrogate for 'illumination' that is apparent illumination. One reason for anything like this is of course to do with attention. Some things are cusp or ambiguous, that is you can treat them as having the characteristics of a grouping that involves physical things that allow you to treat them as illumination. And other grouping will not have those characteristics. So what are those characteristics of the stimulus that allow you to treat it as illumination?. What sort of characteristics change a grouping from "Here's another grouping" to "Here's a grouping that has to do with illumination"?

If I can speak to that (this will require a long discussion with more slides for me to really give this answer). There's a dispute, I think it's still going on, between Alan and Acussim Zide . and everyone else ... Acuspi and Brocha Swehar came up with a demonstration where the classical grouping rules predict an illusion in one direction but sort of more heuristic T-junction type rules predict the other direction. And the classical grouping rules loose. And their argument, I think they still hold this view, is that Alan is wrong and there's no grouping involved in the computation of lightness. And my interpretation is that there is grouping involved but the grouping for illumination has different rules from classical Gestalt belongingness and it actually makes sense that it should have different rules because the kinds of cues that tell you that things are in the same plane or that they belong to the same object are different from the cues that tell you they are in the same single illumination. I can't show you without showing a bunch of slides, but I think that there are multiple grouping systems with their own rules...

What Acussim did was he had a stimulus where his pitted grouping by T-junctions against co-planary. I would say it's one kind of grouping pitting against another kind of groupings. He was referring to the T-junctions as one kind of grouping or coplanarity.

With his stimulus, if you ask a group of people what is being grouped here they say "it's all in the same plane".

I think the experiment that he really did was he set one group factor against another. I would interpret T junctions as a grouping principle ... (xxx and others have got very similar ideas). When you have a T-junction I think it tends to group the parts that are on the two sides of the stem and segregate both of those from what's above the top of the T. That's a rule that can be generally applied. So I see T-junctions as a grouping factor and I see co-planary as a grouping factor. So cossim pitted one against the other. And they say in that paper that they get both effects. It actually says, literally that what you see is a mixture of the two effects. And that's exactly what I'm saying, that you have these different groups and you have a computation in each group and then you get a weighted average of them. I would take CCCCs data to be an excellent example of the anchoring proposal and I wish I had done that experiment myself..

Question: the easiest way to frame it is in terms of Jim Bergman's grid, the Bergman Grid and it's clear as Ted said, that you can't explain that just with filtering but my next reaction was, well if you have filtering followed by some non-linear collinear facilitation which would operate if the lines were continuous but not if they were offset plus dynamics that you obviously need for things that come and go, that you would build up low level things into a high level thing involving neurally plausible grouping that was dynamic and you'd explain it and so (a) has anyone done that and (b) does that sound like the kind of thing your comfortable with as an approach in principle, if not in detail?

I agree that if I was going to try to model it I'd do something like what you describe and surprisingly to me it seems as though there is almost no work on this Jim presented this at a couple of conferences in Los Angeles and ... published in Vision Research. These are phenomenal effects and almost no work has been done either psychophysically or in modelling and that surprises me.

But I would have no desire to model it ... except for the fact it's a basic psychophysical effect.... But if I were trying to design a machine to do lightness, it would not occur to me to piece it together with these complicated non-linearities. It maybe that I'm just not thinking what's good about it but it seems to me you can model it but you have to just make up a weird model just to get the effect. Do you feel like there's good reason to do this?

Well I think, carrying it a step further, it seems to me that that kind of model could then be perhaps directly compared to what Nikos Logothetis or someone else might extract from single units in a behaving monkey. Whereas the kinds of models with linguistic rules if you like, are much more difficult to associate with the underlying hardware. So that would be the reason, I would argue.

Yes I also would also be very interested if physiologists would go look for correlates of Grid Psychophysics. That could be very interesting. And it's also true that it is not at all clear where you would go, in physiology, to look for the midlevel stuff. Its not at all clear where to look for any mid-level cell in physiology and that's a problem. Even though we talk about these things with words, we talk about [inaudible] and judgements and so on, I think that the actual computations are fuzzier and more robust and really not so cleanly defined. In Russell Frazell's [?] work for example, I think it IS possible to formalize and compute a lot of these things that are based on judgements, [inaudible] and so on. One doesn't need to go to all the way to a linguistic-level description.

The thing that seems to be common to cover the range in the three talks was the approach to low-level mechanisms all the way from full embrace to extremist rejection. And I wonder if there isn't there some simple figure that can just solve that question. To take the extremist point, Alan's point, that low level is not contributing. So I have two questions and here's the first one. Didn't one of Ted's slides have simultaneous contrast with fuzzy patches where the illusion increased?

Cavanagh

And it doesn't seem to invoke any change in grouping or [inaduible] or change in illusion..

If your asking me.. Well first of all I think there is a grouping factor there in my model and in these data. I would call a fuzzy edge a segregating factor. So what a fuzzy edge does in my opinion is that it segregates two regions and says these are parts of different levels of illumination .You shouldn't use something on one side of a fuzzy edge to anchor something on the other side of a fuzzy edge. Now that doesn't really apply to Fred's case because there's just a single target and a fuzzy edge but it does apply for example if you take the difference between a black background and a white target. If you make that fuzzy it's well known that you get a stronger illusion and so it handles that case because it segregates those regions even more.

And so you've got a fuzzy target?

Yeah. I don't know what to say about the fuzzy targets. I don't know how to explain that case. You can't say there's no grouping factors in it. Yeah OK... (laughter)

You've ruined my second question which was given that you have explained [inaudible] can't you explain everything and so its not falsifiable. But since you've admitted it's falsifiable....

So what's the conclusion? What do we want to take away from here? Not low level, not high level, might be midlevel. What do we think, all of the above?

We should take a vote of the three speakers (laughter)

For what? Its not absolutely clear what the cases are for each person. Is that what you want? Low, Mid, High? Is that what you want to be the three levels?

Kingdom

One general point that I'd like to make concerns the idea of trying to find a unified model for a whole set of phenomenon.  In principle it is obviously a good idea, but what worries me is that it carries with it a tendency to reject a model or mechanism as soon as a phenomenon comes along which the model/mechanism cannot explain.  Imagine we applied that principle to all aspects of vision, for example colour vision. It would mean that since we can't explain all of colour vision with trichromatic theory we should reject the theory of trichromacy. Or because we can't explain all of colour vision on the basis of colour opponency, we should reject colour opponency. Or because we can't explain all of colour vision in terms of colour constancy, we should reject colour constancy. My concern here is that we might end up rejecting all multi-component-system approaches to vision.

Neither Ted nor I are rejecting the idea that there are low level mechanisms.

What!! (laughter) May I quote you on that?

I made the point that I think lateral inhibition is a mechanism that allows you to pick up the edges. I'm not saying that lateral inhibition is not a fact! It's there. And I think its part of the mechanism to pick up the edges. I'm simply saying that when you pick up the edge you don't have a lightness level, there's no lightness values.

I think when you say that the higher level lightness computation fails to override the lightness levels at the first level, there have to be lightness values at the first level and I don't know where they come from...

If you are going to disrupt the signal at low level, it's going to carry through the lightness level. So the error can certainly come from [the lower levels] ... They are not transparent to lightness.

Well I don't know what's disrupted. I think what the system does is it picks up the edge ratios at the edge and it uses lateral inhibition to do that. Where's the disruption in that?

Where's the lightness getting computed from by the time you get to your midlevel if you haven't got something like a 'signal' (if you don’t want to call it a lightness level) where's it coming from ? The low level.

If you want to find the highest luminance you have to integrate a bunch of edges because highest is a relative thing

OK, take the case of the grating induction stimulus. You can predict the illusory variations by the output of a simple filter. Now its true that if you want to assign to any part of the stimulus an absolute grey level then you have to have an anchor, but that in no way negates the possibility that the illusory variations in lightness are signalled by a filtering mechanism. In other words in order to say "that bit is white, that bit is mid-grey and that bit is black", as opposed to "that bit is lighter than that bit" then you have to have anchoring. But this does not conflict with the idea of a low level mechanism that signals the variations in lightness across the stimulus.

Incidentally what is the Gestaltist explanation for grating induction? Iíve never understood what it possibly could be.

My suggestion would be that it's based on proximity. We always find these very dumb groupings, they're very weak I think, but they're based on just local proximity. I think its true that things are influenced by their local surrounds regardless of whether that's in the same plane or whatever. .. sometimes there are very small effects there that are irreducible, you can never get rid of them. But that's consistent with the whole principle. Paul Whittle said attributing lightness to lateral inhibition is like attributing the jerky driving of a student driver to the explosions inside the pistons. Yes there are explosions inside the pistons (cylinders ... laughter ... you know how a car works) but the explosions are there and jerky driving is there but that's a coincidence and also that the pistons do have to be moving in order for the driver to drive. But it's a distortion to say, it's ridiculous to say that jerky driving is due to those explosions.

Wilson

I can tell we're getting close to a resolution now...

Here is a question to all three of you. These low level mechanisms, they are characterised by their local processing too: two things. Low or high there is also the fact that they are very local. So the problem is basically of finding where is the transition: when can you rely on totally local information and when you have to have an integration rule, might be much more complicated than you think. May be much more difficult to derive just from seeing the local properties of the low level mechanism.

If local means small distances then there's a problem because certainly there are induction phenomena over long distances and most people who do low-level models have sort of a multi-scale component... If local means with respect to the scale of the object then that would still be ok. I'm just listening to Alan whom I generally agree with, say that we do have a problem in that if we say "well, you are doing normalization within a neighbourhood and when the neighbourhood becomes small it's just normalization within a small neighbourhood" . That's really not distinguishable from looking at a centre/surround kind of organisation. A mid level explanation does has this unfair advantage in that if it includes centre/surround as a special case, it becomes possible for us to say that we don't need low level mechanisms.

OK that seems to be a lull there… (applause)