Studies in Visual Perception, I
Some observations on the psychophysics of Glass patterns and related visual phenomena

by Jack Schwartz

Courant Institute of Mathematical Sciences. New York University

Home . . . . E-Mail

The aim of the extensive set of visual examples to be developed in this paper and its sequels is to illuminate and refine a view of the way in which the initial stages of human (doubtless also primate, and,with some modifications, many other mammalian) visual system is organized, In broad terms, this is conceived of as a bottom-up process, carried by multiple families of cells all laid out in essentially retinotopic correspondence.

The architecture which the images and animations given below seem to suggest is as follows (we speak below of the 'polarity' of an image element, regarding it as 'positive' if it is lighter than the general background, and as 'negative' if darker than the background.)

'Early' processes [pre-edge perception, 'dots' only]

These are high-speed, timing sensitive, and polarity sensitive.
  1. Slant detection.
  2. Texture detection.
  3. Color detection.
  4. Random-dot motion detection.
  5. Glass-pattern detection.
  6. Random-dot stereo depth detection, formation o a 'cyclopean' image.

****** Formation of edges, including illusory edges ******

'Late' processes [edge dependent]

These are slower, less timing sensitive, and relatively insensitive to polarity.
  1. Edge-determined region motion.
  2. Phi motion.
  3. Edge-determined stereo.
  4. Region color and luminance ascription.
  5. Color constancy management.
  6. Binocular fusion.

****** Formation of regions ******

****** (Onward to higher-level, gestalt-based visual processes) ******

This architectural assumption suggests that something like the following number of distinct cell populations may be involved.

For the early processes [pre-edge]

  1. Slant detection. [20 populations, by preferred angle]
  2. Texture detection. [10 populations, various feature, possibly randomized]
  3. Color detection. [3 populations]
  4. Random-dot motion detection. [16 populations, by preferred angle]
  5. Glass-pattern detection. [1 population]
  6. Random-dot stereo depth detection. [4 populations]
  1. Edge-determined region motion detection. [16 populations, by preferred angle]
  2. Phi motion detection. [10 populations, by preferred angle]
  3. Edge-determined stereo detection. [4 populations]
  4. Region color and luminance ascription. [4 populations]
  5. Color constancy management. [?]
  6. Binocular fusion. [1 population]
Since we assume that all but the color detection mechanisms of the early set are doubled populations, one handling dark dots and the other handling light dots, we conclude that as many as 117 + 35 = 152 architecturally distinct populations of neurons may be involved in these stages of visual processing alone.

Examples of light-dot/dark-dot segregation. Many psychophysical observations suggest that perceived 'light dots' and 'dark dots' are segregated in the early stages of the visual system: perhaps the signals representing them are carried by separate families of visual neurons. One class of images in which this phenomenon is salient are the well-known 'Glass patterns' introduced in 1969 by Leon Glass. These are formed when two copies of a random dot pattern are superimposed on each other, as shown below. A clear perception of 'swirl' emerges. However if the background is grey and the dots in one of the groups are darker than the background while those in the other group are lighter than the background, then the perception disappears, as seen in the second figure below. The perception reappears if the background is made dark enough for both groups of dots to be lighter than the background, i.e. to be seen as 'light dots'. All together these figures suggest that the neural mechanism by which the perception of 'swirl' is formed does not combine light dots with nearby dark dots to create the basic 'pairs of nearby dots' out of which the 'swirl' gestalt emerges.

Glass patterns are particularly interesting since they are formed in the complete absence of any edge-related perceptions. Most of the other global percepts formed by the visual system are arguably edge-derived. Thus Glass patterns may serve as a means for studying the emergence of higher-level percepts in a particularly simple case.

Figure 1. A Glass pattern

If the background is grey and the dots in one of the groups are darker than the background while those in the other group are lighter than the background, then the 'swirl' percept disappears. See below.

Figure 2. The same Glass pattern, seen against an intermediate grey background

'Swirl' reappears if the background is made dark enough for both groups of dots to be lighter than the background. See below. It is also worth noting that the swirl percept emerges even if one of the groups of dots is represented by hollow circles, as Figure 3b indicates. Figures 3c and 3d indicate that the swirl percept can persist, though in much weakened form, even if the dots in one of the groups of dots are replaced by short line segments, but that it is disrupted completely if these line segments generate a direction field that contraindicates the swirl pattern too strongly. The 'swirl' percept can resist, but is weakened by, the presence of some other image features generating different direction fields. Figure 3f below indicates that it is important for the formation of the swirl percept that the two items in every pair being matched lie on one side of the other, rather than surrounding it. Figures 3g and 3h demonstrate that the visual mechanism which detects 'swirl' is capable of detecting many other global patterns in the small displacements between two groups of dots. A 'spiral' instead of a 'swirl' percept can be generated if one of the groups is changed in size instead of merely being rotated. An 'exploding out' percept can be generated in place of 'swirl' if the rotation is omitted and only an expansion used. Figures 3i and 3j shows that, unsurprisingly, the 'swirl' percept is retained even if the groups of dots involved are binocularly tilted (but not separated) in 3-dimensional space, or even paced on a bent surface. Figure 3k shows an anaglyphic Glass figure obscured by an overlapping random dot pattern separable from it in 3-space. In monocular view, or viewed anaglyphically with one eye shut, the swirl percept is at best weak. But swirl emerges clearly in the anaglyphic binocular view, again indicating that swirl is perceived at a visual stage subsequent to binocular fusion. The Figure is to be viewed twice, reversing the red-green glasses. Note that the swirl percept is more readily seen when it lies in front of, rather than behind the layer of obscuring dots, indicating the visual system's tendency to emphasize near-plane percepts.

Figure 3l is an animated variant of Figure 3k contrived to illustrate the rule that coherencies detectable in Glass patterns are detectable as coherent motions, and vice-versa. Viewed normally or anaglyphically with either eye closed it generates only a weak impression of coherent motion. But an anaglyphic view with both eyes (start with the red filter over the right eye) separates it into a two groups of animated dots, in one of which a coherent rotation corresponding to the Glass swirl of Figure 3k is visible. With the glasses reversed the other group is easily seen to be moving in and out in 3 dimensions. Note that in each anaglyphic view one of the coherent motions is salient and the other considerably harder to see. These observations strongly suggest that coherent motion is detected after binocular depth, and even after image elements deemed to belong to the 3-D foreground have partially repressed those lying near them but further back.

Figure 3. The same Glass pattern, seen against a dark background

Figure 3b. A Glass pattern in which one of the dot groups is hollow.

Figure 3c. A Glass pattern in which one of the dot groups is replaced by line segments.

Figure 3d. A Glass pattern in which replacement by line segments disrupts the 'swirl' perception.

Figure 3e. The 'swirl' percept resists the presence of other image features generating different direction fields.

Figure 3f. Formation of a 'swirl' requires that the each of the two items in each pair being matched should lie on one side of the other, rather than surrounding it. The two figures seen here differ only in the presence of circles surrounding the dots in one of the groups.

Figure 3g. A 'spiral' instead of a 'swirl' precept can be generated.

Figure 3h. An 'exploding out' percept can be generated in place of 'swirl'.

Figure 3i. A Glass figure anaglyphically tilted in 3-space.

Figure 3j. A Glass figure placed anaglyphically on a parabolic surface.

Figure 3k. An anaglyphic Glass figure obscured by an overlapping random dot pattern seperable in 3-space. (View twice, reversing the red-green glasses.)

Figure 3l. An animated version of Figure 3k. (See coments above. View twice, reversing the red-green glasses.)

If either Figure 1 or figure 3 is animated by moving one of the groups of dots horizontally relative to the other (which for geometric reasons can move the center of swirl between the bottom and the top of an image, it will be seen that the 'swirl' percept is badly disrupted unless the step time between frames of the animation is brought down to approximately 5 frames per second. This indicates that swirl detection is a considerably slower, and hence presumably later, process in neural image analysis than motion detection, which can easily deal with a frame rate of 30 frames/sec. Both depth and swirl perceptions seem to require about 1/10 of a second to form.

The preferential association of 'light dots' and 'dark dots' with each other is further demonstrated by the modified Glass patterns seen in the following three figures, which contain three groups of dots rotated relative to each other, one very light, the second intermediate in brightness, the third dark. If (see below) the background is light, then the intermediate dots associate with the dots of the darkest group, and one sees a 'swirl' centered at the bottom of the figure. If (see the second figure below) the background is darker than the intermediate dots, then the intermediate dots associate with the dots of the lightest group, and one sees a swirl centered at the mid-right the figure. Note that nothing has changed between these two figures except the brightness of the background.

Figure 4. A Glass pattern with three groups of dots - light background

If the background is darker than the intermediate dots, then the intermediate dots associate with the dots of the lightest group, and one sees a swirl centered at the mid-right the figure.

Figure 5. Same Glass pattern as Figure 4 - dark background

What happens if we if construct a Glass pattern figure against a grey background using four relatively rotated groups of dots, one pair darker than and one pair lighter than the background, where each of the pairs implies a different center of swirl? Is it then possible to see two distinct swirls simultaneously (which might happen if the posited swirl-detecting mechanism is duplicated in the light-dot and the dark-dot system)? The next four figures explore this question. In the first of these figures it seems possible to notice both the 'dark dots' swirl (here centered near the upper-left corner of the figure) and the 'light dots' swirl (centered at the mid-right of the figure) almost simultaneously, but somehow not to focus on them at the same time. The perception is bistable, and one can shift ones visual attention between stably seeing the 'dark dots' and the 'light dots' swirls. A surmise concerning the reason for this phenomenon is offered below.

Figure 6. A Glass pattern exhibiting two centers of swirl.

In the figure below the background is very close in intensity to the darker dots in the 'light dots' group, so the 'dark dots' swirl stands out even more strongly. The 'light dots' swirl is still detectable, but not very salient.

Figure 7. A light background emphasizes the 'black swirl' percept.

In our next figure below the background is darkened, making the 'light dots' swirl stands out more strongly, but still retaining a salient perception of 'dark dots' swirl.

Figure 8. A dark background emphasizes the 'white swirl' percept.

In the final figure of this group the background is darkened still more, making the 'light dots' swirl stand out still more strongly, and leaving 'dark dots' swirl perceptible but distinctly less salient.

Figure 9. A still darker background emphasizes the 'white swirl' percept even more.

The dark-light segregation that can be inferred from the preceding observations can also be demonstrated in other visual contexts, including motion perception and binocular vision.

Separation by binocular depth. It is interesting to see what happens if we separate our two groups of dots in some other dimension than light-dark, e.g. give them different binocular depths. Our next figure is an anaglyph which does this (set up for viewing through the usual red-green glasses.) A clear perception of swirl (centered toward the mid-bottom of the figure) emerges if the figure is viewed monocularly. When the image is viewed through red-green glasses which separate the two groups of dots binocularly, the perception of swirl is greatly weakened. However if one shuts one's right eye (assumed to be behind the blue filter), the red dots, seen in isolation, form a clear swirl gestalt centered toward the middle-top of the figure. Similarly, if one shuts one's left eye, the green dots, seen alone, form a swirl gestalt centered toward the middle-bottom of the figure.

Figure 10. Figure 2 with the two groups of dots separated binocularly. Different swirl percepts emerge in the ordinary, anaglyphic, and single-eye views.

It is worth reflecting on the interesting effects which this example exhibits. The fact that the perception of swirl is disrupted when the groups of dots entering into it are separated in depth shows that the swirl-detection mechanism (in this example, the posited light-dots mechanism) tends not to associate nearby dots that differ significantly in binocular depth. This suggests that detection of binocular depth comes before swirl detection in the visual processing flow. The fact that the swirl percepts clearly present if either eye is shut disappear in the binocular view also suggests that swirl detection cannot come before combination of the images from the two eyes, since if it did we would presumably see two bistably competing swirls, as in Figures 6-9.

Animation. If, as our next figure illustrates, we animate the preceding image by presenting its two groups of dots alternately, then various differing perceptions of coherent motion emerge in the ordinary, anaglyphic, and filtered single-eye views. In the both the ordinary and the green-filtered view the coherent rotating motion seen is centered near the lower middle of the figure. In the red-filtered view the rotation seen is centered closer to the middle of the figure. In the binocular view, the perception of rotation is greatly disrupted, but a coherent in-and-out motion is seen instead.

Figure 11. The preceding Figure with the two groups of dots presented alternately. Different percepts of coherent motion emerge in the ordinary, anaglyphic, and filtered single-eye views.

Dark-light separation in the detection of coherent motion. Here is an animated image formed by alternating between the 'dark' and 'light' groups of our first Glass image. In spite of the disturbing alternation in intensity, a definite perception of coherent motion emerges.

Figure 12. Coherent Psi motion between two groups of dark dots rotated relative to each other.

Our next image shows exactly the same groups of animated dots, but against a background dark enough for one of the groups of dots to be darker than the background and one to be lighter. This corresponds to the static Figure 2, in which we noted that the perception of swirl disappears. in this animated version, we see that the perception of coherent motion is much weaker, leaving us with an impression more likely to seem that of a random group of chaotically flashing dots.

Figure 13. Motion between two groups of dots, of opposite brightnesses, rotated relative to each other.

Next we present the same alternating groups of dots against a background dark enough for them both to be seen as light dots, as in the third of our initial static images. As in that case we observe the reappearance of a strong global gestalt, in this case the perception of coherent motion rather than swirl.

Figure 14. Coherent Psi motion between two groups of light dots rotated relative to each other.

By animating Figures 4 and 5, which involve three groups of dots at different levels of intensity, rather than Figures 1, 2, and 3, which involve two intensity levels, we can use the preferential association effect just described to generating two quite different motion percepts from exactly the same alternation of dots, depending on whether this is seen against a light or a dark background. This is demonstrated by our next Figure, which shows otherwise identical light-background and dark background animations side by side, together with the two frames that enter into the light-background animation. Note that one of the two alternating frames must contain both the lightest and the darkest dots, while the other contains all the intermediate dots.

Figure 14b. A Glass pattern alternation generating different motion percepts against light and dark backgrounds.

Further confirmation of the tendency of the binocular system to fuse bright with bright and dark with dark dots. Next we present two identical pairs of lines set up for red-green anaglyphic viewing. However, these are seen against different backgrounds, one darker, one lighter. View these with the blue filter on the right, the red filter on the left. Against the dark background the right eye sees a single bright line (deriving from the green line on the right), while the left eye sees a bright line (deriving from the red line on the left) and a (weaker) dark line (deriving from the green.) Accordingly the right-eye image fuses to the right, so in the anaglyphic binocular view we see a bright line recessed in depth.

Now consider the situation when these same two lines are viewed in the same way against the bright background seen on the right. Here the right eye sees two lines, one bright (deriving from the green line on the right), one dark. The left eye sees one dark line (deriving from the green), plus a much fainter (dark) line deriving from the red. In this case the dark line seen by the fuses to the right, and so in the anaglyphic binocular view we see a raised dark line. Note however that since in these binocular matches at least one unmatched line is always seen by at least one eye, one or more additional lines of no particular binocular depth accompany the raised or lowered line seen.

Figure 15. Identical anaglyphic line pairs seen against backgrounds of different brightness. See discussion.

The following very simple anaglyph involves only two lines, one moving horizontally relative to the other. This is what is seen if the image is viewed binocularly, or monocularly through the blue filter. But, if the same animated figure is viewed anaglyphically, the perception which emerges is that of a single line moving forward. Since it is known that motion detecting cells are present in the retina, it therefore seems that the perceptions of changing binocular depth generally represses sensations of planar motion when the two occur in register. This is fact a general requirement of binocular depth perception, since the changing disparities that hint at changing depth typically occur at edges, which seen monocularly, are moving left-to right.

Figure 16. A two-line animated anaglyph.

To confirm the impression that swirl detection must follow at least the earliest stages of motion detection in the visual flow, we present the following figure, which presents this same relatively rotated groups of dots as Figure 1, but now marks them only as small patches of incoherent motion within a random-dot field. The 'swirl' percept still emerges, though less distinctly. Further experimentation shows that if one of the groups of dots is represented by pure motion patches, while the other is shown statically by either dark of light dots, they do not combine to form a percept of swirl.

Figure 17. A Glass figure marked by patches of chaotic motion.

The following figure provides evidence that there are at least two motion-detection systems of the visual system, one of which detects motion at the 'micro' level while the other detects the macro-scale motion of pre-detected gestalts. It consists of a patch of chaotic motion (formed by rapidly reversing the black-and white dots within a rectangle) which moves more slowly form side to side. The motion of the edges of patch is seen quite distinctly, and, interestingly,communicates itself to the interior of the patch. This is an instance of a familiar, but profoundly mysterious, property of visual perception which can be demonstrated in many settings: purely edge-derived properties of regions, for which these exists no evidence withing the region itself. This fact must somehow relate to the question, touched on again below,of the way in which an entirely distributed system of neurons can represent global gestalts capable of influencing the perception of the distributed element out of which they are formed.

Figure 18. Side-to-side movement of a rectangle of chaotic motion.

Two other phenomena appearing in this image are worth noting. Within the chaotically moving zone formed by dot reversal, one sees a curious texture, different from the surrounding random-dot texture, and distinctly 'coarser' than the latter. This is only seen when the two reversed frames alternate at between 10-40 cycles per second. At too high a reversal rate, one start to see something closer to the flat grey that would be seen if the two reversed images were simply blended.

We also note that, as they move away, the edges of the moving region of chaotic motion sometimes leave behind evanescent strips of 'decaying motion'. The perception of local motion seems to take about 1/4 of a second to decay, and its evanescent after-image is often seen when motion within a delimited field ceases. The following simpler animation, which alternates intermittently between just two of the images which constitute figure 18, explores this idea.

Figure 19. Intermittent chaotic motion within a rectangle.

The following figure, presented anaglyphically, contains a binocularly matched random-dot fields into which fields of chaotic motion displaced from one another are inserted. In the anaglyphic view, a weak but perceptible of a raised rectangle emerges. (The percept is considerably more definite if the same figure is presented, not anaglyphically, but for mirror viewing. This suggests that edge perceptions, even of pure motion edges, are formed prior to some stages of depth detection.

We may note a subtle feature of this image, which is best viewed with the red filter on the right. Close anaglyphic inspection of the characteristic 'reversing motion' texture seen within the rectangle of chaotic motion will show that it often seems to lie, not at the level of the raised rectangle seen, but rater in the same plane as the stationary outer areas. When this is the case, the raised plane is seen as a 'transparent film' standing in front of the more or less flat background, within which a patch is moving chaotically. 'Films' of this sort often appear, perhaps as representations of pure boundary-generated gestalts within empty regions, when such gestalts can be raised in depth over the background region they fill.

Figure 20. Depth perception created by offset patches of chaotic motion.

The following simple binocular-color figure demonstrates the extent to which even a thin boundary can affect the perception of a large region which it surrounds. The figure shows two thin (2 pixel wide) rectangles on uniform red background. Viewed monocularly the interior of the yellow rectangle looks little different from the interior of the rectangle on the left: - perhaps a touch brighter. But viewed anaglyphically the right hand rectangle seems far darker. This can be explained by reference to the two single-eye images which are being combined (close your eyes alternately to see these). In the blue-filter image the background is very dark, the green rectangle stands out brightly, and the rectangle on the left is nearly invisible. In the red-filter image the background is much brighter, the green rectangle is barely visible, and the rectangle on the left dark. Thus on the left there is a strong edge only in the red-filter image, and on the right there is a strong edge only in the blue-filter image. Hence the following general rule applies: when two differing images are combined binocularly, and if one contains a boundary which the other does not, the color derived from the image containing the boundary will be preferred to that derived from the image containing none. On the let this rule selects the bright color of the red-filter background; on the right it selects the dark color of the blue-filter background. Note that this phenomenon argues that the ascription of color to regions is a late visual process, possibly occurring after edge detection.

Figure 21. Effect of edge color on identical regions viewed anaglyphically.

Interactions between Glasss 'swirl' and motion perception. The suspicion that 'swirl' perception may somehow arise as a perceptual parasite of motion perception makes it interesting to look for interactions between these two kinds of perception. Can coherent motion of glass patterns be seen, and how do fixed Glass patterns behave if the dots constituting them move chaotically? Our next few figures explore these questions. Figure 22 confirms the unsurprising fact that swirl perception is not incompatible with small coherent motions. (The figure, like many others in this group, should be viewed anaglyphically.)

Figure 22. A small coherent motion of a glass pattern.

Figure 23, which is constructed by moving the first group of dots of a glass figure an average of 15 pixels at random while holding the swirl pattern itself fixed, shows a more interesting effect. The invariant swirl is still seen, but the chaotic motion of the dots constituting it are mistakenly interpreted as an approximately coherent swirling motion of the entire pattern, which upon continued gaze seems to swirl,now clockwise, then again counterclockwise. It is as if the two percepts 'motion' and 'coherent swirl' ae being mis-associated into 'coherent swirling motion', which , since it has no intrinsic direction of swirl, switches between the two possible directions, perhaps as the perception of one of the two possible directions fatigues. It is also interesting that the active motion of the anaglyphic dots has little effect on the placid 3-dimensional surface on which they lie.

Figure 23. A mis-associated motion signal generated by a random motion creates an illusory coherent motion percept.

Motion signals can mis-associate with higher level gestalts in other contexts, as illustrated by our next example (which can be viewed only as a "Shockwave for Director" movie). It contains a random-dot background moving continuously and coherently upward. Within this, two random-dot subregions carrying motions of a different kind are embedded. The one at the top is a simple rectangle filled with a purely chaotic motion. The second is a stationary region of coherent horizontal motion, forming the letters "RED". At the boundary between this region and the vertically moving background, the eye detects the sudden change of motion direction and forms pure motion edges which are distinct enough for the letters to be readable. But the coherent horizontal motion of the random dots within the letters is mistakenly reattached to the letter gestalts formed from the motion edges, so that the letters themselves seem to be continually moving to the right, even though continued attention to their position shows that they never get anywhere. A less striking but noticeable illusion attaches to the bar of chaotic motion at the top of the random- dot area: the surrounding coherent motion somehow makes it appear to be 'rolling'. Note that this figure is interactive: you can toggle various features of it by clicking on the buttons shown.

Figure 24. A mis-associated motion signal creates an illusory global motion percept.

We can explain the perceived 'rolling' of the chaotically moving bar in the above figure as follows. The sensation of upward motion of the upper edge of the bar emits a weaker upward-motion signal than the surrounding region, and so ay betaken to be moving more slowly. The same is true at the lower edge of the bar. This is just the pattern of speeds which would be seen at the edge of a rolling cylinder, hence this percept is plausible. The direction of roll seen would be correct if the bar were an obscuring cylinder above the background.

The preceding figure raises an interesting question which is explored more fully in the variant of it given next. Since the at the boundary between the letter gestalts seen in Figure 24 and the surrounding random-dot area the two areas are symmetrically related, why is it that we see the letter gestalts moving to the right relative to its surround rather than the surround moving upward as a global object past the letter gestalts embedded in it? One may suspect that this perception is influenced by the immovable outer edges of the figure. Our next figure tests this idea by replacing the letter gestalts by a box surrounded by a rectangle which leave a sufficiently large rectangular section S of the coherently moving surround visible between them. here, one of two illusory motion percepts can form. The easiest seen is that in which the inner box, along with the surrounding outer rectangle R which follows it seem to be continually moving (as an object) to the right. But, upon protracted gaze, the 'object' perception will transfer to the rectangular section S lying between R and the inner box, remain stably there, and then suddenly transfer back (with something of a sickening 'lurch', possibly ascribable to a momentary interaction with visual system's fundamental the inner-ear based gaze stabilization system) to the original gestalt ('object' is inner rectangle plus S.) This sequence of perceptual events seems to be largely independent of conscious will, one simply needs to maintain gaze and wait for it to happen.

Figure 25. A mis-associated motion signal which can be see in either of two metastable ways.

It would be very interesting to know how the gestalts between which perception shifts in this example, and may others like it (e.g. the Necker cube) are held together enough for the shifts that take place always to affect a whole gestalt rather than a fragment of it. This question may prove to be fundamental to the understanding of visual perception. The instance seen in Figure 25 is particularly interesting since both of the gestalts between perception shifts are illusory, and would be disrupted at once by more visual evidence, e.g. a black line trace along the edges bounding the motion-delimited regions.

If, as in our next figure, the modest side-to-side jumps seen in Figure 23 are made larger, the eye no longer sees the motion as sufficiently coherent and the illusory rotation seen in Figure 23 starts to reappear. There is also a curious tendency to fill in the blank areas of the figure with evanescent 'motion trails'.

Figure 26. A side-to-side motion of a Glass pattern, with a larger jump.

The next figure in this collection shows the dots in a glass pattern moving endlessly to the right while subtly adjusting themselves to leave the 'swirl' itself stationary. The motion interferes enough with the swirl perception for the fact that the center of swirl remains stationary not to confuse the eye.

Figure 27. A Glass pattern whose dots move steadily to the right while the 'swirl' itself is stationary.

Discussion. The wildly speculative diagram seen below represents the kind of model suggested by the preceding observations. It assumes that the retina emits several kinds of signal groupings, segregated within corresponding cell populations with a genetically determined preference for connecting to cells of the same kind rather than of different kinds. Each of these signals is (like the cell population which carries it) retinotopically organized, and so can be thought of as an abstract 'image'. The separate signals are kept in register for processing by subsequent stages of the visual system. Among these signals we find one which reports on the presence of image microfeatures which are lighter than their surrounding background ('light dots'), and another which reports on microfeatures which are darker than the background ('dark dots'). As indicated by the fact that each of the three rectangular 'process boxes] in the Figure has a separate 'dark section' and 'light section', these signals are input to separate, duplicated mechanisms which detect certain global gestalts which they suggest, e.g. swirl, coherent motion (in the dynamic case) and binocular depth (which involves bringing together corresponding signals from the two eyes.) These separate mechanisms may detect separate gestalts and emit signals representing their presence. However, since consciousness needs to be unitary (so that coherent behavior (rather than 'Balaam's Ass' paralysis) can be result, a further competitive mechanism allows one of the perceptions in a given modality to repress the others. The inter-gestalt repressions which take place at this stage of visual processing are subject to a measure of modulation by the higher processes of conscious attention, allowing perception to be shifted among then by an act of will.

Figure 30. A possible architectural view of some parts of the visual system.

Note that the assumption made in the immediately preceding discussion, that identical structural mechanisms can be duplicated within biochemically segregated cell populations, is quite natural biologically. For example very similar toes grow on both our legs, very similar suckers on the eight arms of an octopus, and similar claws on the many legs of a centipede. I advance the suggestion that the brain also consists of a variety of different suborgans within which certain sub-suborgans can readily be duplicated, or nearly so. However, these suborgans are not geometrically separate in the sense that legs are. Rather, I imagine them to be geometrically intermingled, but separated by the biochemically determined pattern of linkages between them.

The model just advanced is far too simple to represent many of the amazing subtleties of visual psychophysics, and indeed requires many immediate extensions and corrections. The presence in the visual cortex of cells with receptive fields of different size may be taken to suggest that signals carried by cells sensitive to 'dots' differing in size remain segregated in early stages of the visual system, and experiments with Glass patterns seem to support this idea. A more adequate model would need to represent this fact. The model shown also fails to represent the fact, noted in our examination of Figure 10 above, that depth detection modulates swirl perception and so is presumably prior to it.

In all of the above the profoundest puzzle is the manner in which global gestalts like the presence of 'swirl' can be represented neurally. The same fundamental question can be asked about many other, more common and basic gestalts that the eye demonstrably forms, e.g. edge percepts having insides and outsides which can affect the perception of their content in many ways. Experiments capable of elucidating the fundamental problem need to be devised.

It may also be asked why we have retained the visual mechanisms necessary to detect the global gestalts considered in this paper, given that they seem to play no discernable rule in the natural life either of humans or of our simian ancestors. The normal assumption is that unused mechanisms are destroyed by accumulating mutations in the course of time. Perhaps the perception of Glass patterns is parasitic on some other, more necessary mechanism of perception, for example perception of coherent motion. It would be interesting in this connection to know whether any part of the population has a genetically determined 'Glass pattern blindness', analogous to the red-green colorbindness which affects 8% of the male human population.