PereGaea

Laslo Godel

2: VISION

We don't really know quite what senses the Dynism might eventually evolve, whether they will include those we are familiar with such as vision, sound, smell; or others difficult to imagine. But let's assume for the sake of simplicity that it will acquire the sense of vision, and concentrate our attention on that exclusively. Although it may well become the Dynism's most complex sense, it will at least by its nature be easier to illustrate.

We will also assume at this stage that it is monochrome only. I will use color however to make it easier for you to see how it operates.

To begin, let's assume that the Dynism has by now extended its simple photocell into a hemispherical `retina' made up of hundreds of them. However, as well as registering the actual brightness of the light they receive, the cells now also register sudden changes in it. If an object moves in such a way as to produce such a sudden change, a reflex uses the retinal positions of the cells that detect it to turn the Dynism towards the object and move it in its direction.

This reflex eventually leads to the Dynism evolving a new mechanism we'll dignify with the name `Visual Processor'. Through a `gang-switching' sub-mechanism, this centers an 8 x 8 array on any stimulated such cell or group of cells and extracts an 8x8 Input Pattern from them.

These cells now also become sensitive to eight different levels of light intensity, which we'll call `tones'. Each cell must be at least 50% covered by a particular tone to register that tone, otherwise it will register no tone at all. The Processor can now extract eight Input Patterns from the same part of the Dynism's visual field, one for each tonal level.

To illustrate, let's imagine that the continuous shading a sphere reveals to our eye when it is illuminated by a point source can be translated into eight tonal `stimuli' something like this:

We'll also imagine that a Retinal Array happens to capture such a sphere to extract a crude eight-tone image. The Visual Processor can then extract a set of eight Input Patterns, one for each tone. As you can see, the first pattern captures the tone of the background, the others the tonal stimuli of the sphere. Since each Pattern can only consist of black and white squares, the Processor must attach `Tone Flags' to them that records each tonal level. These Tone Flags are nothing more than three-place digital numbers like those we saw earlier; I'll come back to how they are used later on.

The Dynism now attempts to identify the object by matching the Input Patterns to Templates as before. However, in order to match like with like, it must `standardize' both by forming them according to certain rules. These are:

RULE 1: There can be only one pattern per frame. Black squares singly or in groups not attached to the pattern by at least one side or corner are deleted.

RULE 2: A pattern must fit its `frame' with respect to at least one dimension. As you can see in this example, the Dynism will have to move closer to the `target' object to extract an Input that fits in this way.

RULE 3: Patterns must be centered. Those that are an odd number of squares wide or high will be biased to the left or bottom of the frame respectively.

RULE 4: Any white squares fully enclosed by black squares will themselves be converted into black squares. Such squares represent stimuli in other tones and will have their own patterns extracted from them.

In order to conform to Rules 2 and 3, the Dynism must usually be moving directly towards a stimulus. Indeed, just as most terrestrial sharks must keep swimming to keep breathing, the Dynism must keep moving to keep seeing. To keep the stimulus centered in its retina, it continuously extracts `trial' patterns as it goes. A Reflex Loop can then alter the Dynism's angle of approach according to the positions of these patterns within their frames.

The Dynism begins to minimize the need for `perceptual motion' by subdividing its retina into a much larger number of tinier cells. The Processor can then gang-switch these into groups of four, eight, 16 or even larger powers of two to correspondingly double, quadruple, or otherwise increase the size of an array. It can then extract a pattern from a stimulus at several distances, though some motion may be needed for an exact fit at points in between.

To reduce the need for motion further, the retina sub-divides into much larger numbers of tinier cells still. However, smaller cells on an external retina entails greater risk of damage from the environment. So let's assume at this point that the mantle of `Dynism' is now taken over by another species that solves this problem by evolving an eye with a single lens and an internal retina. Although this produces an inverted image on the retina, matching is not affected since the Templates are simply stored in its ROM the same way.

The advantages of this eye are considerable. For instance, instead of the arrays being gang-switched across the retinal surface to capture a stimulus, a Centering Reflex responding to a light fluctuation can move the entire eye to place the stimulus in the center of its retina. This allows the retinal arrays to remain stationary so that they form a set of Chinese Boxes, one inside the other. The largest, Box 0, subtends an arc we'll assume for the sake of illustration to be 128 degrees. Box 1 subtends an arc of 64 degrees, Box 2 32 degrees, and so on down to Box 7 which subtends an arc of just one degree. As you can see, the center 16 cells of each Box are each made up of four smaller cells from the Box within it acting in unison.

To be able to fit a stimulus closely enough to eliminate the need for motion entirely, the photocells from which each Box is made are themselves formed from 32x32 arrays of subcells. These allow a Box to shrink to nearly half its size half a cell at a time to fit a stimulus. It only need shrink to this size since capturing smaller stimuli is the role of the immediately smaller Box, the inner colored square. To ensure a stimulus does not extend beyond its boundaries, it also extracts a trial 32x32 pattern derived from the subcells themselves, then checks that this is wholly contained within the outer colored square. The two colored squares therefore define a `Limit Zone' (yellow) within which a stimulus must fit in at least one dimension before the Box can extract an Input Pattern from it.

Nor need a stimulus be matched to its template with total accuracy, within a few squares will do. These can be bunched together, or spread throughout the template with little risk of mismatch.

This flexibility becomes possible through the new mechanisms which now evolve as described below.

How does a Dynism identify stimuli that are partially obscured? By acquiring a mechanism that allows it to identify groups of them in the near vicinity. We will call such a group an `object', and the mechanism itself an Object Identifier.

To see how this works, let's imagine one of its Prey species actually looks like this (albeit somewhat contrived) object. As you can see, the further away its stimuli lie from its center, the more distorted they become. The Dynism can therefore only match the center ones - unless it evolves predistorted Templates for the others. Also, several `stimuli' that we see as `complete' are bisected by the tones of the underlying sphere, preventing the Dynism from matching them. Others may be common to other objects, so that they can't be used as `identifiers' even when they are successfully matched. The three stimuli here may well be the only ones that will allow the Dynism to Identify the object as an object, perhaps as a member of a particular Prey species.

But in order to capture any set of identifying stimuli quickly enough to identify a Predator object, the Dynism's eye must become able to extract patterns from several stimuli all but simultaneously. A Reflex Loop may evolve to move the eye in `raster scan' fashion over its Visual Field, but such an eye would quickly be superseded by one that scans the Boxes across the retina instead, just as its external-retina predecessor did.

To speed things up, they needn't scan it in its entirety. Each Box need only scans its immediately larger one. When it completes its scan, it moves to the center of the Retina so that the next smaller Box can scan within it like this:

How though can such a scanning Retinal Box adjust its size to fit the stimulus? This would surely make the Retina impossibly complex and the pattern extraction process once again lethally slow.

The Visual Processor must therefore take over as many of the functions of the Retinal Box as possible. When the Retinal Boxes stop at their scan-points, they no longer either shrink or extract patterns. Instead they copy whatever image fragments they receive direct to a stationary `Centering Box' within the Processor itself. The sole task of this Box is to fit itself to a stimulus in the way we saw earlier. If it succeeds, it copies the stimulus to a `Virtual Box', which then does the actual Input Pattern extraction ready for matching to a Template.

This new eye and it Visual Processor allow the Dynism to acquire many other survival-enhancing capabilities. For instance, it can now also distinguish between one or more Predator or Prey Objects in a cluster of them. To illustrate this, let's imagine that the Dynism's field of view now encompasses this segment of PereGaea's Ocean - which perhaps owes more to Joan Miró than Earth. The cubic object, in moving into this position here, produced a light fluctuation which caused the Dynism's eye to center its Retinal Boxes on it instead of the sphere. This allows the smallest Box I've drawn here to capture most of the Cube's undistorted identifying stimuli as it scans within its larger one. All the other boxes, where they capture valid stimuli at all, will probably only capture unidentifiable ones; the sphere tonings in the case of the larger, or the Cube's fine surface details in the case of the smaller.

All the Stimulus Templates in the ROM now have Object Numbers attached to them. Those that belong to the same object each have the same Number. To keep things simple we will assume that these Numbers have four digital places so that the Dynism will be able to distinguish between 16 objects.

These Object Numbers then allow a Dynism to detect sufficient visual stimuli to Identify and respond to an object that is partly obscured by another.

Certain stimuli may be presented by more than one object however. This is especially the case with `universal' shapes like `lines', squares, disks, and triangles.

If an object has an overall shape similar to these universals, its Object Number can have a `Universal Flag' attached to it indicating this to aid in its identification. These flags are similar to the Tone Flags we saw earlier.

The three-dimensional equivalents of universals like cylinders, cubes, spheres and pyramids are objects in themselves however and require their own Object Numbers. More advanced Dynisms may attach such Numbers however as Universal Flags to other more essential Object Numbers.

Dynisms are not at this stage able to identify objects like `rocks' or `cliffs', they can still only identify Objects that contain at least three unique stimuli such as presented by Predators or Prey.

`Attribute Flags' such as the Tone and Universal Flags we have just seen quickly become a major feature of the Dynism's Object Identifier For instance, the Dynism's ability to match a stimulus independently of its distance, oddly enough, prevents it from determining that distance. And to make things even worse, if the Dynism cannot determine distance, it cannot determine size. Determining such attributes are essential to a Dynism's survival.

Let's assume that most PereGaean objects, like many on Earth, tend to have limited size ranges. More advanced Dynisms can then in effect `learn' their sizes via the experience of evolution and attach Size Flags to all their ROM templates. It can then use these to determine distance.

Like the Boxes, the Numbers of these Flags range from 0 to 7. If a Stimulus fills the largest Box, Box 0, at a single unit of distance based on the focal length of the eye's lens like this, it will have a `zero' Size Flag attached to it. If this stimulus is then moved to two such Units, it will then fill Box One, and if it moves to four, then Box Two.

Conversely, a Size Three stimulus will fill Box 3 at Distance One, a Size Two will fill it at Distance 2, Size 1 at Distance 4, and so on. The smallest stimulus the Dynism can perceive, a Size Seven, can only have a pattern extracted from it if it is no more than one Unit away from the lens of its eye.

Now, when the Visual Processor extracts a Pattern from a stimulus, it also attaches a `Box' Number Flag to it to indicate which Box was used to extract it. A Distance `Tester' can compare this Flag to the one attached to the Template it is matched to and extract a `difference' Number Flag. If a stimulus's Box Flag is smaller than its Size Flag it is a short distance away; if the Flag is larger the stimulus is distant. Difference Flags therefore become Distance Flags.

Up until now the Dynism has only been able to respond to objects presented to it at a single rotation, what we might describe as a `vertical' one. In order to perceive objects at other rotations, it must become able to match component stimuli that have been rotated relative to its line of sight. As this drawing shows however, the Dynism need only evolve a few rotated templates for each stimulus, in this case perhaps only the first, the third, and the fifth, with little risk of mismatch. It may also only need them for those stimuli most significant to it. Rotation Flags are also added to these Templates so that the extent of a stimulus's rotation can be perceived if necessary.

This capability may then evolve into one that allows a Dynism to determine an object's orientation. As I said earlier, certain objects will contain stimuli that are more significant to the Dynism than others. Obviously such an object must be oriented the right way before before the Dynism can perceive and respond to such stimuli, and up until now this has been a matter of chance. Those Dynisms that evolve the ability to reorient objects or adjust their own spatial position to reveal such stimuli will acquire an important evolutionary advantage.

When an object rotates around the x or y axes, various stimuli come into view, undergo certain changes determined by the object's three-dimensional shape, then disappear. The Dynism's nearest-matching system only permits a stimulus to alter in shape to a limited extent before the pattern extracted from it will be mismatched.

Let's suppose an object can be enclosed in a spherical grid made of 18 near-equal `domains' like this at some arbitrary orientation. These grid lines represent the boundaries beyond which alternate templates for a particular stimulus become necessary.The Dynism's Visual Processor attaches an Orientation Flag to each of its Object Numbers according to the stimuli visible at each grid domain.

Looking now at all the Attribute Flags we have now so far seen, each Object Number may come to have several `Attribute Flags' attached to it; size, distance, rotation, orientation, as well as the Tone and Proportion Flags we saw earlier. A Flag Tester can then use these Flags to select between the several Output Patterns that may now be attached to a single Object Number.

A Dynism can now respond either to a single stimulus or an entire object. It may for instance flee from a Predator whatever stimulus it presents, but select its response to a Prey according to whether its eye happens to fall upon a defensive mechanism, or some kind of defensive shield. The Dynism in effect behaves as if it was applying a new form of Stimulus Response to determine how to respond to a particular object in a particular circumstance. The Dynism's ROM is now made up of constructs like this:

Obviously, since the Dynism's Visual Processor really only deals with numbers, they need be little different from Earth's microprocessors. However they are still relatively crude and error-prone, so Dynisms will still need a high reproduction rate to enable the species to survive. This makes them unsuitable for use in the conventional robotic world here on Earth where only a few robots are built at any one time, often only the one. In the next chapter, we will look at the evolution of visual mechanisms which are rather more efficient than those outlined here.