5.7 Imagination

"We don't see things as they are. We see things as we are." — Anais Nin

When Carol picks up one of her blocks, that action seems utterly simple to her: she just reaches out, grasps it, and lifts it up. She just sees that block and knows how to act. No ‘thinking’ seems to intervene.

However, the seeming ‘directness’ of seeing the world is an illusion that comes from our failure to sense its complexity. For, most of what we think we see comes from our knowledge and from our imaginations. Thus, consider this portrait of Abraham Lincoln made by my old friend Leon Harmon, a pioneer in computerized graphics. (To the right is a portrait that I made of Leon.)

How do you recognize features in pictures so sparse that a nose or an eye is but three or four patches of darkness or light? Clearly, you do this by using additional knowledge. For example, when you sit at a table across from your friends, you cannot see their backs or legs—but your knowledge-based systems assume by default that all those body-parts are present. Thus we take our perceptual talents for granted—but ‘seeing’ seems simple only because the rest of our minds are virtually blind to the processes that we use to do it.

In 1965 it was our goal was to develop machines that could do some of the things that most children can do—such as pouring a liquid into a cup, or building arches and towers like this from disorderly clutters of building blocks. [5] To do this, we built a variety of mechanical hands and electronic eyes—and we connected these to our computer.

When we built that first robot for building with blocks, it made hundreds of different kinds of mistakes.[6] It would try to put blocks on top of themselves, or try to put two of them in the same place, because it did not yet have the commonsense knowledge one needs to manipulate physical objects! Even today, no one has yet made a visual system that behaves in anything close to humanlike ways to distinguish the objects in everyday visual scenes. But eventually, our army of students developed programs that could “see” arrangements of plain wooden blocks well enough to recognize that a structure like this is “a horizontal block on top of two upright ones.”

It took several years for us to make a computer-based robot called Builder that could do a variety of such things—such as to build an arch or tower of blocks from a disorderly pile of children’s blocks, after seeing a single example of it. We encountered troubles at every stage but sometime those programs managed to work when arranged into a sequence of levels of processes. (Note that these do not much resemble the levels of Model Six, but do tend to progress from highly specific to very abstract.)

Begin with an image of separate points.

Identify these as textures and edges, etc.

Group these into regions and shapes.

Assemble these into possible objects.

Try to identify them as familiar things.

Describe their spatial relationships.

However, those low-level stages would frequently fail to find enough usable features. Look at this magnified digital view of the lower front edge of the top of that arch:

That particular edge is hard to see because the two regions that it bounds have almost identical textures. [7] We tried a dozen different ways to recognize edges, but no single method worked well by itself. Eventually we got better results by finding ways to combine them. We had the same experience at every level: no single method ever sufficed, but it helped to combine several different ones. Still, in the end, that step-by-step model failed, because Builder still made too many mistakes. We concluded that this was because the information in our system flowed only in the input-to-output direction — so if any level made a mistake, there was no further chance to correct it. To fix this we had to add many ‘top-down’ paths, so that knowledge could flow both down and up.

The same applies to the actions we take because, when we want to change the situation we’re in, then we’ll need plans for what we will do, so all this applies to the Do's of our rules. For example a rule like, “If you see a block, Do pick it up” leads to a complex sequence of acts: before you begin to lift a block, you need to form an action-plan to direct your shoulder, arm, and hand to do this without upsetting the objects surrounding that block). So again, one needs high-level processes, and making these plans will equally need to use multiple levels of processing—so our diagram must become something like this:

Each Action Planner reacts to a scene by composing a sequence of Motion-Goals, which in turn will execute Motor Skills like ‘reach for,’ 'grasp,' ‘lift up,’ and then ‘move’. Each Motor-Skill is a specialist at controlling how certain muscles and joints will move—so what started out as a simple Reaction-Machine turned into a large and complex system in which each If and Do involves multiple steps and the processes at every stage exchange signals from both below and above.

In earlier times the most common view was that our visual systems work from “bottom to top,” first by discerning the low-level features of scenes, then assembling them into regions and shapes, and finally recognizing the objects. However, in recent years it has become clear that our highest-level expectations affect what happens in the “earliest” stages.

V.S. Ramachandran: “[Most old theories of perception] are based on a now largely discredited “bucket brigade” model of vision, the sequential hierarchical model that ascribes our esthetic response only to the very last stage — the big jolt of recognition. In my view … there are minijolts at each stage of visual segmentation before the final ‘Aha’. . Indeed the very act of perceptual groping for objectlike entities may be pleasurable in the same way a jigsaw puzzle is. Art, in other words, is visual foreplay before the final climax of recognition.”[8]

In fact, today we know that visual systems in our brains receive many more signals from the rest of the brain than signals that come in from our eyes.

Richard Gregory:“Such a major contribution of stored knowledge to perception is consistent with the recently discovered richness of downgoing pathways in brain anatomy. Some 80% of fibers to the lateral geniculate nucleus relay station come downwards from the cortex, and only about 20% from the retinas.”[9]

Presumably those signals suggest which kinds of features to detect or which kinds of objects might be in sight. Thus, once you suspect that you're inside a kitchen, you will be more disposed to recognize objects as saucers or cups.

All this means that the higher levels of your brain never perceive a visual scene as just a collection of pigment spots; instead, your Scene-Describing resources must represent this block-arch in terms (for example) like “horizontal block on top of two upright ones.” Without the use of such ‘high-level’ Ifs, reaction-rules would rarely be practical.

Accordingly, for Builder to use sensory evidence, it needed some knowledge of what that data might possibly mean, so we provided Builder with representations of the shapes of the objects that it was to face. Then, from assuming that something was made of rectangular blocks, one of those programs could frequently ‘figure out’ just which blocks appeared in a scene, based only on seeing its silhouette! It did this by making a series of guesses like these:

Once that program discerns a few of those edges, it imagines more parts of the blocks they belong to, and then uses those guesses to search for more clues, moving up and down among those stages. The program was frequently better at this than were the researchers who programmed it. [10]

We also gave Builder additional knowledge about the most usual ‘meanings’ of corners and edges. For example, if the program found edges like these then it could guess that they all might belong to a single block; then the program would try to find an object that might be hiding the rest of those edges.[11]

Our low-level systems see patches and fragments, but then we use ‘context’ to guess what they mean—and then confirm those conjectures by using several levels and types of intermediate processes. In other words, we ‘re-cognize’ things by being ‘re-minded’ of familiar objects that could match fragments of incomplete evidence. But we still do not know enough about how our high-level expectations affect which features our low-level systems detect. For example, why don’t we see the middle figure below as having the same shape as its neighbors?

In an excellent survey of this subject, Zenon Pylyshyn describes several theories about such things, but concludes that we still have a great deal to learn.[12]