Braitenberg now introduces a number of design innovations related to visual processing. He proposes that sensors be organized into a regular configuration, which permits the representation of relevant spatial information. He also proposes "visual processing connections", which include motion, object, and velocity sensors, and which include lateral inhibition. Finally, he proposes the use of multidimensional spatial representations.
For Braitenberg, a "map" is a topographic arrangement of sensory input. "You will get a picture. It will not be scrambled information about the outside world; it will be a representation of the order of things, of their neighborhood relations and, roughly, of the distances between them." In other words, the responses of adjacent sensors now becomes important *because* of their adjacency.
By feeding the signals of many of these sensors into a single threshold unit, an object detector can be built. (Question: Can you think of the limitations of this approach?). Movement detectors can be built by including a delay device into the system; this is exactly the logic underlying Reichardt's influential models of motion detection that were introduced in the 1950s. Velocity sensors are also possible. Inhibitory and excitatory connections can also be arranged in a pattern so that threshold units are under the influence of lateral inhibition, which will result in the vehicle paying little attention to uniformity, which is typically uninteresting. "It is quite clear that these tricks, and a number of other tricks that you might invent, are only possible when there is an orderly representation of the `sensory space' somewhere in the body of the vehicle."
Importantly, the internal spatial representations that are built into the vehicle could be very complicated -- where a detected object can be thought of as a single point located in an N-dimensional space. Such spaces are very difficult to conceptualize visually.
"On the other hand, it is quite easy to imagine or to draw networks of more than 3 dimensions" -- the basic move is to embed dimensions into nodes of a representation that we can visualize. Braitenberg uses as his example a 4-D network, where three dimensions are used to specify the location of a cluster of units, and the fourth dimension is then used to specify which unit in the cluster is to be identified. "Now, you could even build the network, or a piece of it, out of spheres and wires: you would be able to hold in your hands a structure that is intrinsically 4-dimensional, though of course collapsed (`projected') into the 3 dimensions of space in which your hands move." In other words, networks permit construction of spaces for solids that are impossible to imagine.
Now, the issue becomes whether if such a network is built into a vehicle, does it have some concept of space? Braitenberg suggests behavioral observations might be important here in deciding this issue. For instance, move the vehicle to a new location. Does it move back to a preferred location? And if so, does it follow a specific trajectory? Indeed, the path that is taken might give information about the kind of space internally represented in the vehicle. Roger Shepard has used this concept in his study of "mental manifolds" for the representing of spatial information in mental imagery and in motion perception.
Braitenberg goes on to argue that such spatial representations could indeed be built into vehicles -- 2D sheets, 3D blocks, 4D nets -- where distances between points in a specific spatial direction ultimately are represented in terms of an electric pathway between points.
"The point I am making is more than just convenience of construction. It provides for easy tests of reality." For Braitenberg reality is the relationship between internal and external space. "In our vehicle, just as in the physics of relativity, the recognition, or even the existence, of objects is related to the dimensionality of space, internal and external."