Design of Everyday Things [65]
THE CONNECTIONIST APPROACH
We still are a long way from understanding human memory and cognition. Today, in the developing field of cognitive science, two different views are emerging. The traditional view considers thought to be rational, logical, and orderly; this approach uses mathematical logic as the scientific means to explain thought. Adherents of this approach have pioneered the development of schemas as the mechanism of human memory. A newer approach is rooted in the working of the brain itself. Those of us who follow this new approach call it “connectionism,” but it also goes under the names of “neural nets,” “neural models,” and “parallel distributed processing.” It is an attempt to model the way in which the brain itself is structured, with billions of brain cells connected into groups, many cells connected to tens of thousands of others, many all working at the same time. This approach follows the rules of thermodynamics more than it does the rules of logic. Connectionism is still tentative, still unproven. I believe that it has the potential to explain much of what puzzled us before, but part of the scientific community thinks that it is fundamentally flawed.7
The brain consists of billions of nerve cells—neurons—each connected to thousands of other cells. Each neuron sends simple signals to the neurons to which it is connected, each signal attempting to increase or decrease the activity of its recipient. The connectionist approach to the study of thought mimics these connections. Each connectionist unit is connected to many other units. The signals are either positive in value (called “activation” signals) or negative in value (called “inhibition”). Each unit adds up the total influence of the signals that it receives and then sends along its outward connections a signal whose value is a function of that sum. That’s about all there is to it. The elements are all simple: the complexity and power come from the fact that there are a large number of interconnected units trying to influence the activities of the others. All this interconnection leads to massive interaction among the units, with the signals sometimes leading to fights and conflicts, sometimes to cooperation and stability. After a while, however, the system of interconnected units will eventually settle down to a stable state that represents a compromise among the opposing forces.
Thoughts are represented by stable patterns of activity. New thoughts are triggered whenever there is some change in the system, oftentimes because some new information arrives at the senses and changes the pattern of activation and inhibition. We can think of the interactions as the computational part of thought: when one set of units sends signals activating another, this can be interpreted as offering support for a cooperative interpretation of events; when one set of units sends signals suppressing another, it is because the two usually offer competing interpretations. The result of all this support and competition is a compromise: not the correct interpretation, simply one that is as consistent as possible with all possibilities under active consideration. This approach suggests that much of thought results from a kind of pattern matching system, one that forces its solutions to be analogous to past experiences, and one that does not necessarily follow the formal rules of logical inference.
The relaxation of interacting connectionist structures into patterns happens relatively quickly and automatically, below the surface of consciousness. We are conscious only of the end states, not of the means for getting there. As a result, in this view of the mind, our explanations of our own behavior are always suspect, for they amount to stories made up after the fact to explain the thoughts that we already have.
Much of our knowledge is hidden beneath the surface of our minds, inaccessible to