In previous vehicles, Braitenberg used "Mnemotrix" wire to implement one of the basic laws of association, the law of simultaneity. In this latest vehicle, Braitenberg's innovation is to implement a second important law of association, the law of contiguity. He uses "Ergotrix" wire to code the association between events that are temporally successive (e.g., lightning & thunder).
Braitenberg's goal for vehicle 11 is to provide it a brain that provides a better model of the world than the models available for simpler vehicles. In particular, he desires this model to be richer in the sense that it includes dynamic (or at least temporal) aspects. Specifically, we need to capture associations between events that reliably occur in temporal succession.
The key design need here is for an antisymmetric association. For example, if event A usually precedes event B, but B rarely precedes A, then we require that "when the representative A is activated by the input, the representative B is activated by the connection, but not vice versa." Here Braitenberg invents "Ergotrix" wire, which has the same properties as Mnemotrix, but which only allows signals to be conducted in a single direction. (In this sense, Ergotrix is much more analogous to real synapses than is Mnemotrix.) This should provide an increase in the complexity of the vehicle's behavior, because the world in which the vehicle operates provides rich opportunities for temporal learning.
At this point, Braitenberg raises an interesting issue pertaining to the design decisions that he has been using for his vehicular development. It is a version of the principle of least commitment. You see, with his invention of Ergotrix, Braitenberg could go back to all his vehicles equipped with Mnemotrix, and replace each Mnemotrix wire with two Ergotrix wires (one running in one direction, the other running in the other). But, he doesn't want to do this. He offers two reasons. "First of all, we don't want to go back in evolution and change things that have already provided to be convenient, since we might lose some advantage that we have not even realized (Remember the law of uphill analysis and downhill synthesis: we run the risk of not understanding any longer what we previously put together.)"
Second, Braitenberg succumbs to a clear theoretical motivation, namely, the assumption that one should keep the association of properties distinct from the association of events. "It would seem that the process of abstracting things from the environment-- concept formation at the most elementary level -- must occur prior to the process of discovering the dynamic properties of these things." However, Braitenberg understands that the two types of association are interrelated, in the sense that the kinds of dynamic concepts that one uses might dictate the kinds of static concepts, and vice versa. "Abstracting meaningful chunks from the environment (things, events) and discovering the rules of their behavior are two processes that condition each other and are necessarily interlaced, like the learning of the vocabulary and the learning of grammar in a language course."
Braitenberg incorporates this interrelationship in an updated learning rule. Specifically, "whenever the Ergotrix wires become strengthened, the Mnemotrix wires within each of these groups will also become strengthened. Thus concepts are established in the vehicle especially when they appear in regular sequences." (NB: Interestingly, the converse of this learning rule is not used. Is there a good reason for this??)
"But it would take prolonged observation to notice this particular aspect of learning in the vehicles. As a matter of fact, we might not have suspected it if we had not introduced a piece of our own philosophy into the construction of these vehicles. As our brain children become more efficient, we notice that the `law of uphill analysis and downhill synthesis' becomes more and more compelling."
(NB: But should we be skeptical here? Note that Braitenberg might be subject here to being carried away with the potential complexities of the behaviors now possible for his vehicles. After all, much of the complexity depends crucially on associationism. Should we be concerned about potential limitations that are placed on machines that depend (exclusively?) on this kind of learning?
The answer to this question requires some additional reading. Here are a couple of intriguing, old papers that challenge the utility of associationistic principles:
Bever, T.G., Fodor, J.A., & Garrett, M. (1968). A formal limitation of associationism. In T.R. Dixon and D.L. Horton (Eds.) Verbal behavior and general behavior theory. Englewood Cliffs, NJ: Prentice-Hall. Chomsky, N. (1957). Review of B.F. Skinners Verbal Behavior. In J.A. Fodor and J.J. Katz (Eds.) The structure of language. Englewood Cliffs, NJ: Prentice-Hall.
For a balanced picture, a response to the terminal metapostulate problem that is raised by Bever, Fodor, and Garrett (1968) is available in:
Anderson, J.R., & Bower, G.H. (1973). Human associative memory. Hillsdale, NJ: Lawrence Erlbaum Associates. Paivio, A. (1986). Mental representation: A dual-coding perspective. New York: Oxford University Press.
Pearl Street | Vehicles Home Page | Dawson Home Page |