"Our senses react mainly to how things change in time." This is the key to Minsky's solution to the duplication problem. "Any agent that is sensitive to changes in time can also be used to detect differences. For whenever we expose such an agent, first to a situation A and then to a situation B, any output from that agent will signify some difference between A and B." As a result, only one agent is required to detect a difference.
NB: I'm quite interested in this as I write these notes (Aug. 16/98). For a long time now I've thought that there is an important relationship between motion detection mechanisms and learning in PDP networks, and I've been exploring some ideas which should lead this fall to the development of a new architecture. Here's the cool idea -- finding a way in which a network could learn to detect differences!