Training a randomly wired net is like pretending that the training data won't give it pre-conceptions, when the training data is what drives the model.
(Pretending that closing ones eyes, removes all items from existence)
Training a randomly wired net is like pretending that the training data won't give it pre-conceptions, when the training data is what drives the model.
(Pretending that closing ones eyes, removes all items from existence)
Sussman has written a whole program to play Tic-Tac-Toe: two players, a 3x3 grid, players alternate turns, the game ends under such-and-such conditions, etc. etc.; and then he thinks that by such a superficial gesture as wiring the net randomly he somehow creates a tabula rasa with no "preconceptions of how to play." Minsky points out, by an equally superficial gesture, that giving up control over one minor aspect of the program (the wiring, the ability to see with one's eyes) really has nothing to do with the issue at hand (how much "understanding" of Tic-Tac-Toe has been baked into the program, how many things exist in the room).
Or, to put it back into koan form:
One day, Sussman wrote a program that had no preconceptions of how to play Tic-Tac-Toe. The following day, Sussman wrote a program that had no preconceptions of how to play chess. His two programs were not the same.