unlike logical inference, the inner workings of a neural network are inscrutable. Most of the millions of hidden units don’t stand for any coherent concept that we can make sense of, and the computer scientists who train them can’t explain how they arrive at any particular answer.
