Stardate 20011113.1448 (On Screen): Chris is writing a series of articles about the problems of Artificial Intelligence. The second of five is now out, and spends some time talking about neural nets.
Neural nets are an interesting approach and it is indeed the case that they can do some kinds of pattern recognition -- but only if you train them correctly. There's a classic story of an attempt to use neural nets for automated analysis of aerial reconnaissance photos to try to spot the presence of Warsaw Pact tanks. So they decided to train it with pictures that they knew contained such tanks and others containing pictures of NATO tanks, so that it could learn to tell them apart. Then they started feeding it test pictures, and it seemed to get the wrong answer far too often. Finally someone analyzed the contents of the net to find out what it thought it was seeing. It turns out that the training pictures of Warsaw Pact tanks had all been taken on an overcast day, whereas the training pictures of NATO tanks had all been taken in bright sunlight, and the neural net had learned that NATO tanks had shadows, whereas Warsaw Pact tanks didn't have. (discuss)
