Achim Schneider wrote:
Erm...
There is this story about some military (US afair) training a neural net to detect tanks in images, I can't find the link right now.
It worked, with amazing 100% accuracy.
Then they threw another batch of images at the net.
It worked, with devastating 50% accuracy.
It turned out that in the first batch, all the pictures with tanks on them were shot in sunny weather, all the pictures without tanks were shot with cloudy sky.
I have some similar stories to tell. I don't have references or links, but they are reproducible - in fact you have probably experienced them first-hand. A. Students were trained to do math. At first, they did well with high accuracy. Then more math was taught and tested, and they did poorly. It turned out that the first batch was algebra, where -(x+y) = -x + -y z*(x+y) = z*x + z*y and the second batch was trig, and students assumed sin(x+y) = sin(x) + sin(y) B. A friend of mine is a safe, attentive driver. But one day, the car in front of us stopped, and he almost didn't stopped, I had to remind him. It turned out that normally (by legal requirement) any car that brakes must put on red light at the back. But that day, that car, it showed no red light.