patrick.net

www.youtube.com/embed/wj8mVe-tPB4

Machine learning yields unpredictable results. It's not a bug. It's pattern recognition. It's just not always the pattern you intended to be recognized.

Case example. The army trained a neural network to determine whether or not pictures contained images of tanks. It had 100% accuracy in the test data, but only 50% accuracy afterwards. The reason? All the test images that had tanks in them were taken on a cloudy day, and the non-tank pictures on a sunny day. The neural network learned to detect cloudy days, not tanks.

Machine learning may involve learning unintended and arbitrary things.

#scitech