The Neural Net Tank Urban Legend - Gwern.net

Let's consider a more sophisticated example, that of determining whether a tank is hiding in a photograph. A neural net can be configured so that each output value correlates to exactly one pixel. If the pixel is part of the image of a tank, the net should output a one; otherwise, the net should output a zero. The input information would most likely consist of the color of the pixel. The network would be trained by feeding it many pictures with and without tanks.