Comments by "Taxtro" (@MrCmon113) on "Gradient descent, how neural networks learn | DL2" video.

  1. Not quite, it is more like identifying a five in something that vaguely resembles a five. Firstly the input was not "random data", but generated specifically to trick the network (it is always possible to make any sufficiently complex learner fail this way). Secondly superstition or pareidolia in humans is not a catastrophic failure, but a bias towards false positives. Moon craters and radio static does resemble other things and if we disregard them, we might have false negatives in other situations. False positives seem to have had a lower cost in our evolutionary history, which is why people tend to see patterns where there are none.
    1