3 ms·

This is really great research and interesting: (very roughly) how to compute a very small mask which, when applied to any image, makes the neural network miscla

• hammock 10y ago
  It says these universal vectors are the same across different classifiers. Why would that be?
  • morenoh149 10y ago
    I'm not an expert but it seems these perturbations are fiddling with the fundamental approach used with a NN. Mainly, that a NN works in layers. So these perturbations must be messing up the lowest layers and then the higher layers end up generating the wrong features and ultimately the model misclassifys. See http://i.stack.imgur.com/jpYdN.png http://i.stack.imgur.com/jpYdN.png
  • Eridrus 10y ago
    They're still trained on the same dataset, so these neural nets should be learning roughly similar distributions.