3 ms·

Intriguing Properties of Neural Networks [pdf]

• SlipperySlope 12y ago
  This paper demonstrates that deep neural networks have surprising blind spots when inputs are only slightly perturbed in a certain manner. I wonder if these algorithmically determined adversarial examples can be fed back into the network in the training set with correct tagging to make the network more robust with regard to blind spots?
  • simonhughes22 12y ago
    That's a good idea but there's other papers where they try that. Note that this paper is almost a year old now.
  • anko 12y ago
    I think the problem will remain because the "bits" (perceptrons etc.) of representation are always less than the sum of the inputs. So the training is lossy, which also means not perfect. As humans, we probably have less "blindspots" because we have learned extra double checking mechanisms, such as applying logic and our world knowledge to analyzing an image. We still fall prey to optical illusions though.
• jostmey 12y ago
  The reason the Neural Networks are suffering from blindspots is because they were never trained as generative models. A straightforward classifier cannot be expected to correctly identify something it has never seen before---so of course you can find adversary examples. Generative models are computationally more expensive to build, which is why they are not always used.