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?
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.
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?