“despite years of work we have failed to find neural network models that can be considered robust on MNIST from the point of human perception." important conclusion from @bethgelab, h/t @dileeplearning https://arxiv.org/abs/1805.09190
-
I don’t think NNs are in danger of abandonment again because they have more than proved their utility, and many more potential applications remain. (I have partially experienced the other side. In 2003, the only NN class at Stanford was by Widrow and it had only ~15 students)
i for one have repeatedly said that neural nets are useful tool, and never argued that they should be abandoned. (cf the straw version of my claims that people often like to challenge.)
