20 ms·

So in concrete terms, does this mean that we show the network an image, choose one layer's output vector, and then back-propagate gradients to the image such th

by gradys 11y ago

So in concrete terms, does this mean that we show the network an image, choose one layer's output vector, and then back-propagate gradients to the image such that the direction of that vector stays the same, but the magnitude increases?

murbard2 11y ago
That is my understanding of the blog post, yes. That plus a prior on the input pixels to keep it image-like.