Visualization of deep conv net

In week 4, neural style transfer, I don’t quite understand how Andrew visualized the “patches” of the pictures that maximized the activation of a certain hidden unit. Especially the ones in deeper layers, can you trace it back to the original pixels that were relevant to a certain hidden unit?

1 Like

Hey so what the pictures here mean are not traced back pixels from the network. The visualisation just shows what kind of pictures produce the maximum activations. So it could simply be a set of pictures tested in the network to see which of them have the highest activation values in the 5th layer here. We can check the individual activations of the layers for any given input picture if you meant to ask that.


Thanks for your reply, that makes sense for the later layers. Does it also apply to for example, layer 1, (far left in picture)? It doesn’t look like they are complete pictures but part of picture that contains edges.

Since these are purely convolutional layers, you could pass crops of images. Perhaps what this is meant to show is, the crops have the highest activation for the earlier layers rather than the full images.