0

when I used yosinski/deep-visualization-toolbox to visualize the conv feature of mnist, I found that the deconv feature is nothing, the left-bottom of the picture enter image description here but the visualization of imagenet is what I want to see

enter image description here

so I am wondering what the problem is? the toolbox's visualization problem or the special thing of mnist's feature?

Hossein Kashiani
  • 330
  • 1
  • 6
  • 18
wang
  • 55
  • 1
  • 11

1 Answers1

1

I use darkon to visualize, it looks good

enter image description here

top is heatmap, bottom is feature map

when feed a noise image, the net get a wrong label '5', we can see that the heatmap and feature map get a wrong location:

enter image description here

wang
  • 55
  • 1
  • 11