Visualizations for Caffe AlexNet

For each unit in each layer of the network, thousands to millions of image patches are searched to find those the drive the most positive (top) and most negative (bottom) responses. Deconv visualization (Zeiler and Fergus, 2014) is then used to highlight features that drive the response.

Select a layer below, then a unit number on the page that appears, and click on the images to toggle between the original image and the visualization. For units that have a limited maximal receptive field (those in Conv2 to Conv5), a red box is superimposed that marks the image patch that could potentially be connected to the unit:

Conv2

Conv3

Conv4

Conv5

FC6

FC7

FC8

List of Categories for the 1000 output units in layer FC8.