Fig. 9: Illustration of the hierarchical feature representation within CNNs.
a Comparison of concepts contained in different network levels (neuron, channel, and layer)158. Figure reprinted from ref. 158 under the CC BY 4.0 license156. b Visualization of CNN filters from different layers of an imageNet96 pretrained VGG16 network97. Note that the filter representations become more and more complicated as filter positions move deep along the network. The early layer filter representations are relatively primitive, while the top layer filter representations are highly complicated. The VGG16 architecture is plotted using the plot neural net159 library. Filter representations are visualized using the convolutional neural network visualizations150 library.
