Visualize image features that maximally activate filters of a CNN
Review
I wanted to try this project to see for myself the various kinds of features detected by CNNs as the network gets deeper. A lot of intuition was given while learning about CNNs but the practical experience is always much better. A pre-trained VGG16 model was taken and the output from a custom layer (one in which the particular filter is present) is selected and the gradient ascent is used to find the image which gives maximum activation from that filter. A trend of increasing complexity in patterns being searched for by the filters in deeper layers of the model, as we had learned in the courses and seen in books.