As we saw in a previous post, some challenges caused in explaining neural network decisions can be overcome via canonization.
However, models with complex topology can pose other challenges in the canonization process. In this case, a focused approach on the network structure needs to be taken. In this post, we will demonstrate the canonization of DenseNets.
For example, in the basic DenseNet architecture, the neural network is defined as a sequence of Dense blocks. Each Dense block includes Dense layers, each of which is composed of a sequence of Batch Norm (BN), ReLU and convolutional layers.