Advanced Deep Learning with Python

Dilated convolutions

Recall the discrete convolution formula we introduced at the beginning of the A quick recap of CNNs section. To explain dilated convolutions (Multi-Scale Context Aggregation by Dilated Convolutions, https://arxiv.org/abs/1511.07122), let's start with the following formula:

We'll denote the dilated convolution with *_l, where l is a positive integer value called the dilation factor. The key is the way we apply the filter over the input. Instead of applying the n×n filter over the n×n receptive field, we apply the same filter sparsely over a receptive field of size (n*l-1)×(n*l-1). We still multiply each filter weight by ...