Advanced Deep Learning with Python

Spatial-based convolutions with attention

The second ConvGNN category is spatial-based methods, which take inspiration from the computer vision convolution (Chapter 2, Understanding Convolutional Networks). We can think of an image as a graph, where each pixel is a node, directly connected to its neighboring pixels (the left-hand image in the following diagram). For example, if we use 3 × 3 as a filter, the neighborhood of each pixel consists of eight pixels. In the image convolution, this 3 × 3 weighted filter is applied over the 3 × 3 patch and the result is a weighted sum of the intensities of all nine pixels. Similarly, the spatial-based graph convolution convolves the representation of the central node with the representations of its ...