Figure 5.

Graph Convolutional Neural Network Architecture. The network’s input is the the polar map data defined on a graph that encodes the neighborhood relationships between the individual data points. We use two layers of convolution filter each followed by graph coarsening and pooling layers. When training with Chebyshev polynomials we use filters of order 16 in the first layer and 32 in the second. Cayley polynomials of order 4 and 6 are used in the first and second layer respectively. Both models employ a pooling size of 4 in the first layer and 2 in the second.