Braid Manifold Discovery using Temporal Graph Networks
Andrew J. Christensen, Ananya Sen Gupta, Ivars Kirsteins · OCEANS 2022, Hampton Roads · 2022
This paper presents initial findings utilizing Graph Neural Networks (GNNs) to perform classification on experimental active sonar data. GNNs enable using neural networks on graphs, which were previously difficult to train on due to the permutation invariant property of graphs. Nodes in a graph are formed by thresholding sonar ping spectrograms. Edges in a graph are formed by first calculating the correlation of the remaining node values across multiple ping spectrograms, and then assigning edges between two nodes if the correlation value exceeds a defined threshold. Both the graph’s adjacency matrix and the graph’s node embeddings are then used as input into the GNN. We use a variant of GNNs called Temporal Graph Networks to allow learning on graphs that change over time.