Graph representation learning on braid manifolds

Abstract

The accuracy of autonomous sonar target recognition systems is usually hindered by morphing target features, unknown target geometry, and uncertainty caused by waveguide distortions to signal. Common “black-box” neural networks are not effective in addressing these challenges since they do not produce physically interpretable features. This work seeks to use recent advancements in machine learning to extract braid features that can be interpreted by a domain expert. We utilize Graph Neural Networks (GNNs) to discover braid manifolds in sonar ping spectra data. This approach represents the sonar ping data as a sequence of timestamped, sparse, dynamic graphs. These dynamic graph sequences are used as input into a GNN to produce feature dictionaries. GNNs ability to learn on complex systems of interactions help make them resilient to environmental uncertainty. To learn the evolving braid-like features of the sonar ping spectra graphs, a modified variation of Temporal Graph Networks (TGNs) is used. TGNs can perform prediction and classification tasks on timestamped dynamic graphs. The modified TGN in this work models the evolution of the sonar ping spectra graph to eventually perform graph-based classification. [Work supported by ONR grant N00014-21-1-2420.]