ON THE BENEFIT OF MULTIPLE REPRESENTATIONS WITH CONVOLUTIONAL NEURAL NETWORKS FOR IMPROVED TARGET CLASSIFICATION USING SONAR DATA

The benefit of using multiple representations of data in the context of convolutional neural networks (CNNs) is demonstrated. We present three variations on this theme of multiple representations, in the form of (i) fundamentally different input data representations obtained from the same raw data, (ii) isometries of a given data representation, and (iii) intermediate representations arising from unique CNN architectures. Taken together, these variants can produce excellent classification performance while relying on orders of magnitude fewer free parameters than used in typical CNNs, thereby reducing training data requirements. The value of this multi-representation approach is demonstrated on a target classification task using real, measured sonar data collected at sea.