Refining the metavariables of convolutional autoencoders for better performance on littoral acoustic backscattering data

Abstract

Convolutional autoencoders have been used in an effort to learn latent encodings of littoral accoustic backscatter, but previous work did not explore a wide range of metavariable variation. Only the dimension of the encoding space, and the learning rate of the training optimizer were adjusted to achieve a functioning initial result. The other metavariables were assigned with some justification but optimality was not verified. We explore the effects of changing the method by which random weights are initialized as well as doubling the amount of random kernels per convolutional layer of the autoencoder. Additionally, we vary the layer count of the network, and the shapes of the kernels in the layers. Finally, various nonlinear activations are compared against the activation used initially. The results based on experimental field data from the TREX13 experiment will be presented. [This work is sponsored by ONR grant N00174-20-1-0016.]