Cost–Benefit Analysis of Metavariable Variation in Convolutional Autoencoders Applied to Acoustic Backscattering Data from Small Underwater Targets

With an end goal of physically interpretable results, we build on progress towards finding generalized low dimension encodings of acoustic backscattering data from sonar measurements of small targets submerged in water. The data set in use is TREX-13. The initial effort succeeded in learning non-invertible mappings (encodings) of the data to a low-dimensional vector space with convolutional autoencoders and sparse convolutional autoencoders. The prior work largely used default values for various “metavariables” according to the Keras and Tensorflow library specifications. Mitigating the failure rate of network training as well as ancillary increases in signal reconstruction accuracy and network training speed are the goals of our current efforts. The metavariables in question are the number of convolutional kernels in each network layer, the layer count, the size of the convolutional kernels, the technique for randomly initializing the values of the kernels, and the choice of nonlinear activation functions between network layers. We perform some cost benefit analysis for the modifications in terms of training time, reconstruction accuracy, and GPU memory utilization, and decide whether to continue using the variations in further research.