Differences in regression, classification, and multi-task deep learning on pressure time series for range and seabed type

Estimating source-receiver range and predicting seabed type are important problems in underwater acoustics. Deep learning solutions for simultaneous predictions have been previously shown some success for these problems due to their ability to learn patterns from large datasets. However, one important question to consider in deep learning is how predictions should be made: A network can be configured to classify or regress predictions. In this study, that question is explored by comparing predictions from networks trained on simulated SUS charge pressure time-series configured to either regress or classify range and seabed predictions from measured data taken with the IVAR system during the 2017 Seabed Characterization Experiment. To further this inquiry, networks configured to regress range and classify seabed type (via “multi-task” learning) are also explored. Separating the two predictions proves to be useful, as the networks using multi-task learning perform better at predicting range and seabed class simultaneously than those configured only for classification or regression. The results of this experiment illustrate the need to use the proper type of network outputs depending on the desired predictions. [Work supported by the Office of Naval Research.]