Feature interpretation in underwater acoustics

Deep Learning algorithms such as Convolutional Neural Networks (CNN) have been widely used recently in underwater acoustics applications including target recognition, passive acoustic monitoring of marine mammals, parameter estimation, etc. In many of these applications, spectrogram images are used to train the algorithms to identify the unique features. CNNs suffer from the black box effect due to their large number of parameters and developing methods to transform this black box algorithm into an interpretable one is an active area of research. This talk will highlight the challenges in interpretation of machine learnt features based on acoustic physics of propagation. How such knowledge of propagation physics may be incorporated into neural networks (e.g., through physics informed Neural Networks) would be explored. [Work supported by Office of Naval Research.]