Investigating parameter importance for different ocean environments using Fisher Information

In machine and deep learning, we often seek a simple and effective model. Overly complex models may be difficult to train and make inaccurate predictions. One way to find an effective model is to consider the relative impact of parameters on predictions as seen in the Fisher Information. The applicability of the Fisher Information is shown using numerically modeled transmission loss as a function of frequency for different range-independent ocean environments. The sensitivities to source-receiver range and depth and environmental parameters are quantified by calculating the Fisher Information. First, the Jacobian matrix of partial derivatives of transmission loss with respect to each of the model’s input parameters is obtained, then the Jacobian matrix is used to calculate the Fisher Information matrix. An eigenvalue decomposition of the Fisher Information matrix shows that this system is “sloppy,” because it exhibits an exponential hierarchy of parameter importance. In many cases, only a small number of parameters are relevant for explaining the model output but the impact of individual geoacoustic parameter varies with both environment and frequency. Our results have implications for learning algorithms and data collection methods while elucidating the relevant physics for different conditions. [Work supported by the Office of Naval Research.]