Enhancing Machine Learning using Synthetic Data Pipeline for Coherent Underwater Acoustic Images

Machine learning (ML), specifically deep learning, works well when abundant amounts of labeled data are available for training. However, in many environments, like undersea remote sensing, obtaining abundant amounts of labeled data is politically and/or logistically expensive. To overcome this hurdle, this work proposes the use of synthetic data to bridge the lack of data gap to improve machine learning performance for undersea remote sensing tasks where little labeled data is often available. Specifically, we work with simulated data provided by the University of Bath to understand the tradeoffs in ML performance versus simulation fidelity, understand the shortcomings of the simulator and determine the #missing physics# of its model, and understand the advantages of training withpurely simulated data in a multi-task learning framework since labels are present and abundant for several aspects of the simulatedscene including object orientation, seafloor environment, distance from sensor, and bathymetry; all scene parameters which may be used in during the machine learning training process. Finally, as a stretch goal, we will investigate the use of reinforcement learning in conjunction with the simulator to teach the learner the ability to manipulate the simulator as to reproduce an input real sonar image of interest thus inverting the real sonar image into the simulator parameter space. Success of our proposed efforts will help the Navy understand how to best leverage simulated data in improving current machine learning systems where labeled training data is sparse, and identify gaps in physical acoustic models which may be resolved in future work in order to generate more fidelitous data for physical simulation, personnel training, and machine learning purposes.