Artificial data augmentation using braided human features for underwater acoustics

Abstract

Underwater acoustics datasets provide rich environmental information that can potentially be mined using popular machine learning architectures. However, such knowledge discovery is typically limited by the high feature uncertainty, lack of robust ground truths, limited availability of public-domain data, and difficulty of reproducing experimental conditions in field experiments. This poses a data science challenge to study persistent target features, as they can only be learnt and classified robustly if large scale robust training and testing datasets are available. This “nano-data” problem can be potentially solved by augmenting the training and testing data repositories with a two-pronged approach: (i) physics-driven simulations of desired features, which offer interpretable ground truths but typically cannot emulate practical open-sea experiments; (ii) geometric-proxy augmentation using creatively constructed non-domain datasets that emulate the desired feature geometries and environmental effects. For example, a skilled dancer can emulate specific features geometries through geometric contortions of the human body. Such choreographed dance movements can be reproduced reliably against different types of stage environments, structured and unstructured, to emulate the oceanic environment under different conditions. We will present preliminary results in data augmentation from both approaches and discuss the trade-offs between physics-oriented simulations and geometry-driven proxy augmentation. [ONR grant N000142312503.]