Uncertainty-aware classification of underwater acoustic signals using Bayesian deep learning

This talk presents our work on applying Bayesian deep learning methods to provide interpretable and uncertainty-aware classification of underwater acoustic signals. We demonstrate how Bayesian convolutional neural networks can accurately classify both passive sonar and synthetic aperture sonar data while quantifying prediction uncertainty—a critical capability for autonomous systems. By examining estimated aleatoric and epistemic uncertainties, we gain insight into the model's decision-making process and areas of low confidence. We leverage predictive entropy and calibrated uncertainties to identify challenging samples, explain misclassifications, and reveal patterns in model behavior. Our results show how uncertainty correlates with factors like range, aspect angle, and seafloor texture in ways that align with acoustic propagation physics. We also discuss how active learning strategies guided by epistemic uncertainty can efficiently train models while providing further model interpretability. This work establishes Bayesian deep learning as a promising approach for developing explainable and reliable underwater acoustic classifiers, addressing the need for transparency in AI-driven underwater sensing applications.