Acoustic Training Data Prioritization

This proposal details the development of a novel AI/ML training data prioritization tool specifically designed for enhancing submarine detection capabilities in naval warfare. The project's core objective is to create a sophisticated tool for the effective, comprehensive, and precise assessment of training data for AI/ML models, with a focus on acoustic data analysis in undersea systems. This project is driven by the Navy's need for advanced data analysis tools capable of efficiently handling large volumes of acoustic data, ensuring the training datasets are diverse, representative, and optimally managed.  Arete’s tool will streamline AI/ML training in naval operations by reducing costs associated with large datasets to include time, money, and resources for data collection, storage, and processing. We propose a two-tier data filtering model utilizing Knowledge Distillation and Latent Analysis (KDLA) paired with Cluster Analysis that informs Diversity-Difficulty (CAD2) Data Pruning. The project involves using sonar data for development and proof-of-capability demonstration.  Our sonar data comes from a Wide-Band Forward-Looking Sonar (WBFLS) sensor and a Multi-Function Side-Looking Sonar (MFSLS) sensor.  We will demonstrate our tool's efficacy through in-depth analysis, modeling, and the Receiver Operating Characteristic (ROC) curve performance monitoring.  Areté’s approach will be validated in Phase I, showcasing training dataset size reduction while maintaining or enhancing AI/ML model performance, with a roadmap for extending this technology in Phase II for the Navy's use.