Acoustic Training Data

This proposal describes the advancement of a novel AI/ML training data prioritization tool specifically developed to enhance submarine detection capabilities in naval warfare. The primary objective is to provide a comprehensive methodology for assessing and managing training datasets for acoustic data analysis of undersea sensing systems. Motivated by the Navy's pressing need for solutions capable of efficiently handling extensive acoustic datasets, this initiative aims to ensure training data remains diverse, representative, and optimally curated, thus significantly streamlining AI/ML training processes while reducing associated costs in data collection, storage, and processing.During the Phase I effort, Arete successfully developed a two-tier data filtering model that leverages Knowledge Distillation and Latent Analysis (KDLA) in combination with Cluster Analysis-informed Diversity-Difficulty (CAD2) Data Pruning techniques. Utilizing surrogate sonar datasets closely analogous to customer operational data, specifically Wide-Band Forward-Looking Sonar (WBFLS) and Multi-Function Side-Looking Sonar (MFSLS). the Phase I effort demonstrated proof-of-concept performance. Validation was accomplished through analysis, modeling, and Receiver Operating Characteristic (ROC) curve comparisons, confirming that Arete’s methodology effectively reduces training dataset sizes while preserving or enhancing AI/ML model accuracy.Building upon Phase I successes, Phase II will focus on algorithm maturation utilizing actual customer-supplied data and comprehensive performance evaluations against currently deployed operational models. Additionally, Arete’s Phase II roadmap includes the development and deployment of an integrated, end-to-end software solution, incorporating SonarSIV technology to meet the Navy's operational requirements.