A critical bottleneck in machine learning efforts continues to be either the lack of sufficiently sized, fully curated data-sets, or availability of the time and resources required to develop the required data-sets/models via manual identification and tagging. Because large curated data-sets are essential to ship identification and classification using machine learning, Rite-Solutions proposes an approach using weakly supervised learning to automatically generate labels for non-curated data-sets to train ship recognition and classification ML models. Several tools exist on the market and in academia that have a range of capabilities that, when integrated, will provide Weak Supervision, e.g., an automated way of curating data-sets. Weakly supervised machine learning shows strong potential to accurately perform the label and training functions through automation and thereby reduce SME effort and time to develop the models required for high confidence threat/non-threat vessel identification. Equally important, this approach will rapidly incorporate new images and data to aid the warfighter in identifying and classifying new and changing threats.
Benefit: Development of tools and techniques to overcome the burden of human, manually-curated data-sets and applying the resultant AI models to edge computing environments potentially benefits a broad range of image recognition applications. Benefits include less SME effort required, less overall investment in AI solutions, improved model quality by virtue of using larger labelled data-sets. Markets include surface Navy, air Navy, other DoD, non-DoD, government and commercial enterprises including the oil and gas industry any organization needing to make AI-driven image recognition applications more effective and more quickly deployed than possible today.
Keywords: ship recognition, ship recognition, Automated Labeling, Snorkel DryBell, Weak Supervision, Ship Classification, Snuba, Snorkel, Machine Learning
