THAMES: Transformer-based Hybrid Architecture for Management of Hypotheses with Scalability

Under this SBIR Phase I project, we propose a scalable transformer-based hybrid framework for hypothesis management including hypothesis generation, strategy generation, and hypothesis ranking. The framework seamlessly integrates multiple component models to analyze diverse data types for comprehensive situation understanding, employs multi-task weak supervision technique to incorporate pretrained models with domain knowledge for efficient training, and integrates criteria-based ranking algorithms and deep learning models for coherent hypothesis ranking. The system will utilize Novateur’s existing deep learning engine, developed and enhanced under multiple DoD-funded projects that combine the output of multiple deep learning algorithms on diverse data streams to perform hypothesis management. The Phase I effort will focus on developing scalable AI models for hypothesis management, integrating hypothesis management system with a user-friendly interface and continuous learning capability, and testing and evaluating publicly available datasets. The proposed effort will build upon the experience and expertise of the Novateur team in the areas of deep learning, weak supervision, hybrid model design, ranking system development, context-based reasoning, and large language model-based context modeling.