Corelated Data Generation (CADET)

The U.S. Army has a need to generate multi modal synthetic data to train and validate various AI/ML tools. AI/ML tools like Object Detection and Automatic Target Recognition have rivaled or surpassed human level performance. However, development and validation of such AI tools requires large amount of data. Before these tools can be deployed on the field, they need to be validated on variety of mission scenario data. Collecting such data can be cost prohibitive, hence there is a need to synthetically generate training and validation data. In response to this Army requirement, Knowledge Based Systems, Inc (KBSI) proposes to design, develop and demonstrate Corelated Data Generation (CADET) to generate multi-modal corelated synthetic data to train and evaluate Army AI/ML capabilities. CADET will utilize recent advances in Generative AI and Reinforcement Learning to generate Multi-Modal Synthetic Data. The innovations in the CADETĀproposed solution are: Using Auxiliary Learning to fine tune multiple data generators in a single step. Applying Reinforcement Learning principles to quickly finetune data generators to novel scenarios with minimal human input. Providing intuitive metrics and visualizations to summarize performance of models relative to human labelling efforts.