Machine Learning for Transfer Learning for Periscopes

Areté and the Machine Learning for Artificial Intelligence (MLAI) Lab at the University of Arizona (UofA) will develop and demonstrate new approaches that improve the performance of in situ machine learning (ML) algorithms as they evolve over time, adapt to new environments, and are capable of transferring their learned experiences across platforms.  Technological advances that will be brought to bear in this solution include: Meta-Learning, Deep Reinforcement Learning with Visual Attention, Grad-CAM, and Single/Few-Shot Learning. Areté’s solution addresses key problems facing many DoD ML implementations in using commercially developed DNNs –  the lack of sufficient, operationally relevant data to train DNN architectures, the need to quickly adapt to new and changing operational conditions, and the desire to shared “learned” experiences across platforms.  Areté’s approach brings three innovations that tackle these criticisms: A CNN capable of vessel detection, vessel identification, and vessel re-acquisition trained using a model agnostic meta-learning paradigm. An “online” learning capability so that the CNN can learn in situ, adapting to changing situations using a network visualization technique and deep reinforcement learning with visual attention  A method for optimal transfer learning from more experienced networks to less experienced networks. Areté will develop this capability in suite of algorithms to as a software package that implements these innovations for transition to the Navy to improve on the tasks of timely vessel detecting, target classification, target re-acquisition using periscope imagery.