Teams selected to teach AI Agents to interact with people and learn
Use cases, such as mission-critical analysis, require AI with suitable trustworthiness and competence. ECOLE will tackle one of the three areas DARPA experts say is vital to producing such systems - human-AI teaming.
"The ability to acquire conceptual knowledge is core to the future of reliable AI automation," said Dr. Wil Corvey, ECOLE program manager in DARPA's Information Innovation Office. "Learning new concepts requires new methods of introspecting about the essential properties of objects and activities as an agent observes them in its environment."
All participating teams will focus on transforming machine learning approaches through curriculum learning, sharing amongst multiple participants, and human-machine collaborative analysis. Rather than relying on handcrafted models as previous efforts have explored, ECOLE will use state-of-the-art data modeling to automatically infer the properties of objects and their role in activities. Research and development from each of the five teams will use complementary techniques, including the following:
Boston Fusion Corp (BFC) will train AI agents to recognize objects and actions in images/videos and identify the features of these objects and actions. The team will further study the role of features in learning through a technique called masking, where researchers remove certain properties (e.g., the yellow of a banana or the treads on a tank) to gauge how important those are to the identity of the object.
GE Research will develop automatic methods for generating object and action curricula for discovering their properties. GE will also develop strategies for handling conflicts between user input and learned knowledge structures.
Systems and Technology Research (STR) LLC will use contrastive learning techniques to help AI agents automatically learn by comparing samples against each other. This approach helps AI to learn attributes that are common between data classes and attributes that set apart one data class from another. STR will also develop a curiosity-driven model to explore stored knowledge and investigate new, unknown concepts.
The University of California San Diego (UCSD) will formulate a graphical representation of object and activity concepts and learn by connecting these concepts with their properties. UCSD will also teach an existing model new concepts by analyzing whether new properties have been introduced to the graph while considering the acquired concepts' quality with automatic metrics and human feedback.
The University of Illinois Urbana-Champaign (UIUC) will develop an interactive curriculum learning platform acquiring symbolic knowledge representations from unconstrained unlabeled multimodal data in a completely unsupervised way. They will also develop a framework to learn and compose the attributes and concepts for reasoning, prediction, and explanations for uncertain domains.
Additional details of the program's motivation, schedule, and metrics are available here