As part of DARPA’s Environment-driven Conceptual Learning (ECOLE) program, several university teams and industry performers will attempt to create artificial intelligence (AI) agents capable of continually learning from linguistic and visual input, the agency said in a statement.
The goal is to develop AI agents that can collaborate with humans to produce analyses of image, video, and multimedia documents for critical national security tasks. The program’s focus on human-AI teaming aims to produce AI systems with the reliability and robustness required for mission-critical analysis, says DARPA.
The program involves five teams, each contributing complementary techniques:
- Boston Fusion Corp (BFC) will train AI agents to recognize objects and actions in images/videos and study the importance of specific features in object identity through masking.
- GE Research will develop automatic methods for generating object and action curricula to discover their properties and strategies for handling conflicts between user input and learned knowledge structures.
- Systems & Technology Research (STR) will utilize contrastive learning techniques to help AI agents learn by comparing samples against each other and develop a curiosity-driven model for exploring new concepts.
- The University of California San Diego (UCSD) will formulate a graphical representation of object and activity concepts, connect them with their properties, and teach existing models new concepts.
- The University of Illinois Urbana-Champaign (UIUC) will develop an interactive curriculum learning platform to acquire symbolic knowledge representations from unlabeled multimodal data in an unsupervised manner and create a framework for reasoning, prediction, and explanations for uncertain domains.
Instead of relying on handcrafted models, ECOLE will leverage state-of-the-art data modeling to automatically infer object properties and their roles in activities.
[Image courtesy: DARPA]