The Defense Advanced Research Projects Agency (DARPA) has embarked on an ambitious project through its Environment-driven Conceptual Learning (ECOLE) program. Several academic and industry teams will try to develop artificial intelligence (AI) agents that can learn continuously from both visual and linguistic input. These AI agents aim to assist humans in time-critical analytical tasks, such as analyzing images, videos, and multimedia documents.
ECOLE: The Next Step in Human-AI Collaboration
The ECOLE program aims to create a trusted and capable AI system to assist in mission-critical analysis. The initiative is set to address one of the three vital areas defined by DARPA experts for the development of such systems: human-AI teaming.
“Conceptual knowledge acquisition is crucial for reliable AI automation in the future,” says Dr. Wil Corvey, ECOLE program manager at DARPA’s Information Innovation Office. This approach calls for innovative ways to comprehend the fundamental properties of objects and activities as an agent observes them.
Rethinking Machine Learning Methods
The focus of all participating teams is to revolutionize machine learning methods through curriculum learning, collaboration, and human-machine analytical cooperation. Unlike previous efforts that depended on preconceived models, ECOLE plans to utilize state-of-the-art data modeling to infer properties of objects and activities automatically.
The five participating teams will implement distinct yet complementary techniques:
Boston Fusion Corp (BFC)
BFC will train AI agents to identify objects and actions in images/videos, emphasizing understanding the features of these objects and actions. The team will also explore the importance of specific attributes in learning through a method known as masking.
GE Research
GE Research intends to create automatic methods to build object and action curricula for discovering their properties. The team will also devise strategies for resolving discrepancies between user input and the AI’s learned knowledge structures.
Systems & Technology Research (STR) LLC
STR will utilize contrastive learning techniques, enabling AI agents to learn by comparing samples. This approach allows AI to distinguish common attributes and differentiate between data classes. The team will also develop a curiosity-driven model for exploring stored knowledge and investigating unknown concepts.
University of California San Diego (UCSD)
UCSD will create a graphical representation of object and activity concepts. They will then teach an existing model new concepts by examining if new properties have been added to the graph, considering the quality of the acquired concepts through automatic metrics and human feedback.
University of Illinois Urbana-Champaign (UIUC)
UIUC will create an interactive curriculum learning platform that acquires symbolic knowledge representations from unlabeled multimodal data in an unsupervised way. They will also construct a framework to learn and combine attributes and concepts for reasoning, prediction, and explanations in uncertain domains.
Related Articles
