PI: Susannah Paletz; CO-PI: Adam Porter
Co-Investigators: Aimée Kane (Duquesne University) and Madeline Diep (Fraunhofer USA Mid-Atlantic)
Student investigators: Sarah Vahlkamp (iSchool), Aayushi Roy (CS)
Human-Agent Teaming on Intelligence Tasks
Description:
Complex sensemaking tasks are a fundamental part of intelligence analysis. Analysts routinely sift through large numbers of source documents, identify key facts and trends, assemble them into coherent models of reality, and use those models to inform action. While some operations can be effectively handled by a single analyst, others require a team effort, which introduces additional cognitive challenges associated with sharing information and aligning mental models across the analysis team. Artificial intelligence (AI) and machine learning are increasingly discussed as essential technologies to solve analysis challenges (Gartin, 2019). For this project, we are creating an experimental infrastructure and task in order to test different types of AI supports for human analysts in the context of a shift handover.
PI: Susan G. Campbell
Student investigators: Victoria Chang and Melissa Carraway (both iSchool Ph.D. students)
Description:
The goal of this project is to advance the science of human machine teaming by developing and testing techniques for communicating the state of machine agents to warfighters in an actionable way and for communicating the relevant information from a flood of battlefield sensors in a comprehensible way. It is divided into two streams: stream A, which concerns understanding and guiding agents, and stream B, which concerns aggregating and distilling sensor information from agents. Year 1 of the project focuses on identifying and documenting use cases that will guide both streams of research, while later years will build infrastructure and prototypes to enable experimentation.
PI: Susan G. Campbell
Co-investigators: Breana Carter-Browne (iSchool affiliate) and Elizabeth Bonsignore
Description:
Future soldiers will need to be able to use AI-enabled systems in most situations, but people vary in their abilities to use technology now, despite widespread use of consumer technology systems. The long-term goal of this project is to improve the measurement of human ability to use and adapt to advanced technologies like machine-learning enabled assistants, autonomous agents, and multi-agent systems. To assess this fluency with AI systems, we plan to collaborate with ARL and the Army Research Institute for the Behavioral and Social Sciences (ARI) to define the capabilities and cognitive skills that comprise AI Technological Fluency (ATF). Once we have targets for measurement, we plan to build testbed systems that will allow us to collect indicators of those capabilities and skills.
PI: Sharon Glazer
Co-PIs = Stephan Gerschewski, Vas Taras, Sharon Glazer, Longzhu Dong
Description:
PI: Joel Chan
Description:
Synthesis — the construction of novel conceptual wholes, such as theories and new questions from previous knowledge components — is the lifeblood of scientific progress. This critical conceptual task requires deep engagement with past work at the level of theories, lines of evidence, and claims, and how they inform, support, or oppose each other. Unfortunately, infrastructures and tools for scientific literature do not support this unit of analysis, privileging instead the coarse unit of the scientific paper. This limitation is a serious impediment to scientific progress. A key building block for a new infrastructure for synthesis already exists: a discourse graph data model centered on richly contextualized networks of questions, claims, and evidence. A critical remaining challenge is how to incentivize the right people to contribute this data at scale: automated systems cannot yet replace human judgment for synthesizing knowledge claims and evidence from the literature. To meet this challenge, this project explores novel interactive tools that seamlessly integrate the work of authoring and sharing discourse graphs into scientists’ natural workflows of reading and synthesizing literature. These data can then be shared, federated, and aggregated into new infrastructures for synthesis that can accelerate scientific discovery, both within and across disciplines
PI: Joel Chan
Description:
Analogy — the ability to find and apply patterns from other domains — is fundamental to innovation. Observing water led the Greek philosopher Chrysippus to speculate that sound was a wave phenomenon; an analogy to a bicycle allowed the Wright brothers to design a steerable aircraft. Today, the opportunities for finding analogies are exploding with the increased availability of online repositories of ideas ranging from scientific papers (Google Scholar) to product ideas (U.S. patent database) to the entire web. Yet, the cognitive effort and time required to fully mine the scale and diversity of these online data far exceed humans’ limited cognitive and temporal resources. Conversely, computational systems can scale to large amounts of data but are limited to a shallow understanding of structure and semantics, which poses a serious challenge to finding analogies that go beyond surface features and word-based matches.
This research will develop interactive search engines that enable scientists and inventors to discover and adapt analogies from very large unstructured text datasets. The intuition behind our approach is that rather than trying to solve the problem of fully structured analogical reasoning, we instead explore the idea that for retrieving practically useful analogies, we can use weaker structural representations that can be learned and reasoned with at scale (in other words, there is a tradeoff between the ease of extraction of a structure and its expressivity). Specifically, we investigate the weaker structural representation of an idea’s purpose and mechanism as a way to find useful analogies. We also explore extensions of this representation to deal with issues of hierarchies of purposes and mechanisms, and levels of abstraction that are present in real-world documents like research papers and R&D documents, as well as interaction techniques for reducing the cognitive overhead of transferring insights across domains. Using these representations and interaction techniques, we will build computational tools enabling users to connect problems in one field with solutions from another field based on their deep structure. The proposed research aims to bridge the gap between the power of large-scale text mining approaches, which excel at detecting surface similarity, and the depth of human cognition which is currently unsurpassed at detecting deep analogical similarity. Our results and algorithms could spur the development of new types of machine-learning techniques that focus on deep structure, and may contribute back to theory in the fields of creativity, problem solving, and innovation.
The development of analogical search engines that retrieve content based on deep structure rather than surface features could also have transformative practical impacts in a variety of fields. Instead of relying on the small number of incidental analogies found by people today, computational tools that scale up analogy finding to large data could dramatically accelerate innovation and discovery. Augmenting human ability to rapidly integrate knowledge across a wide range of data sources and knowledge domains may also have significant benefits for the Navy and any organization that needs to address complex problems without ready-made solutions, especially if the problems require integrating many diverse perspectives. To increase impact, we will transfer our research results to partner organizations including Web of Science, as well as to local research groups, companies and startups.
Organizational Teams and Technology Society (OTTRS)
University of Maryland
Patuxent Building
4161 Fieldhouse Dr
College Park, MD 20742