In August 2018, DARPA released its first AI Exploration (AIE) opportunity called Automating Scientific Knowledge Extraction (ASKE). Unlike DARPA’s typical four-year programs, AIEs are designed to be fast-tracked (~18 months in duration) research efforts that help determine the feasibility of an AI concept. The goal of the ASKE project was to develop AI technologies capable of automating some of the manual processes of scientific knowledge discovery, curation, and application. It identified how and where AI could accelerate the process of scientific modeling, and ultimately improve researchers’ ability to conduct rigorous and timely experimentation and validation.
Scientific models – or conceptual representations of complex systems – are used by myriad communities to understand and explain the world around us. Computationally creating these models is a largely manual, cumbersome task that requires scouring mountains of research for relevant content, and then executing multi-step processes to build, validate, and test the resulting model. The challenges to model creation are compounded by the many opportunities to lose information at each step of the process, or for other errors to occur. ASKE’s goal was to address these challenges by developing approaches to locating new data and scientific resources, comb them for useful information, compare those findings with existing research, and then integrate the relevant data into machine-curated expert models and execute them in robust ways. The project’s research efforts were split across two technical areas. One focused on machine-assisted curation, where researchers explored ways to use AI to extract useful information from research and build it into new models. The second area focused on machine-assisted inference, where AI uses those newly developed models to help researchers understand the modeled system, answer complex questions, or make predictions.
Leveraging streamlined contracting procedures and funding mechanisms, DARPA was able to get researchers on board within three months of the initial opportunity announcement. Things kicked-off quickly, and the ASKE teams began developing a number of novel approaches. Researchers from academic institutions and commercial companies devised ways to automate the extraction of knowledge and information from existing models (including across diverse data types such as written text, equations, and software code), and created technologies to query and link information across literature. They created ways to universally represent and explain different modeling frameworks, while also developing tools that allow computational models to be automatically maintained and/or updated as new discoveries and information becomes available. “The ASKE AIE demonstrated a 50x speedup, extending an existing epidemiology model with additional dimensions and states when compared to state-of-the-art manual processes,” said Joshua Elliott, the DARPA program manager that led ASKE. “Using the same tools, it also showed that a new computational model in a different domain could be created 8x faster with ASKE than with current procedures.”