Towards Reflection Competencies for AI Scientists: Developing a Conceptual Framework and Open Research Platform
Abstract
Approved for Public ReleaseAs scientific questions become more complex, the capabilities of scientists to do research will need to be augmented with intelligent aids. Compare the challenges of finding a cure for polio with finding a cure for cancer. Polio, a scourge that has affected humanity for millennia, was cured through a vaccine that was discovered by one scientist. Glioblastoma, a brain cancer that takes very few months to go to advanced stages and is very hard to detect and to treat, is being studied scores of scientists with different specialties. Different research groups have complementary information about the disease, with disparate data on genomics, proteomics, transcriptomics, MRIs, clinical, therapies and drugs, etc. Each discipline advances very quickly, making ithard for each research group to be up to date with new methods and findings. It is also challenging for each research group to keepup with the latest and best methods.Artificial intelligence (AI) has a longtradition of tackling scientific research as a domain ofstudy. The DENDRAL project at Stanford pioneered this field, where AI was used toautomate the analysis of mass spec and other data. Many aspects of scientific discoveryhave been addressed by artificial intelligence research across many areas of science. Inturn, science challenges most often uncover research challenges for artificial intelligence.Our ultimate goal is to develop AI systems able to carry out scientific research. These AIscientists will be capable of pursuing independently substantial aspects of the research and therefore make their own discoveries. They should be capable of taking on significantproblems by formulating their own research goals, proposing and testing hypotheses,designing theories, debating alternative options, and synthesizing new knowledge. These AI scientists should also be able to explain their reasoning, compare their lines of inference to other possible ones, and situate their findings. They should also be able to communicate with scientists who have different levels of expertise and understanding in any given research topic.The required capabilities will only be possible through substantial research advances in adiversity of areas of AI, including cognitive systems, machine learning, knowledgerepresentation, constraint reasoning, problem solving and planning, meta-reasoning,reasoning under uncertainty, multi-agent systems, natural language processing,collaboration, and robotics. It will also emphasize intelligent capabilities that have beenreceived less attention in the past, such as representational change, creativity, andknowledge discovery.The goal of the proposed work is to prototype core infrastructure to jumpstart an effort todevelop AI scientists. The proposed prototype, named DISK, will focus on the development of core technologies for AI scientists to do hypothesis-driven research. DISK will build on and extend an existing system developed under NIH, DARPA, and NSF funding, that uses AI problem solving and knowledge representation to test hypotheses in cancer omics and in neuroscience. DISK will provide core capabilities, and will have a modular extensible architecture with well-defined APIs that will enable other researchers to plug in components to improve itsperformance.DISK will enable AI advances that have the potential to accelerate all areas of science. Inaddition, they will benefit many other areas of societal interest, since AI advances will beapplicable in education, health, innovation, and security.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112437
Entities
People
- Yolanda Gil
Organizations
- Office of Naval Research
- United States Navy
- University of Southern California