UCI’s Department of Computer Science welcomed Dr. Yolanda Gil, the Director of New Initiatives in Artificial Intelligence (AI) and Data Science at USC’s Viterbi School of Engineering, for a computer science seminar on artificial intelligence on Nov. 12. Gil discussed the future role that artificial intelligence systems may have as authors of scientific papers to produce their own experiments and shape interactions between scientists and publications.
Gil opened the discussion by introducing the topic of artificial intelligence and the ways in which it has been utilized throughout the years.
“AI has had many important contributions in intelligence, modeling, capturing and representing knowledge itself,” Gil said. “In the area of knowledge representation and ontology for instance, AI has been able to connect and integrate so much knowledge in biology through the genome.”
According to Gil, the beneficial use of AI has been favored over a scientist’s ability to conduct scientific work. Inefficiencies in approaching problems from a human perspective allow for errors and biases to arise and cause poor reporting of scientific concepts.
“We think of scientists as people who do very special thinking. As humans however, I see that we suffer sometimes in this area. Humans aren’t very systematic, since we aren’t thorough when doing a task,” Gil said. “We also make mistakes and errors. This causes papers to be written in ways that others can’t build on from. The way we do science is essentially unideal. AI, in this way, can help us become better scientists.”
In order for AI systems to eventually become scientists themselves, generating papers should be a task they can do consistently from a reference or on their own. Highlighting some of the ways scientists carry out research, collect data and write scientific papers. Gil explained these concepts to discuss the possibility of extensive accurate reporting by AIs.
“Particular softwares can be used to carry out experiments. Once data is obtained, results are sorted and combined to eliminate anything that is insignificant. In this way, a report can be written. Highlighting good principles to archive and share data and workflows with AI systems are all significant parts of producing a paper, which humans are not generating,” Gil said.
Gil then highlighted the ways in which AI systems have begun to shape and develop scientific papers. The first example she described was science vocabulary and the ways it can be utilized to describe scientific data in the area of paleoscience, the study of climatic and environmental processes.
“When attempting to describe trends in climate over thousands of years, you want to pull all data together to report the findings. To unify their data, paleo scientists look to speed up the process with AI. We proposed to them a new approach where we could crowdsource the way each community of scientists wanted to describe their data,” Gil said.
In this procedure, one face to face meeting was held, where all professionals agreed to use the system to help write the report. A few months later, a common way to describe their data was developed. Gil explained this system.
“A systematic wiki was developed, where these scientists could adopt terms that existed. They could also browse and add their own terms,” Gil said. “A lot of standard ways to describe factors like location for example, arised as well. This AI approach had to adapt to the changes in terms along the way.”
Another example Gil described was workflow, a multi-step data analysis process. According to her, scientists can apply constraints to workflow to enhance data collection and scientific writing.
“In regular scientific papers, we are able to describe the general method used versus its execution in a study. We use rules and schematics in technologies and languages that place constraints on the workflow, which allow for users to be consistent with the data they present,” Gil said.
For example, time series analysis, in which an analysis of a sequence of data points over a period of time is discussed, can use workflow with AI systems.
“We represent all of the steps as an abstract template that provides specific steps for the constraints that come with them with AI. These developed algorithms for the constraints are good instruments to elicit the knowledge in order to build these workflows, which can be applied in many contexts,” Gil said.
To conclude, Gil explained how tackling complex scientific phenomena is something we will have to do in the future. According to Gil, scientists and AI systems will have to learn to become partners of discovery in the scientific community.
“AI systems can’t just be around being told what to do by scientists. I think AI systems need to learn on their own, grow their knowledge, notice new software and connect with people or other AI systems. Reproducing and writing articles is a very modest goal,” Gil said.
More information about Dr. Yolanda Gil’s team and their research in artificial intelligence and its innovation in scientific discovery is available on the “Knowledge Capture and Discovery” website.
Korintia Espinoza is a STEM Intern for the fall 2021 quarter. She can be reached at korintie@uci.edu.
