Sandia National Laboratories’ Agent Bayes project faces a nine-month test that asks whether AI can help scientists perform Bayesian reasoning, a process that often requires a subject matter expert and a statistician working together.
[Image courtesy of DOE]
Performing Bayesian reasoning for science currently “requires both the subject matter expertise and a statistician to help them in developing data analysis methods and approaches that can quantify uncertainties in many high consequence situations,” Tommie Catanach, a lead investigator on the Agent Bayes project, said. “What Agent Bayes is supposed to do is automate a lot of that kind of reasoning and help a subject matter expert reason through decisions.”
When asked how AI could change Bayesian statistics workflows in the long term, Catanach said part of her goal is to make advanced statistic techniques more approachable for scientists that aren’t statisticians.
“Lots of what we’ve already done is trying to automate many of these methods that we have developed… There’s a whole field in something called probabilistic programming, which tries to do this already. However, they still require a lot of expertise to use. So really, this is trying to be a level above that, so that people like me can focus on the things that we really like,” she said.
Catanach and her colleagues designed the Agent Bayes project with the Genesis mission in mind.
“I do a lot of work in Bayesian methods development in general and have been looking to integrate AI… but we haven’t really had a method to directly pursue that and think about it more holistically, which is what the Genesis Mission provides,” she said. “A lot of [the project] was developed based on what the call asked for and also what resources are available through the Genesis platform.”
In addition to resources, the Genesis mission also changes how Catanach thinks about her work, she said. One change is thinking about delivering something that will be a tool other people can use and fit within their own ecosystem, she added.
The team is also thinking about how Agent Bayes will fit within other developments from Genesis mission projects. Catanach said the project is less about deeply theoretical, foundational work on the algorithms and methods and more about researching how AI can fit within existing lab workflows, broadening the impact beyond what they could do with a traditional grant.
The team intends to plug Agent Bayes into the broader ecosystem of the Genesis Mission, which the mission website describes as “an integrated platform that connects the world’s best supercomputers, experimental facilities, AI systems and unique datasets across every major scientific domain to double the productivity and impact of American research and innovation.”
The first Genesis mission cohort, including Agent Bayes, is moving into the first phase of the mission, for which they will receive funding of $500,000 to $750,000, depending on negotiations with the DOE, for nine months of work. Teams can then apply for Phase II, which will provide between $6 million and $15 million.
To advance to the second phase, teams will need to identify how their model provides an advantage over available models, Catanach said.
“For us, that means what the AI we are developing through Agent Bayes can do better than a naive AI implementation of the system or a default implementation of one of the probabilistic programming languages,” she said.
Proving Agent Bayes can be better than existing approaches in nine months is going to be challenging, she admitted.
“I’m really used to working on these three-year timelines… It is very different. We have a bigger team than we normally work with on something like this… and we tried to scope our project within those needs, but it’s definitely a different type of environment.”
Tell Us What You Think!
You must be logged in to post a comment.