Can we trust scientific discoveries made using machine learning?

Rice University statistician Genevera Allen said scientists should continue to question the accuracy and reproducibility of scientific discoveries made using machine learning techniques until researchers develop new computer systems that can criticize themselves.

Allen, an badociate professor of statistics, computer science and electrical and computer engineering at Rice and pediatric-neurology at Baylor College of Medicine, will address the topic at a press conference and one-on-one session. General held today at the annual meeting of the American Association in 2019. for the Advancement of Science (AAAS).

"The question is:" Can we really trust the discoveries that are currently being made using machine-learning techniques applied to large datasets? Allen said. "The answer in many situations is probably" not without verification, "but work is underway on next-generation machine learning systems that will evaluate the uncertainty and reproducibility of their predictions. "

Machine Learning (ML) is a branch of statistics and computer science in charge of setting up computer systems that learn from data instead of following explicit instructions. According to Allen, a lot of attention in the area of ​​BC has been focused on developing predictive models allowing ML to predict future data based on his understanding of the data he has studied.

"Many of these techniques are designed to always make a prediction," she said. "They never come back with" I do not know "or" I did not discover anything, "because they're not made for."

She said that the uncorroborated, evidence-based findings from recently published cancer studies on ML are a good example.

"In precision medicine, it's important to find groups of patients with similar genomic profiles to be able to develop drug therapies that target the specific genome of their disease," Allen said. "People have applied machine learning to genomic data from clinical cohorts to find groups or groups of patients with similar genomic profiles.

"But there are cases where the discoveries are not reproducible, the clusters found in one study are completely different from those found in another," she said. "Why? Because most of the machine learning techniques of today always say:" I found a group. "Sometimes it would be much more helpful if they said, "I think some of them are really grouped together, but I'm not sure about these others". "

Ms. Allen will discuss today the uncertainty and reproducibility of ML techniques for data-based discoveries at a press conference at 10 am. She will also discuss case studies and research to resolve uncertainty and reproducibility at 3:30 pm. General Session, "Machine Learning and Statistics: Applications in Genomics and Computer Vision". Both sessions take place at the Marriott Wardman Park Hotel.

Provided by
Rice University