Can an artificial intelligence program reveal the most influential neuroscientists?

Deco / Alamy Stock Photo

A computer program just ranked the most influential brain scientists of the modern era

When it comes to influential neuroscience research, University College London (UCL) has a lot to boast about. That's not the opinion of a human but rather the output of a computer program that has now parsed the content of 2.5 million neuroscience articles, mapped all of the citations between them, and calculated a score of each author's influence on the rest. Three of the top 10 most influential (see table below) neuroscientists hail from UCL: Karl Friston (1st), Raymond Dolan (2nd), and Chris Frith (7th). The secret of their success? "We got into human functional brain imaging very early," Frith says. Getting in early made it possible to "be first to do many of the obvious studies."

The program, called Semantic Scholar, is an online tool built at the Allen Institute for Artificial Intelligence (AI2) in Seattle, Washington. When it debuted in April, it calculated the most influential computer scientists based on 2 million papers from that field. Since then, the AI2 team has expanded the corpus to 10 million papers, 25% of which are from neuroscience. They hope to expand that to all of the biomedical literature next year, over 20 million papers.

When Semantic Scholar looks at a paper published online, what does it actually see? Much more than the typical academic search engine, says Oren Etzioni, CEO of AI2 who has led the project. "We are using machine learning, natural language processing, and [machine] vision to begin to delve into the semantics."

Under the influence

These are the top 10 most influential neuroscientists according to a citation analysis by Semantic Scholar
Rank and Name Affiliation Known for ...
1. Karl J. Friston University College London, U.K. Statistical parametric mapping, voxel-based morphometry, dynamic causal modeling
2. Raymond J. Dolan University College London, U.K. Emotional influences on cognition
3. Marcus E. Raichle Washington University in St. Louis, U.S. Default mode, functional neuroimaging
4. Trevor W. Robbins University of Cambridge, U.K. Neural basis of drug addiction and impulsive-compulsive behavior
5. Terrence J. Sejnowski University of California, San Diego, U.S. Computational neurobiology studies of behavior
6. Alan C. Evans McGill University, Montreal, Canada Brain imaging, modeling of structural brain networks and large-scale brain databasing
7. Chris D. Frith University College London, U.K. Social cognition, cognitive basis of schizophrenia
8. Randy L. Buckner Harvard University, U.S. Tethering hypothesis, neuroinformatics, open data
9. Anders M. Dale University of California, San Diego, U.S. Brain imaging software (Free Surfer)
10. Jonathan D. Cohen Princeton University, U.S. Cognitive control, and its disturbance in psychiatric disorders

DATA: Semantic Scholar/Allen Institute for Artificial Intelligence

For example, consider the Semantic Scholar search results for studies of the basal ganglia in songbirds. On the left side of its results page are things that Semantic Scholar parsed from those papers: not just the typical bibliographic data, such as publication date and author affiliations, but the types of cells used in the experiments and even the methods. The goal is to be able to ask the system questions in English, Etzioni says. "Think 'Siri for science' but hopefully better."

Sam Gershman, a computational neuroscientist at Harvard University, tried out Semantic Scholar this week. “This is a cool tool! ... It clearly has some advantages over Google Scholar," he says, such as "the ability to do finer grained sorting of articles and to show an article's references and some of the figures." But Gershman also quickly discovered the problems that plague all search engines: dirty data. Scientists aren't named the same way on all their papers, for example, and that ambiguity still stumps Semantic Scholar. Also, the paper metadata has mistakes. One of Gershman’s papers is listed as being published in 1987 when he was 2 years old.

But what Gershman found most puzzling was what papers turned out to get the highest influence scores. "Looking at 'most influential publications' sometimes gives strange results," he says. "For example, none of the most influential articles listed for [University of California, Berkeley, psychologist] Thomas Griffiths fall into his top five most cited articles."

Etzioni stresses that Semantic Scholar is a work in progress. "We have errors," he acknowledges.

That didn't stop ScienceInsider from wanting to know the most influential neuroscientists based on the current data. It turns out that the three UCL researchers on the top 10 list have known each other from early in their careers. "We all started together working at the MRC Cyclotron Unit, Hammersmith Hospital [UK] in late 1980s," Dolan says. "We moved together to UCL in 1993 and worked in the same department." Frith points out that Semantic Scholar detects—correctly, he says—that his wife and UCL colleague Uta Frith has been one of the strongest influencers of his career.

Friston, who pioneered techniques for analyzing brain imaging data as well as computational models of how the brain works, took the news that he topped the list with wry humor. "My first thought was, 'Who could I tell without appearing immodest?’" he said. "I then realized the only people who would want to hear about this are my children!"