Among the many fields where AI is supposed to have a positive impact, few hold as much promise as science. Indeed, it’s notable that the Nobel prizes for both physics and chemistry went to leading figures from the AI community.
The hope is not new, of course, and in 2019 researchers from Carnegie Mellon showcased how they believed AI could support scientific research. They showcased an AI-driven approach to mine databases of patents and research papers for ideas that can be recombined into solutions for new problems.
Central to the approach was an attempt to find analogies that connect seemingly disparate methods and problems. They used crowdsourcing to understand how people form analogies, then used this to train a deep learning algorithm to mine the intellectual databases for potential innovations.
Making breakthroughs
To date, AI has largely been used in that context, with other startups promising to propose novel ways to develop new drugs or to target different approaches, neither of which have really materialized.
A paper from the Kellogg School examines how AI is currently put to work. The researchers analyzed millions of research papers and found that AI has been deployed in pretty much every scientific area, from biology to geology.
Interestingly, however, while deployment is almost universal, the benefits are not. Indeed, they found that benefits accruing from AI were lower in disciplines where scientists lacked AI skills.
“This is the crucial insight of the paper—the misalignment in terms of supply and demand for AI talents across the disciplines,” the authors explain.
Understanding usage
To assess how artificial intelligence is shaping the sciences, the researchers analyzed a vast dataset of nearly 75 million academic papers across 19 disciplines and 292 fields, published between 1960 and 2019.
First, they defined what “AI” means for scientists today. In the field of computer science, they identified five key areas of AI: machine learning, artificial intelligence, computer vision, natural language processing, and pattern recognition. From these AI-focused papers, they pulled out the most common phrases linked to specific techniques, such as “supervised learning,” “word embedding,” and “generative adversarial network.” They then searched through all the papers across disciplines to see how often these AI phrases, or “n-grams,” appeared.
Their analysis shows that AI use in science has been rising steadily, with a sharp spike starting around 2015. From 2015 to 2019, the use of AI techniques in fields like physics, engineering, geology, and psychology increased by 24% compared to a hypothetical baseline. Other areas, including biology, economics, materials science, and sociology, saw increases of 10% to 30%. The researchers also found that papers mentioning AI were about twice as likely to be highly cited—landing in the top 5% of papers in their field and year.
Moving forward
The researchers then explored how AI could be used in science in the future. They conducted a key-phrase analysis of the scientific literature to pull out various “field tasks”, which are words and phrases that describe what scientists do. They did this both for scientific papers and for AI-related papers to see if there was any overlap. If there was, they deemed this discipline ripe for AI-based disruption.
Perhaps unsurprisingly, they found that the potential for disruption was far from uniform across disciplines, with the field of biological subsystems four times as likely to be affected by AI as other areas of biology.
The researchers argue that the profession is often ill-prepared to make the most of AI, with a lack of skills and training in the use of AI to aid scientific research quite common.
Lack of training
The researchers assessed how well the education system is preparing for the rise of AI by analyzing a database of 4.2 million English-language university syllabi. They looked for references to AI-related papers in course materials, using the frequency of these references as a rough gauge of how much each discipline is investing in AI education.
The results were disappointing. Outside of three computational fields—computer science, mathematics, and engineering—most disciplines were not doing enough to equip graduate students and young scientists with the AI skills needed to fully take advantage of the technology’s potential.
The researchers believe that the promise of AI to advance science won’t be realized until scientists either gain a better understanding of AI itself, or partner with AI researchers. Indeed, this is evidenced by the fact that papers published via collaborations between AI researchers and biologists are growing at a faster pace than by biologists alone.
Giving researchers the opportunity both to develop their AI skills and also collaborate with researchers from AI-related disciplines could be key to getting the most out of the burgeoning technology.
