Indeed, some AI systems have already solved vexing scientific problems, such as predicting the structure of proteins (a breakthrough recently recognized with a Nobel Prize) and discovering new algorithms for numbers. As technology advances rapidly, there is every reason to believe that its impact will grow over time.
“I would argue that the main question today in AI is whether AI can discover new technologies, which is often considered the most important step in general artificial intelligence,” says Dashun Wang, a professor of management and Kellogg organizations, where he also directs the Center for Science and Innovation (CSSI) and directs the Ryan Center on Crisis.
Given the progress of the past—and the promise of more to come—Dashun Wang and Jian Gao, associate professor at CSSI, wanted to better understand how AI benefits science today, whether how it will benefit science in the future, and whether the education system adequately trains the next generation of scientists to use this new opportunity.
After analyzing tens of millions of research papers, Gao and Wang provide the first mathematical answers to these questions.
They find that, as of 2015, the impact of AI has actually spread to almost every field of science—from biology and chemistry to geology and physics. Many researchers who use AI methods also enjoy a “citation premium,” meaning their papers have a strong impact among their peers.
But Gao and Wang also found that the benefits of AI are not evenly distributed: they are lower in studies with a higher percentage of female and minority researchers.
And perhaps most strongly, with little training, there is a huge gap between scientists who are well trained to use AI in their work and the potential benefit of AI in that teaching.
“This is a key insight of the paper—the inaccuracy in the supply and demand of AI talent across disciplines,” says Wang.
The impact of AI is rising rapidly
To measure the use and benefits of AI across the sciences, Gao and Wang analyzed a large dataset containing the titles and abstracts of nearly 75 million academic papers from 19 disciplines and 292 fields, published between 1960 and 2019.
First, the researchers used the data to create a broad definition of what “AI” means to practicing scientists. Within the discipline of computer science, Gao and Wang identified five areas of AI – machine learning, artificial intelligence, computer vision, natural language processing, and pattern recognition. From this set of AI papers, the researchers extracted frequently used key phrases that correspond to specific AI techniques (such as “supervised learning,” “word input,” and “adversarial networks “). The researchers then searched the entire publication database—that is, all papers published in every discipline and field—to see where and how often AI-related phrases or “n-grams” appeared. how much.
Their analysis showed that while the use of AI in science has been increasing steadily over the past two decades, the “go” like hockey sticks began in many ways around 2015. ( Maybe it wasn’t a coincidence, that was a year. Nature published an influential article on the deep learning of three pioneering AI researchers, and AI algorithms outperformed human-level performance on the ImageNet cluster.)
From 2015 to 2019, direct AI scores using papers in physics, engineering, geology and psychology each increased by 24 percent compared to hypothetical controls. Other subjects, from biology and economics to materials science and sociology, also saw increases from ten to thirty percent. The researchers also found that papers discussing AI n-grams were almost twice as likely to be “hit” within their field (defined as the top 5 percent of the number of papers within the field and year).
AI and the technologies of tomorrow
Next, the researchers wanted to estimate the potential benefits of AI for scientific studies going forward.
To do so, they conducted a cross-sectional analysis of keywords in scientific papers to extract “fieldwork” – pairs of verbs and nouns that describe what scientists in each discipline do. For biologists, one such task might be “identifying genes”; for chemists, it can be “catalyze reaction”. Gao and Wang also collected similar verb pairs from AI-related papers (as well as AI patents selected from the 7.1 million patents issued by the United States Patent and Trademark Office of Business between 1976 and 2019) to create “AI capabilities.” Finally, they compared two sets of subject phrases (ie, verb-noun pairs) and searched for overlap. If the “ability of AI” also appeared as “field work” in a particular discipline, that discipline was seen as an opportunity to benefit from AI in the future.
Using this measure, Gao and Wang found that AI has the potential to benefit every science discipline.
They found significant differences between sub-regions within a particular discipline. For example, jobs in the systems biology subfield (which seeks to model complex interactions within organisms) were four times more likely to be affected by AI compared to jobs in in other areas of biology, such as horticulture or food science. But overall, says Gao, “AI has widespread influence and scientific benefits in all disciplines.”
An uneven journey
That said, Gao and Wang also found that disciplines with a high proportion of women and underrepresented minorities were least likely to benefit from AI—both in terms of the direct use of AI today and the benefits. possible AI of tomorrow. For example, in sociology—where nearly half of researchers with PhDs are women, and 16 percent identify as an underrepresented ethnic group—the current benefits of AI for the discipline it’s almost half physics, with a very high stakes. of male, Caucasian, and Asian researchers. Career-level surveys also revealed that non-professional scientists who engage in AI-related research see a smaller increase in their “hit” rate (as measured by citations) than non-professionals. other scientists.
Gao says: “We have known for a long time that technological change often causes labor inequality. “If we predict that AI will continue to benefit scientific research in the future, we should be concerned about how those benefits are distributed.”
Training gap
But the biggest finding of all may be how ill-prepared many disciplines are to take advantage of AI advances. After all, AI can only benefit a discipline if its scientists have the experience and training to use AI properly.
The researchers assessed the education system’s readiness for AI by scanning a database of 4.2 million English language university syllabi for scientific references to AI-related papers. “We wanted to know how ready the next generation of scientists are to use AI advances,” Gao explains. By measuring the frequency of AI references in the discipline, we can estimate [that discipline’s] level of investment in AI education. ”
Here the results can only be described as disappointing. With the exception of three computer courses (computer science, mathematics and engineering), universities were not investing enough in teaching AI-related skills to graduate students and junior scientists to achieve the full benefits in from AI.
For Wang, this finding should be a clear appeal to policy makers around the world: “What kinds of science policy can solve this problem?”
Indeed, their research points to one answer. When Gao and Wang examined collaboration patterns in disciplines other than computer science (e.g., biology), they found that the number of AI publications produced through collaborations between scientists computer and, say, biologists were growing faster than what was produced by biologists alone.
In other words, scientists from different disciplines are finding it useful to rely on their peers who have specialized knowledge in AI. This suggests that fully utilizing AI in science may require not only more funding to train scientists but also more opportunities for interdisciplinary collaboration.
To some extent, this is already happening. “Some organizations are starting interdisciplinary research centers, which encourage scientists from different disciplines to have discussions about how to use different AI tools and developments,” says Gao. “That would give researchers an opportunity to learn from each other, while doing research together.”
But to maximize AI’s potential, say Gao and Wang, collaboration will need to happen on a much larger scale. To that end, Gao and Wang’s findings are included in a larger report presented to the National Academy of Sciences, which advises the US government on science-related policy.
Regarding the benefits of AI in his research field, Gao is optimistic. “I’m excited about how AI can help automate difficult tasks and improve our creativity,” he says. “It would free us up to have more time to ask new questions, explore difficult areas, and push the boundaries of knowledge.”
#Changing #Science #Scientists #Ready