Artificial intelligence will not solve everything
Even if great information and machine learning techniques (machine learning) sometimes give surprising results, they do not allow to understand the mechanisms involved. These can only be clarified by fundamental research, which remains very important to ensure scientific progress. It is mentioned in the editorial article published in the magazine Science alarm Holden Thorp, magazine editor Scienceand magazine editor Michael Yaffe Science alarm.
“We are faced with the collection of large scientific data banks and new methods [informatiques et statistiques] Analyzing this big data, it is tempting to believe that major advances in biomedical science will come from translating these troves of information directly into applications in health, agriculture, and climate change strategies rather than discoveries. generated by basic research,” said Thorp and Yaffe. At the same time, they note that major national, international, and private funding agencies are devoting a significant portion of their grants to this approach, which focuses on developing immediate practical applications.
Both are concerned with the future of basic research, which aims to increase our knowledge of the world and discover mechanisms underlying the phenomena we observe, which can then be used to develop biomedical applications.
Long-term scientific losses
They argue that investment in basic science remains important, although many argue that given the incredible potential of intelligence technologies, artificial intelligence (AI) and the wealth of scientific data, scientific research must focus on more pragmatic concerns. All this to meet the monumental challenges facing our species in particular, if not the entire planet. They “warn against research that focuses too much on short-term technological gains that will lead to long-term scientific losses.”
“The greatest advances in science are still the result of proven research methods,” they say, citing viral enzymology and traditional medicinal chemistry, as well as our understanding of the antiviral Paklovide used against COVID-19. immunotherapy for the treatment of cancer developed from new knowledge gained in immunology during basic research.
Using advanced machine learning methods such as deep learning (deep learning), they say, revealed to us how much our fundamental understanding is still lacking in the biological sciences.
Vital basic research
Using deep learning methods, the computing programs AlphaFold and RoseTTAFold can accurately predict the three-dimensional structure of a protein from its amino acid sequence: an amazing feat that humans have never achieved, even though they have known physicochemical principles since the 1950s. , they report.
“This example shows us that there are very fundamental aspects of the protein folding process that we do not yet understand. “Continuing basic research to understand this process is vital if we are to close the gap between our scientific understanding and the prediction of artificial intelligence.”
Another example: machine learning is better than doctors at detecting abnormalities in mammograms, chest X-rays and CT images. “However, what these approaches cannot do enough is explain exactly what the computer is seeing when making a diagnosis or classification. »
And machine learning could not have predicted that a coronavirus, a major research object since the 1960s, would become the pathogen that most threatens humans in the last hundred years. Nor can mRNA vaccines protect us from it. They note that the fruits of basic scientific research have never been more decisive than in this episode.
In which direction to look
“If we have RNA vaccines today, it’s because of the fundamental research done 25 years ago by the Hungarian Katalin Kariko, who tried to understand how mRNA works, when it was not fashionable. […] The results of the main study are unpredictable. Maybe they won’t lead anywhere, maybe the answer will be negative, but it’s important to know because it tells us not to look in that direction again,” adds Yves Gingras, director of the Science and Technology Observatory.
He believes that with deep machine learning algorithms and big data, we will no longer need theories because the data will tell us “a kind of regression towards empiricism”.
“AI techniques are based solely on correlational algorithms that look for relationships between different elements in a data set. And once established, these relationships allow us to predict future events through induction and extrapolation with a certain probability,” says the professor of history and sociology of science at UQAM.
“Prediction does not explain. If we are satisfied with making predictions, we can continue to use the planetary model of Ptolemy’s epicycles [astronome et mathématicien grec du IIe siècle de notre ère], this is wrong, but it works. If the computer is provided with many epicycles, it makes really good predictions about the position of the planets as seen from Earth. But this is just an empirical prediction, even allowing the Babylonians to predict eclipses without understanding how they happen.
“It took Kepler and Newton to really explain what Ptolemy and the Babylonian scribes before him had contented themselves with predicting, after the formation of a physical theory based on the attraction between the planets orbiting the Sun,” he recalled in a published column. in the magazine for science.
False predictions
Mr. Gingras also notes that “forecasts [produites par les méthodes de l’IA] never works 100%. They usually stop at 80%. In 20% of cases they are false. That is, there is uncertainty.”
“Unless we know the mechanism, we cannot have complete confidence,” he reminds. On the other hand, for example, we can trust computers planning trips to the moon today because they apply Newton’s, Kepler’s, and Einstein’s equations, which we know to be reliable. »
published editorial authors Science alarm do not condemn the use of AI, which they see as a technique to speed up more scientific discoveries. “Algorithms are technology, they are useful, but they cannot replace science,” adds Mr. Gingras.
Thorp and Yaffe conclude, “If we rely on a better understanding of biology to guide data analysis in the first place, more discoveries will emerge than if we naively assumed they would otherwise.”
