19 February, 2020
“Vapnik, Guyon and Schölkopf,” the citation continues, “have collectively performed ground-breaking research that transcends traditional boundaries in computer science, and deservedly rank among the world’s leading innovators.”
The awardees have developed methods that endow machines with the essential human skill of recognizing patterns in large volumes of data, enabling them to be sorted into categories. The goal is for the machine to learn through being presented with numerous examples.
Vapnik and Guyon came up with the mathematical tools known as support vector machines (SVM), while Schölkopf extended their range and power through the use of kernel methods, which allow for the input of much more specific categories, thereby multiplying applications. These models are described by the committee as representing “a major machine learning paradigm in both research and applications.”
Thanks to SVM and kernel methods, intelligent machines can now be trained to classify data sets with human precision, or at times even better, enabling them to recognize everything from voices, handwriting or faces to cancer cells or the fraudulent use of credit cards.
SVMs are now being used in genomics, cancer research, neurology, diagnostic imaging, and even in HIV drug cocktail optimization, as well as finding diverse applications in climate research, geophysics and astrophysics.
Schölkopf, Director of the Max Planck Institute for Intelligent Systems in Tübingen (Germany), has recently been employing these methods to analyze data from the NASA satellite Kepler 2. This has helped with the discovery of 21 extrasolar planets, including one with an atmosphere in which signs of water vapor have been detected for the first time.
In the words of one of the nominators, Martin Stratmann, President of the Max Planck Society, machine learning is a core discipline of modern artificial intelligence, consisting of “the study of how to extract patterns or regularities from empirical data.” And Vapnik, Guyon and Schölkopf, he adds in his letter, are “the three scientists who have jointly shaped the field.”
Machines that learn “from examples”
The three laureates coincided at America’s Bell Laboratories in the early 1990s. Yet their backgrounds had little in common. Vapnik, born in 1936 in the former USSR, had worked until 1990 at the Institute of Control Sciences in Moscow, where he and his then pupil Alexey Chervonenkis – who died in 2014 – laid the mathematical foundations for automated pattern recognition methods. Vapnik is now widely recognized, the committee remarks, as “a living legend in machine learning.”
Isabelle Guyon (Paris, France, 1961) was a postdoctoral researcher when she met Vapnik in 1991. Together they two would create the first proven method enabling optimal classification of data, the support vector machine. In 1994 they were joined in their efforts by one of Vapnik’s doctoral students, Bernhard Schölkopf, who expanded the application range of SVMs.
For Vapnik, speaking on the phone after hearing of the award, “the main problem in artificial intelligence is how to get the machine to recognize things, how to distinguish, for instance, between men and women or between different medical diagnoses. You cannot give the machine the rule, the machine has to learn the rule. The support vector machine was developed precisely to fulfil this condition. The aim of all these methods of machine learning is simply so the machine can learn from examples.”
“Machine learning technology is the foundation of almost everything in AI-related business, and will become more and more important in future,” affirms Vapnik, who confesses that he too has been greatly surprised at how much the field has expanded in the past few decades.
The awardee also recalled the initial problem that gave rise to the machine learning field “over fifty years ago,” when he was asked “to solve a small, practical problem that consisted of distinguishing between oil and water, so as to achieve greater precision during prospections based on geological observations.”
In search of cause-effect
Although their paths have diverged, the three laureates have continued to collaborate. Specifically, Guyon and Schölkopf are working on what is seen as one of the key problems in the area: how to identify not just statistical correlations in a forest of data but also relations of causality. Advances in this terrain could tell us, for example, whether a genetic mutation is the cause or consequence of a cancerous process.
The committee says in this respect that “Schölkopf and Guyon have independently and cooperatively advanced the science of causal discovery, uncovering cause-and-effect relationships in observed data, a problem considered by many to be unsolvable.”
In conversation yesterday, Schölkopf explained how his work on causality aided in the discovery of exoplanets: “We had a causal model to distinguish between the signals coming from the star and planet, and the noise produced by the space telescope itself. With this model we managed to filter out the ‘noise’ and determine which signals really originated in outer space.”
A fan of astronomy since childhood, Schölkopf cannot hide his satisfaction at the fact that one of the 21 exoplanets discovered with his aid lies in the habitable zone, and has been found to have water vapor in its atmosphere.
Guyon, meantime, has founded not-for-profit educational initiatives, developed crowd-sourced research and worked on the application of AI techniques to address problems in electricity transmission.
The future of AI
Both Vapnik and Schölkopf are convinced that the AI-led transformation of society is only just starting, and that many of the tasks now performed by humans will be taken over by machines. But this is not to say that machines will actually surpass us in intelligence.
“The machine can already do better than humans at recognizing things, for example, in cases of medical diagnosis or facial recognition,” says Vapnik. “But for me that doesn’t mean that the machine is intelligent. Intelligence is a whole lot more, and we are only just beginning to understand what it is.”
Schölkopf agrees that “we are extremely far away from a machine being more intelligent than a human being.” It is true, he adds, that “in very specific applications, like playing chess or Go, or perhaps in optical recognition scenarios like the diagnosis of skin cancer, machines can do better than humans.” Like Vapnik, however, Schölkopf does not believe that pattern recognition “should really be defined as intelligence,” since what we are talking about is “performance in one very restricted task.”
“What is interesting about our intelligence,” Schölkopf continues, “is that we can play Go then get up from the table and make dinner, which a machine cannot do.” By this general standard of intelligence, the German scientist insists that “machines are still much more stupid than humans.” That said, advances in machine learning are compelling enough, he believes, to cause “legitimate concern that the technology may in future take some kinds of jobs away,” and this is something “we should begin to think about as a society.”