Machine learning, and the way ahead

At the very peak of the most recent 2016, Gartner Hype Cycle resides Machine Learning, which subsumes what used to be categorised as Big Data, Data Science, Artificial Intelligence, Data Mining, or Predictive Analytics. I would even include Statistics and Nonlinear Dynamics, two more traditional fields which continue to remain important, within the broad definition. While there may be slight technical differences, in common parlance they mean the same thing: the ability to use computers (i.e., "machines") and sophisticated mathematics to extract actionable predictive insights from data (i.e., "learn"). In our day and age, almost all mathematicians, statisticians, and physicists use computers, while all computer scientists who work with data usually learn the basics of traditional data science disciplines such as statistics and applied mathematics, nonlinear dynamics and network science in physics, signal processing in engineering, and econometrics in economics. Thus, any difference between the Machine Learning and what may have been once called traditional data sciences has been fast disappearing.

Machine Learning experts get the high-paying jobs, and companies from Google, Facebook, and Amazon, to Oracle, IBM and Microsoft, increasingly identify themselves with this technology. A wide swath of companies claims to benefit, such as those in fast food, retail, packaged goods, and finance, to high tech, heavy equipment and construction. As Google's CEO Sundar Pichai wrote in his 2016 vision letter to shareholders and employees: "A key driver behind [the success of Google] has been our long-term investment in machine learning and AI [Artificial Intelligence]." He went on to say that "the implications … are, literally, game-changing—and the ultimate winner is humanity. This is another important step toward creating artificial intelligence that can help us in everything from accomplishing our daily tasks and travels, to eventually tackling even bigger challenges like climate change and cancer diagnosis."

Reading the above, an ornery reviewer may argue that it is quite a journey from "daily tasks and travels" to "climate change and cancer diagnosis." However, let us not spoil the mood. At least, not yet. Let us instead ask a simple question: If you had the power of Machine Learning with you, would you rather solve problems in business such as online advertising, recommendation systems for movies or friend networks, and stock picks for financial hedge funds, or would you rather address challenges in climate and the environment, hazards and security, infrastructures and the supply chain, and/or medicine and public health, which represent national and societal priorities, and may even relate to fundamental questions in science and in engineering principles? The cover of the 7-7-2017 issue of the reputed Science magazine boldly states: "AI transforms science". Would you contribute to this transformation?

This is a question I often ask my graduate and undergraduate students and research mentees in another form: Would you rather make money or save lives? I would not judge: My own interests and passions have shifted too wildly, and swung too far into extremes over time, for me to judge rashly. However, I would be curious to know. In some cases, I would ask a variant of the same question that Steve Jobs is supposed to have asked Pepsi executive John Sculley, apparently to attract him to Apple: "Do you want to sell sugar water for the rest of your life, or do you want to come with me and change the world?"

Back to the ornery reviewer. Is it an easy leap from solving business problems to addressing societal priorities and answering fundamental science or engineering questions? One of the largest, oldest and most prestigious computer science venues for Machine Learning or Data Mining is the annual KDD (Knowledge Discovery and Data Mining) conference. A few years back, a set of related questions were posed to a KDD expert panel. There were arguments that the size and nature of data available in recent years and the Machine Learning tools we have at our disposal render the old statistical concepts of signal (the part that is predictable) versus noise (the random and non-repeatable variations in the data) null and void. However, questions were raised as to why many so-called groundbreaking discoveries cease to last even a few years. The issue of false discovery was discussed. In online advertising, or friend suggestions, or movie recommendations, such falsity has limited costs and is relatively easily remedied once known to be erroneous. However, the costs may be more deeply felt in finance or in, for example, self-driving cars. In cases like climate change and the environment, or health and medicine, or resilience to natural and man-made hazards, misleading or wrong decisions can entail long terms costs to human lives and well-being. Furthermore, in situations where a baseline process, or the physical, or biological, or engineering, the principle is even partially known, wrong decisions guided by false data-driven discoveries can incur significant avoidable costs, which would be difficult to justify even on hindsight. Nevertheless, Machine Learning is being used in climate and cancer research already. One of our own work, joint with NASA and led by my PhD student Thomas Vandal, develops innovative Machine Learning (or, more specifically, Deep Learning, a special class of Machine Learning) methods for a difficult climate problem. This will be presented at the 2017 KDD conference to be held in August in Halifax, Nova Scotia, Canada.

I teach a graduate course called Applied Time Series and Spatial Statistics, where I cover aspects of what would fall under Machine Learning (broadly construed). When I asked a class of about 40 students this question (would they rather save lives or make money) about half of the students in basic sciences and engineering, and some in policy, answered one way, unlike the rest in applied engineering and operations research fields. The question is perhaps a bit unfair. The choice is usually not that stark. Plus, you can always do both. This is not Star Wars, and you do not have to be either on the Dark Side or on the Light side. You can do a lot of good by wealth creation and eventually use some of your wealth for societal good, and a lot of harm by misguiding climate and environmental policy or by developing false remedies in medicine, even unintentionally. However, once you learn Machine Learning and/or understand how to benefit from the technology, you are undoubtedly empowered by a special Force!