Machine Learning Pioneer Questions A.I. and Forges New Engineering Path
In the century since microeconomics and statistical machine learning came to the fore divides between the two have grown in parallel to their importance for Artificial Intelligence. Hi! PARIS invited University of California, Berkeley professor Michael I. Jordan to share his foundational work on the next generation of machine learning algorithms addressing challenges like privacy and the corruption of the learning process. In late October, Jordan led a masterclass in machine intelligence, statistics and AI in front of students and academics from HEC Paris and Institut Polytechnique de Paris (IP). In a wide-ranging interview before the conference, the American scientist exchanged with HEC on his background, ongoing research and his fundamental questioning of ethics and AI...
Michael I. Jordan (the “I.” is one of several factors distinguishing him from his basketball namesake) has been described as many things. The University of California, Berkeley, alone lists 25 biographical highlights for him, ranging from MIT professor for a decade to Member of both the National Academies of Sciences and of Engineering. Which might go some to explaining why he was named the “most influential computer scientist” worldwide in a Science magazine article. An attribute which, ironically, rankles the research academic despite being awarded the first-ever World Laureates Association Prize last year for his “fundamental contributions to the foundations of machine learning and its application”. “I’m an academic who’s dabbled in every discipline for at least a little while,” he says matter-of-factly. “I began in philosophy which led me to cognitive science that led to neuroscience.” Indeed, Jordan’s interest in how the brain might work in order to better understand humans dominated his research for the 10 years he was in the Department of Brain and Cognitive Sciences at MIT. But he got restless: “I started to realize that it was slow progress in important but slightly static fields. And, in the meantime, data, intelligent algorithms and adaptive systems began to become more prominent in my work. I moved into those dimensions as I realized that I was really interested in statistical inference and induction. You know, basically trying to take partial observations of the world and figure out what’s going on out there.”
The 67-year-old has a wonderful way of simplifying his dense trajectory. And, in the course of his visit at the invitation of Hi!PARIS, he illustrates it candidly: “Specialists try to decide if a drug works or not, or make a prediction about the weather tomorrow. But these are questions you can only make inferences about based on gathered evidence. And I realized that what humans are particularly good at is logical thinking or chess playing. And we’re especially good at dealing with uncertainty and the vast complexity of the world, including the social world.” So, Jordan moved on to how humans deal with this complexity, which led him to machine learning. “This meant understanding uncertainty with algorithms that would manipulate uncertainty. And that brought me to control theory and other areas of mathematics and, most recently, to economics.”
What’s Under the Hood
In the domain of economy, the researcher started talking to multiple decision-makers who cope with uncertainty collectively, leading to different applications. “I’ve spent a lot of time in industry as well as academia,” he continues. “And I’ve seen the perfect storm of computation, algorithms, and data, which has changed the world. Companies that build very big supply chains, like Alibaba and Amazon provide services on a scale that was unimagined before in history. They built markets, billions of products sent to hundreds of millions of people. The supply chains behind that process are amazingly complex. You have to build models which can integrate a strike in Singapore, a storm in the Indian Ocean, a war in Europe, all these pieces that need to come together. And this model on a vast scale has to look seamless so that people hardly pay attention to it.”
Jordan began noticing that the same companies were building recommendation systems, interacting with people based on vast amounts of data, analyzing past behavior and purchases. “I found some very interesting patterns which are having a big effect on the world and its economy. It’s your billion-dollar production gains from supply chain modeling. Something was really changing under the hood. And now that we’ve had the latest wave of machine learning with speech, visual and language data, we start seeing that it’s not so much under the hood anymore. It’s up front, doing things that seem human, but not quite human. There’s been a lot of progress these past 20-30 years, but it’s part of a 50-year trend, with networks coming together with data and statistics.”
The 50 Year Cycles Dominating Engineering
This cycle mirrors one Michael I. Jordan also sees in the engineering field. Looking back over the past century, the researcher (who is a member of the Hi! Paris scientific Advisory Board), has seen important developments which reverberate into the present times. “50 years ago, chemical engineering emerged. Before that, there was the coming together of chemistry, quantum mechanics and fluid flow. We were able to understand what happens when you put molecules together at a huge scale and create objects like polymers, plastics, or chlorine to clean our pools. And the field emerged called chemical engineering, to develop principles and think of chemistry at scale. But new thermodynamics were needed, what’s now called medical engineering. Similarly, 50 years before that you saw the emergence of electrical engineering with its own mathematics. It allowed people to build systems with modules. You can go back even further, another 50 years, to see the emergence of mechanical engineering, and there was civil engineering, and so on.”
So what is emerging at present? “There’s a new engineering field developing. It’s based on the principles of markets and interactions and incentives. And, economics, the principles of markets, interactions, and incentives.” What Jordan finds most stimulating is that this new dynamic goes beyond the science and maths behind these developments: “It’s the actual building of systems that exhibit all these properties. And it’s worth being excited about healthcare that’s being transformed by algorithms, finance, and commerce. It’s key to realize that we’re building actual artifacts that are designed to make human life better and safer. There’s an exciting interaction between new mathematics and real-world problems that are provoking the changes. I don’t think we’ve found a term for it yet, but it is a coherent new engineering field which sees companies plunging in, new firms building ChatGPT and large language models, designing infrastructure for commerce, finance or healthcare.”
An Interdisciplinary Approach
It's hard not to get infected by Michael I. Jordan’s unquenchable thirst for pushing back the boundaries. It was perceptible in the Masterclass went on to lead after our exchange. In his own words, these aim to get young people excited about the bridging of algorithms, computation and statistics. “This brings in the uncertainty. With economics, you introduce ideas of incentives, fairness and people being justly rewarded for their efforts. I want students taking my classes to perceive these new challenges, to combine these ideas in new ways, to try to make them operative in real-world systems.”
Yet, Jordan doesn’t underestimate the challenge this poses, demanding as it does a high degree of mathematical training and an excellent grasp of classical maths, physics and economics. But not only: “If I had to advise students today, I’d encourage them to follow their own nose. That’s what I did. I studied some philosophy and psychology, I learned languages, I explored widely following my own interests. Nowadays, I think it’s just as important to enjoy the breadth of experience and expose oneself to fields of social sciences and humanities as it is to learn science and maths.” All these broad explorations need to lead to real world problems, however, with a direct and immediate impact on the present era. “This new direction will be nourished by a collective effort, bringing together a broad palette of people. So, you need to understand how you can help with climate change, healthcare, democracy. It’s not just about making money or being famous. As you get older, you don’t care about these things, you care about having an impact. With this interdisciplinary approach, can you share a project or an area where this interdisciplinarity can lead to an unexpected insight or breakthroughs?”
Exploring Asymmetrical Information
At present, Jordan’s own deep well of interdisciplinary experiences has motivated him to explore the question of information asymmetry. The researcher disputes the idea that the gathering and processing of information is done centrally with a multinational building a system to channel it for everyone’s profit. “The reality is that everyone has valuable information of their own or they have a creative cultural artifact. They’ve composed a song, written a poem or whatever. And it's not just to be aggregated and processed. You want to think about information more in an economic sense. We're linked in various ways, and we want our creativity to be respected. I want to be connected to people who are interested in the things I've done. And so, the kind of things I've been interested in are these information asymmetries. That's an economics term where one person knows a lot of things and some other person doesn't know so many things, but the latter person wants to incentivize or to motivate the former to work on something. Well, that's how human societies often work, and we haven't recognized that very much in our machine learning and statistics.”
With his Berkeley research team, Jordan is devising a menu of statistical contracts that balance out the information asymmetries. He believes contracts are embedded in everyone when they make a choice for, say, more leg space on an airplane. One of his focuses has been on drug regulation in the US. “I've been interested in how a regulatory agency for medicine decides whether a drug goes to market, or judges if it is even a medication? It's an asymmetrical situation because the candidate medicines are coming from various pharmaceutical companies who would love to have their drug on the market because they would make money. But that's not the right incentive. They should also want the drug to work and not to hurt people. In the United States, it’s the Federal Drug Administration, the FDA, which screens and tests them. We’ve developed a statistical contract theory which sets up protocols whereby you can interact and align the incentives. And the result is that good drugs go to market.” The recipient of the IEEE Pioneer Award hopes his new mathematical theory will address the design problems that pharmaceutical drug companies pose.
The Misnomer Called A.I.
It is with the hot topics of Artificial Intelligence and ethics that the Professor in the Department of Electrical Engineering and Computer Sciences at Berkeley really warms up. With the former, Jordan has been making waves since 2018 when he declared that the AI revolution has yet to take place. “Most of us working in this field for the last 20, 30, 40 years don’t use this terminology,” he says firmly. “We've called it either machine learning, control theory, engineering or just statistics. And I think it's a bit unfortunate that that term has now been brought back to refer to all these engineering developments. The term was originally used for a philosophical aspiration, and that’s not really what's happening. Those who use it think of the computer as having a brain hardware with something like a brain and a mind that's resembles the software. They believe we can figure out the algorithms in a human brain, put them in a computer and the computer will think, reason, and plan in the same way a human does. I don't think that's what's happening.”
“Admittedly, we've seen search engines and recommendation systems arise,” he pursues. “They’ve changed our lives. But that isn't so much about an intelligent computer. A computer with a search engine indexes the web, and so it's not that intelligent, it's just indexes things. But it aids our intelligence, so it makes humans more intelligent, more interactive with each other. It connects us in new ways. It provides the world with knowledge at our fingertips, so it augments human intelligence.”
Jordan believes that the focus is more on what computation has done to people’s lives and not imposing super intelligent entities in our midst. “But,” he adds, “people are fascinated by the idea of interacting with this entity and the terminology that guides our lives. The problem is that the real challenges are elsewhere. They’re centered on building a system that includes computation, people, and the real world. And this system’s impact on transportation, finance, healthcare and so on. This means bigger systems involving networks, data, and algorithms.” He pauses: “We have to make it clear what the business model will look like and what effect it will have on society. I think we're only now realizing the problems with not thinking through what the impact of the technology might be. And so, I don't like AI because it prevents you from thinking through the real impact of the technology and the effect on collectives of people instead of on single individuals.”
Where an Economy of Data Meets Ethics
Michael I. Jordan has similar issues with the question of ethics in science. “I find the essential discussion about ethics to be extremely limited. Ethics is about doing right by people and making sure that people are treated fairly and there can't be just a technology-meets-moral-philosophy approach. There has to be domains in between including law, economics, a sense of fairness, respecting privacy. And I want my utility, my values assessed as part of that determination.”
The co-author of “How AI Fails Us”, published by the Edmond J. Safra Center for Ethics, elaborates on an all-encompassing vision: “I should decide how much privacy I want. Maybe I don't want a huge amount of privacy for some things. Maybe I want to share some things widely. There should be an economy of data that takes into account my desires, and so if you don't build systems like that, that's not ethical. If someone decided for me or someone laid down the rule of law, I would consider that unethical even if they think it's ethical. I need to be consulted on this, all of society does. Treating everyone fairly does not mean treating everyone the same.”
"And so," he continues”, “ethics also includes just building good systems. If you were an electrical engineer and you put wires in houses without checking that they don't burn the whole place down, well, that would be unethical. You would have been responsible for the deaths it caused. Similarly, building systems and hospitals that don't test out drugs effectively or don't think about allocations, well, at scale, that's unethical.”
Jordan is convinced the dialogue must be broadened beyond the classical notions of an intelligent agent making ethical decisions. And this goes back to the limitations in AI: “Think about the main goal of intelligent computers: being autonomous in our world and making sure that they are ethical like we are. That's missing the entire point of the technology we have at our disposal. It has to be much broader and richer and deeper. And so, we have to think about the overall goal of the system, making sure it works, that it's fair and respects human desires.”
With those final words, the professor smiles his broad and disarming smile, shakes hands and leaves for a packed amphitheater in the IP de Paris campus where students from two of France’s Ivy League schools wait for a Masterclass labeled one of the year’s “must-attend” sessions on data science.