We speak to the eminent mathematician about machine learning, artificial intelligence, and their limitations.
BY DEBDUTTA PAUL
Michael I. Jordan is the Pehong Chen Distinguished Professor in the Department of Electrical Engineering and Computer Science and the Department of Statistics at the University of California, Berkeley. He has made foundational contributions in several fields, such as cognitive science, machine learning, and computer science. Professor Jordan is a member of the National Academy of Sciences, a member of the National Academy of Engineering, a member of the American Academy of Arts and Sciences, and a Foreign Member of the Royal Society, and has won several awards, such as Ulf Grenander Prize from the American Mathematical Society in 2021, the IEEE John von Neumann Medal in 2020, and the David E. Rumelhart Prize in 2015 among others. Professor Michael Jordan (MJ) spoke to Debdutta Paul (DP) on his recent visit to the ICTS-TIFR for the discussion meeting on Data Science: Probabilistic and Optimization Methods.
The full text of the interview is reproduced below. The answers are lightly edited, and long paragraphs have been split up for readability. The questions and initials are in bold, and DP’s additions are in square brackets. A shorter version of this interview first appeared in the ICTS Newsletter, volume IX, issue 1, 2023.
DP: Welcome, Professor Jordan. How do you explain to a high school student about your current field of research?
MJ: Well, I’m a statistician, fundamentally. So I’m interested in collecting data and using that to infer things about the world. And that happens everywhere. So medicine collects data on a patient and [figures out] whether a treatment works or not. In commerce, you try out various business models and see how much sales you receive. And all these domains are changing because of the availability of huge amounts of data. And this includes commerce that brings supplies around the world and puts it all together efficiently. And these are [the] kind of issues that help us to scale with a planet where there are billions of people, and there’s going to be more and more over the future, and giving everyone access and using data to help provide new services to people and to make things more transparent, and so on. So a social agenda [is] put together with large-scale data analysis.
DP: We’d love to know about your journey from how you came into it, from psychology to maths to cognitive science to being a statistician now.
MJ: I don’t think it’s a compelling story; it’s random at some level. Just following my curiosity. I was curious, early on, about philosophy and neuroscience, how the brain might work, and psychology. And eventually, you have to pick a career, and it didn’t feel quite right for the career; it didn’t feel like it would move fast enough. And I would like mathematics. So if you think about humans and how we’re intelligent in some ways, we’re like a statistician. We get partial data about the world around us; we’re uncertain. We make plans over time; we try to make good decisions. So our brain is kind of like a statistician. And I thought, well, why don’t I just study statistics a little bit more and think about the principles that allow us to be uncertain about the world but still make good decisions in the world? And if we do that, not just for classical statistical problems but new emerging ones in commerce, transportation, finance, and all that, it feels like a really interesting research agenda. So little by little, I just moved in that direction. I learned enough tools and mathematical ideas that I could start to do it and have new ideas. And then I spent a little time in [the] industry, not full time ever, but partial and sort of solve real problems in the real world. And that drove a bit more of my thinking and the kind of these large-scale social phenomena you see in [the] industry.
DP: But would you suggest that model to new students?
MJ: No, I don’t suggest any model to any students, I mean, everybody is different, and we’re all slightly random. If we rerun my path, it wouldn’t be the same. But I think there are some parts of it that may be useful to keep in mind. One of them is, don’t be in a rush. If you can go a little slowly, that’s probably good. Try to be as broad as you can while not neglecting a little bit of depth. And try to manage that so that you can have a career. So be realistic about your possibilities. But the more you know, the broader and deeper your knowledge is [and] your education [is in], the more you will be ready for new opportunities when they come. And that’s really important. So I think, in my case, there were a few things that arose that I knew what to do immediately, and I could sort of take those opportunities.
DP: As a subject, what is machine learning or artificial intelligence about?
MJ: Well, it’s really just this blend of statistics and computer science. And to me I also in my recent career, I bring in economics — it’s kind of a part of it. How do you connect producers, consumers, or people who have something to offer others? So, machine learning is, in some sense, not new. It’s statistical principles about how I take partial knowledge about the world and try to infer what’s out there like a scientist would do. These are complementary fields. Computer science has been more about how you programme a computer, [and] how you ensure it works. How does it follow certain algorithmic steps and do things? And I was more concerned about the internal world of the computers, that everything was correct and right inside. And statistics is more about the outside world: What’s happening in the outside world? How do I gather evidence about that and try to make decisions about that? And you glue the two together, and that’s a powerful combination. In some sense, that is really what machine learning is. But it’s also a little bit more of an engineering field. Because statistics focus mostly on scientific inference and trying to help scientists discover things, [like] the Higgs-Boson discovery, for example, use statistics to decide if you did discover it or not. And machine learning has a little bit more of a computer science and engineering flavour. How do we build systems that automatically analyse data in various domains [to] help us make our own decisions? Also, things like automatic self-driving cars will use machine learning, and they try to figure out what’s in the world with statistics. And but they do it rapidly and, and do it on a very large scale. So it’s more of an engineering system and less of a science system.
DP: Can you tell us specifically any one problem that excites you nowadays?
MJ: Yeah, sure. So as I alluded to earlier, whilst historically machine learning has been about statistics and computer science together [with] algorithms that make statistical inference, I’ve been more interested recently in economic models being brought to bear and being part of that whole story. So why economics? Well, when you think about where data comes from, who it benefits, and how it is used, you really want to think about a big network. You know, companies collect large amounts of data, often collect them from cell phones or from other sources that individual people have produced. So that’s more like a market. You want to think about the people being agents and main players in that — they should opt-in, [and] they should decide they want to participate. And any contribution they make with their data, which might be something they write or create, should be valued and part of the opt-in process. So there are ideas in economics that definitely talk that kind of language. So one of them I’m interested in right now is something called contract theory or principal-agent model, where there’s one entity, maybe a human, maybe not, maybe an organisation that wants to accomplish some task. And there’s some other entity and agent that has more knowledge about how to achieve that task. And they kind of want to cooperate. So the principal’s got to incentivise the agent, who only accepts if it’s in their interest. If you now think about data in this world, the principal wants to incentivise the agent to provide data. So their data becomes an economic good. And people, in fact, often create data at great expense, or based on their knowledge, and so on. So this model allows the data to be treated as an economic good, traded, aggregated, and used for various purposes. So the mathematics is interesting. And the implications, [and] the use cases are really interesting. That’s the areas that I’m most interested in right now.
DP: In your recent Infosys-ICTS Turing lectures, you talked about social intelligence. What are your thoughts on how ML/AI researchers and policymakers perceive ‘social intelligence’ today?
MJ: I didn’t really use that word, but it is a nice word. It’s a nice phrase, [by which] I mean systems, [let’s say] that bring food into Bangalore every day somehow have social intelligence. They are social. People make individual decisions — I’m going to bring tomatoes over here because I’ve got tomatoes, and you don’t. Everybody’s making simple local decisions, but the overall effect is that enough food arrives for all the people in the city every day. It’s not always efficient, but it’s pretty good. That’s an intelligent system, and it’s a social system. So to be part of what AI should be is to mimic that kind of system. And classical AI didn’t think that way. It was more about mimicking the individual human and trying to be as smart as a human. That’s a different goal. I like this more social goal. So if you put the two together, the computer is now part of the overall social system, and it knows things, and maybe it’s better in some ways than humans in some things but not in other things. Then the overall system could be better and more effective for everybody. But you’ve got to have it clear that your goal here is to have high social welfare and have the system work for people.
DP: Do you think researchers and policymakers look at it that way, or is there a need for change?
MJ: Not enough. I mean, right now, there’s too much mystique about machine learning and AI. Some people think it can just sort of solve all the world’s problems all by itself. Some people fear it because it feels like it’s going to do things that take away things from what humans do. And there’s a lot of misunderstanding about what it can and can’t do, and just not a lot of thoughtfulness about it. It’s true that there are certain kinds of things they can do that are going to take away some jobs. There are also, especially in this social intelligence model, new jobs [that] can be created and ways to think about that. And to me, there’s not enough discussion about that, then those, to me, are the real problems. It’s not really about — Are robots going to take over and kill humans? — And all that kind of science fiction discussions. But to me, it’s more interesting and important to think about: Okay, what kind of jobs are at risk? How fast is that going to happen? Are there other kinds of jobs that are going to be created? How can we incentivise that? How can we make sure this is fair? How can we ensure that everybody participates and it’s not just a small power set? And there is some discussion of that, but just not enough. You don’t see most articles in the newspaper [talking] about those kinds of issues. It’s always about the more exotic fears and dreams.
DP: Some scientists fear that this is the time for meta-principles. Science has been conducted through observations, extracting principles or laws from them. Given that we have machine learning now — was that a limitation? Do we really need laws? Or is the concept of science itself being challenged by ML and AI?
MJ: I totally disagree. I mean, laws can be local and contextual. I think we’re used to laws being, F equals m-a, or [the] law of gravitation, that it applies everywhere, for everything. And that was important and beautiful. But there are also laws that only apply to certain ecological niches, or in certain kinds of species, certain kinds of social interactions, or even certain kinds of fluids and certain kinds of physical systems. And maybe they don’t have the vast reach and power of F equals m-a or [the] law of gravitation. But they’re super interesting and exciting to work on. And there are scientific ideas to be discovered there. And observations are needed, and thinking is needed. Thinking about the immune system of a human — how that works. It’s very rich and complicated. And there are principles required to understand it. It’s not some crazy system; it’s got principles. So there’s lots of science, and it just becomes a little more contextual. And I think that’s actually valuable. Studying genomics, for example. Genomics allowed us to see that DNA is composed of lots of genes, and each gene has its role. And instead of just worrying about a big law for all genes, whatever that might mean, you try to think, what does each gene do? And how does that gene participate with other genes to make an organism function? So these are all more contextual stories, but they’re, to me, just as interesting and powerful and important as the principles that the early physicists and biologists came up with.
DP: And it’s still valid to continue making principles and developing principles?
MJ: Absolutely, unquestionably. Otherwise, you really can’t predict, you can’t have a notion of stability, [and] you can’t also try to build a system that behaves in a desired way. It’s definitely not just the machine takes over, or we write down a list of things we don’t understand. It is about the simplified abstractions that allow us to reason and make some sense of our world.
DP: Coming to the applications of machine learning: ML/AI are used as tools in science. For example, in your talk earlier this week, you mentioned AlphaFold. There’s the Event Horizon Telescope, and you mentioned other astronomy projects. What is its role in automating science in the future?
MJ: I don’t really know. That’s a little far beyond my scope of knowledge. The word “automating” — I’m not sure exactly what that means. I mean, I do think that human curiosity is always going to drive things and human insight. That [if] something is important [it] really requires seeing what all the consequences of that are and see what could change, and doing ‘What if?’ experiments and all. And I think humans will be, for the rest of our lifetimes, at least, really good at that relative to what machines are. So, smart humans will be able to use these machines in new ways — have a bigger scope than they had, just like computers can solve partial differential equations (PDEs) that we can’t solve that helped science. Similarly, here, I think it’s going to drive innovation, and it’s going to drive possibilities we didn’t think of, and some part of that will be more automatic, just like the PDE solution was a little more automated, but [it] doesn’t mean we’re just going to push a button and the computer will solve our science problems for us. I just don’t think we’re aspiring to that. I don’t think it’s realistic. I think it’ll just allow humans to think a little bit more broadly, and maybe a little more carefully, maybe have their ideas have [a] broader scope. Think a little more about safety issues, fairness issues, and things that weren’t on the table.
DP: Right. But in this context, what is the future role of scientists? Should they be adaptable to the changes?
MJ: I don’t think it’s that different from when a telescope came. Instead of the role of a scientist walking outside with your eyes and you look at stars, now you build a better telescope, and you think about what more things? Could I measure the infrared? Could I measure this and that? And then the machine helps [you] see things you could have never seen. And you envisage how you could use that in new ways. And I think, conceptually, it’s not that different. Yeah, AI is still really subservient to the human. I think that’s going to be true for quite some time. The ‘automatic scientist is an AI’ — I still think that’s kind of science fiction. It’s not clear why we would do it. Little by little, yes, it’ll become a little more automatic. But I think it’ll maybe drive us faster to open up new questions that we didn’t even think about before. Someone was talking about chess playing. At some point, the computer got better than humans in chess, okay? But that didn’t mean that all interest in chess went away. In fact, I think it’s been quite the opposite. Humans have seen the computer play. They say, “Well, that’s interesting. I never thought about that.” And then they think about the consequences. And they try it out, and they get pleasure. And in, it’s still a way to help humans grow by playing chess. So just the fact that there exists an entity that could do something automatically doesn’t mean humans won’t want to do it themselves.
DP: But how do we go about training people for the science ecosystem, as well as for the industrial work, the changing job market, especially in a country like India? How do we collectively evolve and adapt to AI for the positive?
MJ: I don’t think you have to ask me. I think the 20-year-olds will figure that out. Honestly, a 20-year-old still needs to learn mathematics; they need to learn sciences; they need to learn something about humanities; they learn to be an evolved person. They may not have to learn as much programming because these systems can do a lot more programming for you. That’s fine. Maybe learning all the details of [the] syntax of a programming language — you could do it, but it’s not necessary. Well, they can do other things. When you see a 20-year-old playing around with ChatGPT, they very quickly understand what it can do, [and] what it can’t do. They play around with it and get some value from it. And they can learn. And we just have to remind them, that’s not the end of the story, that there are other principles to investigate it and use it now to investigate those principles, and think about medicine, and think about commerce, and think about agriculture and all those things. The computer is not going to solve those things for you. But if you are clever at using the computer, you could help contribute to that. So yes, in some ways, it hasn’t changed — [you need to] understand mathematics, understand human history, understand biology and physics and all. Mathematics should also include statistics, as I’ve alluded to. Many previous generations didn’t do much statistics, or treated it just as a mathematical exercise. That’s just changed. Statistics is very much not just a mathematical exercise; it’s really analysing data and making inferences with it. And people need to be more empowered to do that. And think for themselves in the data analyses.
DP: Now that you’re talking about ChatGPT, I had to ask this question. What are your thoughts on this new phenomenon that has taken the world by storm?
MJ: I already alluded to it. I mean, because it’s human data, we already had a fair amount of clarity that machine learning could make great predictions and things like supply chains or recommendation systems. The backbone of many companies has been machine learning for quite some time now, and they’re pretty good. When it became language data, and it’s really doing these things that look like only humans could have done that, that’s definitely surprising. And the architectures because they can scale. They can take in trillions of pieces of data. It is rather amazing how fluent it can be. But it is also true that you have to build things around it. You can’t just take the output of this system as the truth. It doesn’t know the truth. So there has to be work around it to build a better-engineered system. Instead of just outputting the name of the Prime Minister of India as x, well, maybe in the data, it was a certain name, and now it’s changed! ChatGPT doesn’t know that. So instead of having ChatGPT say the name, it might say ‘Prime Minister of India’, and then you go look it up in a database, and someone maintains the database. So there are systems around ChatGPT that will help it do the right thing, say the right thing, be more context-aware, and so on. So that’s like in any engineering field, you take a powerful tool, and you build around it to make it more approachable, more usable, more safe. So ChatGPT itself is fluent but could be dead wrong in many situations. That’s not just fixed by having more data. That’s fixed also by thinking about what to build around it.
DP: What are your final thoughts on these topics that you engage with — for the general public, for school students interested in the topic, as well as for ICTS researchers?
MJ: I would be more excited than fearful right now. It does change certain things. An example for a student is the essays that you have to write in high school and college. A teacher who ignores the fact that ChatGPT is good at writing essays will not be a good teacher. They’re not going to be helping the student very much. But a teacher who says, “Okay, let’s embrace it, it does exist, let a student start with a ChatGPT-generated essay, and show me what they’re starting with, and then they help correct it. Tell me what they could do to make it better. And then I’m going to tell them what they could have done. And I’m going to work with them to sort of see that.” Because a human-written essay on something that’s really written well, we can sort of tell. It communicates in a certain way. There’s depth, [and] there’s a human experience coming through. So teachers can help with that. They don’t have to help as much with all the grammar issues and fluency issues. Just like the calculator helped us not to worry about all the algorithms [with which] we’re adding and subtracting. There were some people who said, “That’s terrible; it’s ruining children’s minds.” I don’t think so. I think that the calculator allowed us to say, “Okay, arithmetic is handled over there. I can now use it in new ways.” Similarly, here the essay can be written, it’s not too bad, [and] it sounds pretty good. It gets you started. Maybe you think, “Okay, I like that. But I would do it a little bit differently.” And again, a teacher can help you with that process. And now you take it and reshape it, and so on. Similarly, for artistic things, it’s not that DALL-E or whatever ChatGPT-like image generators are going to take over art. But they’re a good starting place. Someone could say, “I want to have a scene with a mountain in the back and a horse in the foreground.” It draws, and it comes out looking like that; it’s pretty impressive. But the human will rarely just look at it and say, “I’m done.” They’ll say, “I would like to do this and change this-and-this in this way.” And partly, it’ll just be going back to ChatGPT and working with it. But that could that’ll start to become a very creative act. I can imagine younger people not just being content with just pushing buttons. Good teachers will themselves learn how to do that, [and] they will engage with it. And they will learn this and [the] best practices and things they can teach students. So it’s going to be a demanding process for teachers to embrace it and not fight it. But rather, it’s real, [so] use it, and find new ways to help students engage with it, not be fearful, but also not be intimidated.
DP: And what do you think policymakers should be doing?
MJ: There, I have even less to say, frankly. I think policymakers best in these situations should sort of sit back and wait. I think if they act too quickly, they will be acting on all these kinds of fears and suspicions and misconstruals of what’s really happening. They really should have more dialogue. They should have people talking to [those] who do know a little bit of what they’re saying. So, for example, I’m saying, “Hey, the college essay is going to be a problem now.” They should have a commission that starts talking about that. “Okay, teachers, how will we approach this?” “What if we allow all students to use ChatGPT in the curriculum? How would we have to change?” Have a committee that writes a report, and then have the policymakers look at it and think about it. You’ll start to envisage the world that incorporates this and not just try to regulate it immediately. Also, I do think this whole field is an international cooperative effort. It would be unfortunate if we continue to see [the] nationalist tendencies that “We have to own it.” Or “We have to do better.” Or “We have to not cooperate.” That would be unfortunate for young people and the development of the field and its understanding. So I think policymakers should also be ready to understand the international open nature of this kind of technology and to support that.
DP: But what do you think about the privacy concerns? And the concerns about a few companies having a lot of power over many people?
MJ: Yeah, these are all reasonable concerns. And again, you realise it’s a real concern. Then you think: What’s wrong in the current system? What could be fixed? What’s needed to fix that? So privacy, for example, is absolutely important. But we need to be sophisticated about what privacy means. Okay, so privacy does not mean anonymity — that no one knows anything about me. Most of humanity, throughout history, grew up in villages, not in big cities — they were not anonymous. And that’s still true, I think, in much of India, if you live in a village, you don’t have a lot of privacy. People know a lot about you. And in some sense, that’s very good. Because if something bad is happening to you, they know, and will help you. So humans depend on this lack of privacy in some domains. And we have this myth, I think, that if we went to the big cities and we should be able to be completely in control of all of that. The downside is that if something bad is happening [to you, then] no one knows, and then we don’t get any help. So you want to be able [to] have a little control over the privacy. That means it’s not a zero/one. The company that tells you, “It’s absolutely private, don’t worry,” or, “Hey, I’m not gonna give you any guarantees.” It should be, “Hey, I’m gonna give you a little control; here’s how I’m going to use your data; here’s the privacy guarantee I’m going to give you; I’m going to add some noise to your data or something so that someone can’t tell that you’re in the data or not.” So, there are ways to work on that as a technical person. But there are also ways to convey that to people. And there are also ways to ensure that companies develop models that people agree that, if you want my data, like my medical data, for example, it’s private, obviously, at some level, but if a hospital wants to use it, to study some disease that runs in my family, I probably want to give them that data. So I want to adjust my concerns about privacy depending on the problem and the context and all that. So we need to have transparency and openness about that. That’s just one example of fairness, privacy, [and] cryptography. All these issues require a little bit more of an in-depth discussion and not just saying, “Oh, it’s now probably worse than it was.” Well, certain things are worse, certain things are better.
The other concern you mentioned is [the] concentration of power. And those of us who work in the field, in some ways, actively fight against that. So first of all, all of our ideas are on the arXiv, which is a worldwide open resource. There are no ideas I know of that are important [and] are secrets. The cloud gives access to people for computing. Cell phones are access points for all kinds of AI-related things. But it’s definitely true that there are companies who are incentivised — they’re not bad companies, necessarily, but it’s just that they’re competitive with other big companies — so they try to grab things. One example is just grabbing, you know, Wikipedia or other data that people have created, grabbing that, making a product like ChatGPT, and then selling the product. And I think that’s wrong. Basically, I think that’s economically not healthy. So I work on [the] part of the economics ideas. So I can find ways to ensure that when you create something of value, you’re part of a market where that thing is valued, and you actually get paid for it. So that’s one way of decentralising. Anyway, these are part of overall technological development. It’s something like the steam engine 150 years ago — when it came out, it definitely caused some problems. Chemical factories caused some problems. And if you focus only on the problems and try to regulate [them] immediately, that’s a mistake. You want to think about what’s the real problem and [how] can we mitigate it, [how] can we move away from that, [how] can we move towards something else, and worry about the trade-offs.
DP: ChatGPT takes a lot of data from Stack Exchange. But then, if people only go to ChatGPT, there’s not enough community knowledge being created.
MJ: Yeah, so that’s broken, and it’s wrong. It’s kind of ethically wrong at some level. But just saying that something’s wrong allows you to complain and write a newspaper article, but it doesn’t actually change anything. So changing things means that people are organised, you start to develop an economic model, and you only provide it under certain terms, and that will have to happen. There are things called the gig economy and participatory budgeting, and other mechanisms that are certainly being discussed now. I think those are really important.
DP: Thank you so much for your time, and it was a pleasure having you here.
The author thanks Dr Praneeth Netrapalli for the arrangement.