Welcome to Machine Learning, How Did We Get Here? I'm Tom Mitchell, your podcast host. During the 1990s, one of the key developments in machine learning was the introduction of learning of Bayes nets, or sometimes they were called probabilistic graphical models. The idea there was to represent probabilistic knowledge in a way that had been introduced by Turing Award winner Judea Pearl. But his work left open the question, how would we learn these from data? One of the key contributors to machine learning algorithms for learning Bayes nets and for working with probabilistic methods more generally is Daphne Koller, our guest today. Daphne is an elected member of the US National Academy of Engineering and of the American Academy of Arts and Sciences. I hope you enjoy the conversation with Daphne.
Today I'm happy to have with me Daphne Koller, who is one of the pioneers in machine learning. Daphne, good to have you here.
Great to be here, Tom. Thank you.
Now, you've done a lot of work in different aspects of machine learning, but you're well known on and you wrote a textbook on probabilistic graphical models. So, let me start by just asking, can you describe probabilistic graphical models in a single sentence?
Oof. Probabilistic graphical models are a mathematical framework that allows you to model the interrelationships and specifically importantly causal relationships between different variables in a complex domain. And then be able to ask questions about some while observing others in a way that respects the rules of causality.
That's great. Now you can add as many sentences as you like. And let me just ask, how did you personally get into this field? What was your background that you could have done a hundred thousand things. How did you end up doing this?
So, throughout my life I've had an interest in complexity and putting order and getting order out of chaos and coming up with mathematical frameworks that allow you to think about complex phenomena in a rigorous systematic way that isn't just gut intuition. So, I started out my career in actually thinking about game theory and looking at the interactions between multiple agents and how they can rationally make decisions in that context. And having worked on that for a while, I came to the conclusion that actually there's a lot of complexity and open questions even in just a single agent having to make decisions in a complex varying environment and we needed to think about doing that before we left all the way into the multi-agent scenario. And so I kind of ratcheted back almost and said, 'Okay, fine. I'm going to focus first on single-agent systems.' And then spent a little bit of time saying, 'Okay, well, how do I model a complex environment so that agents can make smart decisions?' And that's where I got into probabilistic graphical models to begin with. And came very quickly to the realization that the big bottleneck was getting people to write down world knowledge in a way that actually reflected the complexities and subtleties of the world. And so this task that used to be called knowledge engineering was really painful, onerous, and frankly impossible for people to execute well, which led me naturally to the path of, 'Well, let's just learn it from data.' And so I came into machine learning through reasoning and needing to come up with a better way of getting models of complex environments. And so when people today ask me, 'How did you start doing AI?' my response is that I didn't ever officially start doing AI. I started doing a thing that was at that point not considered AI at all. I mean, we were at best fringe AI at the time that I started doing it and excluded from AI in the minds of many then practitioners of what was called AI at the time. It's more that AI became the thing that I was doing.
That's great. So, when was this, roughly time, what year was?
So, I started working on probabilistic models around the early '90s, so '91, '92. And by, you know, '93, '94, I was kind of like full-fledged into the machine learning side of that work. And then as I went into Stanford as a faculty member in 1995, I became, I would say, a more broad-based machine learning person in terms of the work that I was doing beyond just the probabilistic graphical models. I also worked, as you may remember, Tom, on active learning and transfer learning and multitask learning and things that are now considered sort of the norm of how we do machine learning. But back in the day, if you remember, those of us who were around in those early days, machine learning was viewed very much as a very simple linear task of you have your input, you have your one target output, and you just want to eke out as much performance as you can from solving that one simple supervised learning problem. And God forbid, active learning, transfer learning, multitask was all sort of considered very much, I would say, outside the scope of normal machine learning work at the time.
So, you got involved right at the very beginning of the work on graphical models. Pearl had just published his book, I guess. So who were the people you interacted with in that subfield? What I mean, you know, did you have people who were inspirational people to you or mentors or was it pretty much you were coming in this from a different angle?
Well, I mean, Judea Pearl obviously is one of the most impressive and forward-thinking people that I've had the fortune to know. And so he was clearly a bit of a role model on how do we think rigorously about questions like this. There were not very many of us at the time. Several of us are still around. So, Eric Horvitz, of course, I still hang out with him. He's the chief scientist of Microsoft and has been super influential in building some of those ideas into Microsoft and also leading a lot of Microsoft's interest in AI more broadly. I mean, Stuart Russell, my postdoc advisor, was also very much an early leader in this space and an advocate for the inclusion of more rigorous systematic ways of thinking about uncertainty and decision-making and rationality that were very foreign to the field at the time. And Stuart's textbook on AI, I think, was the first time someone laid a stake in the ground that said, 'This is a different way to think about AI that goes beyond logic-based systems and expert systems.' Which was, I know, seems so remote when we think about modern-day AI, where no one even barely remembers those days. But that was the norm. I mean, in 1985, which was before I even started my PhD, Judea told me the story, there was this panel at IJCAI, which is the big AI conference. And this was one of the previous AI bubbles that unfortunately got punctured just as I started my PhD. But because it was based all around logic-based systems and expert systems that were never going to scale. But the panel was on the topic of, 'Do numbers have a role to play in artificial intelligence?' And the consensus of that panel, with the sole exception of Judea Pearl, was no. People do not use numbers in how they think about the world and therefore our AI systems should not use numbers either. And when you think that today, where there is really no other AI other than based on numbers. And by the way, also, if you think about neuroscience and how human intelligence actually works, that seems ridiculous when you think about the study. But that was the consensus of the panel back in 1985, which maybe explains why that particular bubble burst and it took a while before things recovered. I, you know, when I graduated with my PhD from Stanford in 1993, you couldn't actually say you were doing AI. You had to say you were doing cognitive computing or statistical learning theory or some other thing that was not constraint satisfaction. You couldn't say you were doing AI because that was considered a fringe discipline at best.
I want to get your picture of how the field evolved since you were involved. But let me start by asking specifically about graphical models. So you were right in the midst of that. How did that feel look like kind of the key steps? How do you see the evolution of that subfield of machine learning?
So probabilistic graphical models for me emerged by realizing that the problems that we needed to solve to address most real-world applications went beyond you have a vector representation of an input and a single often times binary or at best continuous output. There was so much more opportunity to think about richly structured environments and richly structured problems. So even if you think about problems like understanding what is in an image. That's not a single label problem of there is a dog because images are complex and there is interrelationships between the different objects in them and you wanted to get beyond the yes no is there a dog in this image to something that is much more rich of, you know, there's a dog and a frisbee and a beach and three kids building a sand castle. And so you have a rich input and a rich output. And so thinking about these richly structured domains gave rise to well, we have to think about multiple variables. We have to think about the interactions between those variables and leverage that structure both in our input and our output space in order to get to much better conclusions and deal with the problems that really matter.
Right. And so the field kind of evolved from maybe first having this notion of we can represent these joint probability distributions over these multiple components with some kind of graphical representation. But then the question of how to do inference, how to do learning on that. There were some theoretical developments about tractability, learnability. There were algorithmic developments about tractable approximations for inference. If you think about where we are today in terms of our understanding of how to build such systems compared to when we began, what are the kind of key developments that you saw or see over the years as that evolved?
