Microsoft Research Podcast - 045 - Leading Labs with Dr. Jennifer Chayes

Episode Date: October 10, 2018

2018 marks the 10th anniversary of Microsoft Research New England in Cambridge, Massachusetts, so it’s the perfect time to talk with someone who was there from the lab’s beginning: Technical Fellow, Managing Director and Co-founder, Dr. Jennifer Chayes. But not only does Dr. Chayes run the New England lab of MSR, she also directs two other highly renowned, interdisciplinary research labs in New York City and Montreal, Quebec. Add to that a full slate of personal research projects and service on numerous boards, committees and foundations, and you’ve got one of the busiest and most influential women in high tech. On today’s podcast, Dr. Chayes shares her passion for the value of undirected inquiry, talks about her unlikely journey from rebel to researcher, and explains how she believes her research philosophy – more botanist than boss – prepares the fertile ground necessary for important, innovative and impactful research.

At first, I really knew nothing. I didn't know the acronyms. I knew no computer science. When I first met Bill Gates, a few months after I came to Microsoft, I was asked to give him a talk and I didn't know how to use PowerPoint. So I did it with handwritten overhead transparencies. They told me it was the first time an overhead projector had ever been in Bill's conference room. And when I met him, I said, Bill, I want to congratulate you for hiring a group that won't pay off for 100 years. And Nathan told me to shut up. And Bill said, No, no, it's fine. There aren't enough of you to worry about. You're listening to the Microsoft Research Podcast, a show that brings you closer to the cutting edge of technology

research and the scientists behind it. I'm your host, Gretchen Huizenga. 2018 marks the 10th anniversary of Microsoft Research New England in Cambridge, Massachusetts, so it's the perfect time to talk with someone who was there from the lab's beginning. Technical fellow, managing director, and co-founder, Dr. Jennifer Chayes. But not only does Dr. Chayes run the New England Lab of MSR, she also directs two other highly renowned interdisciplinary research labs in New York City and Montreal, Quebec.

Add to that a full slate of personal research projects and service on numerous boards, committees, and foundations, and you've got one of the busiest and most influential women in high tech. On today's podcast, Dr. Chais shares her passion for the value of undirected inquiry, talks about her unlikely journey from rebel to researcher, and explains how she believes her research philosophy, more botanist than boss, prepares the fertile ground necessary for important, innovative, and impactful research. That and much more on this episode of the Microsoft Research

Podcast. Jennifer Chays, welcome to the podcast. Thanks for joining us today. Thank you. You are a busy woman. You're a technical fellow, managing director of three Microsoft Research Labs, and you're actively involved in ongoing research projects yourself. So I'd ask you what gets you up in the morning, but it's possible you never go to bed. What drives you? Why do you do what you do? First of all, I love the people I work with. I just have such a creative bunch of people here in all three labs, actually. So that is definitely something that motivates me. A second thing that motivates me is my own research. I love doing research in different fields and in new fields and in helping to take real world problems and

abstract those problems so that other people can start looking at them. Other academics can start looking at them. And finally, I also get really excited about the impact that we can have. You know, Microsoft has about a billion customers. And if there's something that we do that can impact even a small fraction of Microsoft's customers, that's really exciting. Yeah. So I want to talk a little bit about the breadth of work you do. I mean, you're not just doing your job in the labs and with research, but you're involved in the broader community. What drives you to be so involved outside of the company as well? Well, I have a lot of opinions. And so when I'm asked to be on a board,

I think, okay, well, I have an opinion about the issues that that board deals with. And I either have to get involved now with this opportunity to maybe change things, or I have to forever hold my peace. And I'm really not good at forever holding my peace. So I get involved and it's really fun because I can have an impact in other ways and I can bring the insight of Microsoft to some of those institutions. I can tell them about things I learn about cloud computing,

about fairness and AI, about all kinds of things to those institutions. And I can also make sure that women and minorities are represented in all of these activities because, you know, sometimes people just forget. It's the 10th anniversary of the MSR Lab in New England. That's in Cambridge, Massachusetts, which you co-founded in 2008. In 2012, you co-founded the New York City Lab. And then Montreal came into the fold in 2017.

What I want to know is what's the story of how and maybe more importantly, why MSR started these labs in these different cities and what each one brings to the research party, so to speak? So first, starting with the New England Lab in 2008, I had been in the Redmond Lab for 11 years and had been seeing increasingly that there were opportunities for computer scientists to work with people in economics and in social media. And so I looked around and I said, where is a real center of economics, for example? And in Cambridge, Massachusetts, we have MIT and we have Harvard

and we have the National Bureau of Economic Research. So I thought if we could be in Cambridge and if we could be interacting day in and day out with some of the top economists in the country, what could we do? Similarly in social media, the Media Lab is at MIT and there are fantastic people in many of the social sciences in the Boston area. I mean, you know, Boston, Cambridge has over 50 universities. And so I just thought there was great potential there, really fertile ground. And I thought it was also an opportunity for Microsoft to be in on the ground floor as these

interdisciplinary fields started to develop. So that was the impetus for pitching the New England Lab. And since then, what we've done at the boundary of AI and economics or AI and social media, AI and other social sciences, AI and biology, has been incredibly exciting, more than I ever anticipated. Wow. So then talk a bit about New York City and Montreal. What was the story behind those coming into the MSR fold? So for the New York City lab, there was an amazing group of researchers in New York City, fantastic group of researchers. They were very focused on data science in the early days of data science, bringing together economics and social science

with computer science. They're really one of the founding groups in computational social science and in algorithmic economics. So I just was so keen on bringing them into the MSR fold. And just last year, I am so excited about the Montreal Lab. It had come into Microsoft as a startup. It's filled with very young people. They know how to have impact in ways that I never imagined. And I'm learning so much from them. I always bring in the Venn diagram. But how do you see the mix of those two methodologies, quantitative and qualitative, overlapping at this point? So computational social science is precisely the intersection in the Venn diagram of social science and computer science. But what distinguishes it is that over decades, even centuries, social scientists have learned how to ask certain kinds of questions and what are the important factors that determine how people interact with one another. And so I think it's really this

coming together of the kind of people who in the past have done a lot of qualitative work. And by the way, there were, even before computer scientists got involved, there were quantitative social scientists, but the scale of it is so much larger now. Does qualitative research play much of a role in your labs? Qualitative research absolutely plays a role. When I first opened the New England lab, I hired several people who had backgrounds in anthropology and communications precisely because we want to understand what are the right questions to ask. And the qualitative social scientists give us some idea of the shape of this space. So I think we get to much deeper results when the qualitative and the quantitative work together. So we've established that you're not just a managing director of three research labs, but that you're actually still doing research yourself.

Tell us why you think it's important to stay actively involved in your own research, even as you lead, supervise, and direct other researchers in theirs. For me, being actively involved in research is the only way that I can do it. I believe that it keeps me closer to my creative roots and it allows me to connect with my researchers in aated by my research and frustrated by my research. And it allows me to relate to researchers in a way that would be very difficult if I had stopped doing research a decade or two ago. I understand what's necessary to create what I call a fertile environment for research. I think you have to remove the pressure. I think you have to surround people with interesting questions and phenomena, and you have to embrace making progress slowly. If you push people to make progress too quickly, I think they come up with less deep, less transformative, less impactful results. So being a researcher myself allows me to empathize with what my researchers do

and create what for me would be my dream environment if I could spend all my time doing research. Let's talk about your current research then. One of the projects you're working on has to do with machine learning and large networks. Talk about that. What got you interested? What's going on? I actually started working on things that look like networks 30 years ago when I was doing physics. And I would work on these random systems in which, you know, something would percolate through something else. And interestingly, about 15 plus years ago, I was at Microsoft Research Redmond, and my manager at the time, Dan Ling, said to me, you know, I just heard John Hopcroft, who is a legendary computer scientist, he's a Turing Award winner, giving a talk in our China lab about how he wanted to model the internet as a random network. And Dan said, this sounds a lot like what you do.

And so I looked at the text of John's talk and I said, wow, it really does look a lot like what I do. And so we invited John to come visit and Christian Borgs and I, Christian is my principal collaborator and my husband and co-founder of the New England Lab with me. We started working with John and we started modeling the internet and the worldwide web and coming up with algorithms to prevent web spam and all kinds of things like that. And then we took a step back and we said, wow, when we were physicists, we took limits of everything. And so why don't you just take a limit of a network, which sounded like a crazy thing to do. But we started developing that. And I say, we went off into math land for like 10 years. And then I was at the NIPS conference. NIPS is

Neural Information Processing Systems. And somebody came up to me and he said, oh, I'm using your graphons, which were my limits of graphs, to model certain kinds of networks. And I said, are you kidding me? And he said, no, no, no, everybody's doing it. It's the way to do machine learning of large networks because with your limits, we don't overfit. And at first I thought he was kidding. Then I went and looked at his poster and then that just opened up a whole new world for us where we tried to prove theorems about statistical estimation of these large large-scale networks and now it's very very widely used and you know it just shows you that you never know what the applications of something are going to be we went off into math land for 10 years,

and there are whole branches of mathematics now that have to do with this. And yet there are people at all these companies coding up machine learning algorithms for networks based on this framework. So, you know, who knows where things are going to lead. That's true. And what's going on now? I mean, is it an active thread of research for you as we speak? It is a very active thread of research for me as we speak. I'm interacting a lot with the statistics community, which I had never done before. I'd interact with the math community, but not the statistics community. And they are concerned with how do you do consistent estimation of quantities. And networks are totally different from databases, totally different, because you've got all these edges in between entities. even if I know nothing about you, if I know about whom you're connected to on LinkedIn or Facebook, I know a lot about you. And so it's a totally different data structure and there's a different

way in which you need to estimate this. And so we are now doing a lot of work on this. We've worked with fantastic grad students and postdocs, and it's just really fun. And it's so great that people are using it so widely. That's super cool. I mean, it's like discovering somebody likes your song that you wrote. You're singing my song? All right, well, moving on from math, you're also working on a super important project in machine learning for cancer immunotherapy. Tell us about that. Who are you working with? What's new and what's hopeful about this line of research in machine learning. So cancer immunotherapy is, I think, the future of cancer therapy.

What it is, is it's enlisting your own immune system to go after cancer cells. The way we go after cancer cells now is we do surgery. We remove cancer, which has a lot of collateral damage. We do chemotherapy, which is throwing poison throughout your body and just hoping that the cancer cells pick it up faster and die faster than your other cells. Or radiation therapy, which is a little more targeted, but still you get all kinds of collateral damage to the surrounding tissue. What cancer immunotherapy does is it takes your immune system and goes after the cancer cell by cell. When we're younger, in most cases, our immune system does go after mutations that produce cancer and just kills off the cells before they can even

take off. As we age, our cancers, our mutations that lead to cancer have ways of deceiving your immune system using the same thing that we use to make sure that we don't have an autoimmune disease. And so they masquerade using that mechanism and cancer immunotherapy goes in and re-identifies these cells as cancer cells and tells your immune system to go after them. Now, the problem is that it very often does not work, or worse yet, it creates an autoimmune disease that can kill you off all by itself. And so what we really want to understand is which cancer immunotherapies might be effective and not toxic for which individuals. And for that, it is a big, big data problem. You have all your clinical data, you have your genomic data, and you have your immune profile, which is the profile of all of your T cells. So this is a very big data problem. And there are two aspects of that we're working on.

One, we're working with Stand Up to Cancer to fund cutting-edge cancer research. Stand Up to Cancer has projects going at 150 universities with about 1,500 researchers. And they came to us a couple of years ago and they said, we really need machine learning. These are interesting problems because in a clinical trial, you might have only 25 or 50 people. So that doesn't sound like big data, but for each person, you have thousands and thousands and thousands of pieces of information. So it's very deep data on very few people. So you really can't use off-the-shelf ML for that. And then in a different project, which we're working on with people in MSR Next, we are trying to map out what the interaction is of the protein fragments that identify potential cancer cells with T cells. And so that's a huge matrix completion problem and it's super exciting.

So if you think that you could actually make a difference, that you could come up with a therapy which could take what was a death sentence and allow someone to have a long-term healthy life, you know, it kind of dwarfs everything else I work on at times. Finally, there's a third major project on your research dance card, so to speak, and it has to do with fate, which some other people on this podcast have explained as fairness, accountability, transparency, and ethics. Talk about what you're doing in this area and why it's important to you. Well, first of all, I want to just say that in all three of my research labs, I have incredible people working on this and really some of the founders of the field of faith. So I am inspired by them. And just recently, I began working on this. Let me give you one example of

what we've done. When I try to bring in a more diverse workforce, for example, I have learned that I should broaden the space over which I'm looking. So instead of saying, I want to hire somebody in machine learning, if I also look in information retrieval and in information systems, there tend to be a lot more women in those areas. These are areas which are very close to machine learning. People move from one field to another. So if I say I'm looking for someone in machine learning or information systems or information retrieval, I'm much more likely to find a woman or minority than if I just look for machine learning. And so what I worked on with an amazing intern and some other researchers in the lab is something which we call greenlining. We let the machine learning look in the space of all the attributes, and we try to find

alternative criteria which will give us a much more diverse outcome. So for example, if we look in the space we have of data on students, I could make my criterion based on an SAT or an ACT score, or I could base it on class rank. If I assume that every population independent of their race or their economic background is equally intelligent, it's just that some people have received more training than others, then the top 10% of a school in an area which doesn't have as much money should inherently be as good as the top 10% of a school in a very wealthy area. So what our algorithm does is it goes in and it tries to find different directions of the space where if you use that criteria, you get people or results which are just as effective, but much more diverse. And we've done that in several contexts, and we call it greenlining because it does

the opposite of what redlining does. Redlining is a way in which you don't violate some protected attribute, and yet you manage to exclude people. So let's say I'm not allowed to use race in making a decision on a loan, but I am allowed to use income. And income and race are highly correlated in certain areas. So I say, oh, I'm not basing my decision to not give this person a mortgage on their race. I'm basing it on their zip code. And so redlining is a way of excluding people using other criteria. And what we have come up with is algorithms for being more inclusive by using different criteria. Let's talk for a minute about your research philosophy. There are many models and many metaphors to describe them along the spectrum from pure to applied research. Where do

you fall on that spectrum? What's your vision for the kinds of people in research you want to foster? So I am very much a believer in fundamental research, even though it seems like people are just doing kind of crazy blue sky research. Some of that blue sky research turns out to be the PageRank algorithm for Google or other things, which have brought tremendous wealth to the country. And we have VCs who come in, look over what's being done at universities and help to commercialize part of it. So that's one metaphor for different kinds of research. I also like a more creative metaphor, which is let a thousand flowers bloom. So I believe you hire really smart people in areas that are fertile ground. And then I let a thousand flowers bloom. They all go and they follow their passion and they do their research.

And, you know, a percentage of it doesn't result in anything practical, but there are gems in there. There is what I call the long tail of research that if I were to decide in advance, I would cut off that long tail. If I said there are only 10 projects my labs are going to work on, then, you know, the one that's really valuable would just get excluded because it sounded too crazy. So I let a thousand flowers bloom. I look over them. I'm the VC here. And, you know, and I say, wow, that one looks like it might really pay off. Let me put some more water on that one.

The rest of you just keep going and follow your instincts and be creative because we never know which one is going to result in something positive. So I feel that I have been able to more than justify the investment that Microsoft has made in my labs. There are several projects that are just phenomenal that came out of an individual working by himself or herself, doing something that was so crazy that they didn't even tell me about it for the better part of a year, which is fine with me. I don't look over people's shoulders. And probably I don't hear about all the times these things don't work. But if something does seem to work, then they come to me and we try to

find the right match and we try to resource it. And so that is very much my research philosophy. And I think it is based on the research philosophy that has led to all the innovation in the United States over the years. And I think it just makes for happier people and in the end, much more impactful research. Yeah. And you know, I've had several of your people on this podcast, one of whom just recently described your research philosophy as a portfolio and you have different stocks in your portfolio. And at some point that stock may rise and you have the lion's share of it. So I remember when Nathan Myhrvold first hired me, he said, research is a portfolio. And I came

in and started a very theoretical group 21 years ago. And he said, and you're my junk bonds. That's awesome. So now I have a lot more resources and I have an entire portfolio myself. So I've got my junk bonds and I've got my stodgy dividend paying stocks. But yes, I do believe that a portfolio makes sense. And it makes sense, especially to a company like Microsoft, because we have so many people doing development that if we have research just mirror development, we're not balancing our overall Microsoft portfolio correctly. You spend a good part of your time

recruiting talented people to come work for you. And there's a lot of other places they could go. So I imagine you have a pretty tight value proposition for Microsoft Research. What is it? So I think that there are so many opportunities now for most of the people we hire at Microsoft Research. There are many other companies, there are universities, there's doing startups. And I believe the balance we have

here of being able to do fundamental curiosity-driven research and yet also have a tremendous impact if it turns out that what you do can scale to something that can be used by 100 million or a billion people. We're really at what I find is a wonderful sweet spot. We have most of the advantages of a university. And then we also have the ability to hear about problems that the real world brings to us. We can be on the cutting edge. And if we come up with something, we can have tremendous impact. As I recruit people, I listen to them and I hear what's important to them. And I don't even try to recruit them unless I see a good value proposition for them. I want everybody who comes into my labs to thrive. Given the breadth and scope of the work you supervise and do yourself, Jennifer, is there anything that concerns you or keeps you up at night?

So I do worry at times whether technology is being developed in a responsible fashion. There's so much promise, and yet there are potential pitfalls. We have to think about the new jobs that will be created as some of the older occupations are automated. We need to think about how to value people's labor properly, because there are certain types of labor that are valued so highly, others that are just as important but not valued as highly. I also am really concerned about whether, as we try to optimize in machine learning, we don't

inadvertently increase the bias in our data. The bias in our data is already profound. And by optimizing without thinking about fairness, we will compound that. So I really want to make sure that as AI is moving forward at this incredible pace, that we really bake in the considerations of fairness, accountability, and transparency in the work that we're doing and not just try to undo a mess later. So your path to Microsoft Research was not entirely conventional, and I am a big fan of unconventional stories because I think it gives hope to other people who might consider themselves unconventional and maybe then not really suitable for a career in high tech. Could you tell us your story? Okay, well, I've got a long and winding story.

Actually, let me start way back. I dropped out of high school, which is not the profile you see usually for someone in my position. So when I go and I talk to adolescent girls, I always point out to them that the same instinct that made me a rebel is also the instinct that, well, it still makes me a rebel today, but I work within the system. People who work with me will tell you I'm a little bit of a rebel, but it's okay because I'm also delivering value. So then when I went off to college, I managed to get into college because I got really good board

scores. And then I was first in my class in college and I thought that I wanted to go to med school. But then I took a physics class, totally fell in love with physics and went to physics grad school and became a mathematical physicist, which is the most irrelevant kind of work that you could do. And so I was doing very esoteric mathematics, proving theorems about models in physics. You know, I did well in mathematical physics. I was tenured at the age of 30 and I was blazing ahead. One of my classmates was Nathan Myhrvold, who was the first CTO of Microsoft. And I ran into Nathan when I was 39 or 40. And so Nathan said, you have to come to Microsoft. I'm like, are you crazy? I mean, you know, I took one coding class in college, Fortran and Pascal, and I don't do computer

science. And Nathan said, just don't worry about it. I want you to come and interview. And I went and I interviewed and I just met all of these incredibly creative people and was told that I could have freedom to build something. And so Christian and I left our academic positions and we went to Microsoft 21 years ago. And at first I really knew nothing. I didn't know the acronyms. I knew no computer science. When I first met Bill Gates, a few months after I came to Microsoft, I was asked to give him a talk and I didn't know how to use PowerPoint.

So I did it with handwritten overhead transparencies. They told me it was the first time an overhead projector had ever been in Bill's conference room. And when I met him, I said, Bill, I want to congratulate you for hiring a group that won't pay off for 100 years. And Nathan told me to shut up. And Bill said, no, no, it's fine. There aren't enough of you to worry about. And then being immersed in this environment and learning things like that the internet and the world wide web were incredibly well modeled by the kinds of things I had been using in mathematical physics. I mean, who would have thought, right? You know, all of a sudden the stuff we started doing

became more and more relevant. And one of the first people we hired was a Fields Medal-winning topologist. It's Michael Friedman, who now leads a huge quantum computing effort at Microsoft. So I think if you've got talented, passionate people and you expose them to some of the biggest problems in the world, some of the biggest opportunities, stuff happens.

As we close, what's the most exciting challenge or set of challenges that you see on the horizon for your labs right now? And what advice would you give emerging researchers who might be considering their next steps? I do believe that AI and machine learning are going to change the world. I really, really do. I also believe that we really have not begun to scratch the surface with AI. So there are so many opportunities. There are so many societal problems that we will be able to approach for the first time with AI as it develops. I think there's also a tension between doing fundamental research and doing applications. And I think that as researchers, we live with that tension. I think we have to follow our passion and we have to do things that are a little crazy, but we have an instinct in that tension. I think we have to follow our passion and we have to do things that are

a little crazy, but we have an instinct in that direction. That's really, you know, my experience over years now, both in my own research and in the research of people in my labs has taught me follow those instincts. So there's that. And then there's the tension of finding the time to work, to deploy what you've done, to work with our amazing product groups, to deploy what we've done on a large scale so that it can have impact sooner. So I think that that's the tension, that's the excitement. And the best place to do it is someplace with fertile ground. And, you know, that's my goal, to create research environments that are fertile ground.

The gardening metaphor closes the show. Jennifer Chase, it's truly been a pleasure. Your passion is infectious. Thank you for joining us today. Thank you so much. I really enjoyed it. To learn more about Dr. Jennifer Chayes and the Microsoft research efforts in New England, New York, and Montreal,

visit Microsoft.com slash research.