Interview with Professor Tamara Broderick, MIT

thank you so much for joining us today for interview series professor broderick so you teach the bayesian modeling inference course 6435 and mit numbers could you talk a little bit about how you got involved teaching the course and what was your favorite part about it yeah um well first of all it's great to be here uh thanks for inviting me for this interview um so i got involved with 6435 vision modeling and inference as you say because i do um a lot of research in this area of machine learning bayesian machine learning so maybe just a little background bayesian machine learning is really good for cases where we really care about things like uncertainty we want to know not just what do we know but you know how well do we know it um it's probably most famously used by nate silver for election predictions um and so i think that's that's how a lot of people might have at least a little familiarity with it although it's certainly used in a lot of other applications and the sciences and technology um so so anyway i'm interested in invasion modeling from a research perspective and so i really wanted to teach a graduate course in it but i had a another big motivation which is if you think about the way that knowledge sort of gets imparted to people um there's sort of a pipeline like people do research and in the course of doing research they learn something and then they turn it into a paper like a research paper and they published it and then once you have enough papers and um you know sort of there's a scientific consensus uh that usually turns into some kind of book and then you can have like a course about that book and um and you know somebody could teach from that book and all the details and it's really interesting because when you're a graduate student and when you're doing research you don't even have to be a graduate student just anybody doing research um you're sort of participating in the first part of this pipeline so you're taking things that you find you're turning them into papers which are further sort of digested down the line and so it's really important to understand you know what these papers are like because i think at when you're early in your career you've really only engaged with textbooks um and where everything has sort of already been digested like you might find a typo in a textbook but there's nothing that's sort of in question whereas um when you're when you're dealing with papers the scientific consensus often hasn't been reached yet um and so you have to read them in this much more critical way you have to really think about you know um what do we know in this paper but what also don't we know um and trying to figure that out is a whole new skill and so in this class sort of we're not only learning about um bayesian modeling and inference like learning from data using these basic tools but we're really learning about how to read papers and how to engage with sort of you know science on the cutting edge and machine learning on the cutting edge and so something i just absolutely love about the course is that it's very heavily discussion based like there's a lot of discussion among the students and the students are like an absolute joy i feel like every time i've taught this course um just the insights that people have and the questions that they have and the discussions that we have just make it like a super fun course i mean it's very challenging in that respect too you know you can't just sort of get up there and give a lecture and and everything is um is like you know decided in advance um but i just think it's like so fun to get that feedback and really like talk through ideas with students and so i really i really like both aspects of the course the material obviously i mean there's so many interesting questions in invasion inference um sort of you know how do we do things fast how do we make it easy for the user sort of a lot of the typical questions that we have in machine learning um how do we set up appropriate models for different types of data problems um but i think also just engaging with the research enterprise and reaching that sort of deeper maturity in understanding what do we know and what don't we know on this sort of level of human knowledge which is cool that's really interesting to hear about how you teach that course so shifting gears here it seems like a lot of excitement about deep learning lately has been focused on supervised learning but many of your efforts are around creating scalable unsupervised learning could you talk a little bit more about your research there yeah so so actually um maybe maybe let me just start by talking about my research and then i'll get into some of those questions um so so as i was sort of intimating a little bit my research often focuses on uncertainty and robustness and that can be bayesian but it doesn't have to be patient and i'll maybe give an example um again i think this is about not just what do we know but how well do we know it so like let's say that i was really interested in um learning about you know some some proposed new system for a plane and its engines um you know it's i don't just want to say like on average this plane will take off like if i say oh it'll take off with 70 percent i would be really upset about that right like that's not like an acceptable um certainty level for an airplane to take off um or you know to have a safe flight with 70 so this is a case where we really care about not just some sort of you know rough prediction or best case prediction but really saying how well you know how how certain are we that something's going to happen you know on the other hand if i'm like just typing a text to a friend i can make some mistakes in that and that doesn't matter so much and so you know a lot of times when we're thinking about learning something in science we want to know not just what is some physical constant but how well do we know that physical constant if we're talking about the social sciences um you know maybe like for instance we have a collaborator who's really interested in micro credit these small loans to individuals and poverty areas and whether they're really helping people um and uh and so there you know if you're saying oh i think this small loan you know helps people by giving them an extra three us dollars purchasing power parity every few weeks well okay is that three us dollars plus or minus you know 10 cents in which case you know at least it's a positive effect and you know that um or is it three us dollars plus 100 us dollars in which case you know maybe it's actually really hurting people and you don't know so you really want to get a sense of that um you know i think also we're really interested or i'm you know i'm really interested in that i run a research lab i'm really interested in sort of how do you draw conclusions so again you know thinking about micro credit or any study that you might do in parts of the world you might ask okay well maybe i learned that microcredit helps people in this particular study um or some other societal intervention you know maybe that was related to drinking water help people in this particular study but is that like a general thing that i can trust you know if i if i ran this in different studies usually usually we're trying to say something very general about like oh this intervention will work for other people will help them um you know with their drinking water or their um you know uh livelihoods or what have you and someone ask sort of how robust are our conclusions you know if they're very sensitive to our assumptions then maybe we don't trust them quite as much so i'm very interested in these kinds of things asianism is one tool for this it's sort of this really nice uh framework uh that lets us build models that can incorporate expert information um and sort of naturally using bayes theorem and language of probability give us really good uncertainties that have nice properties um but this comes up in a lot of other ways too and i'm certainly not you know only doing things in the bayesian framework so a really a really common example for anybody who's sort of familiar with the basics of machine learning is cross-validation this is a way of sort of understanding um you know evaluating what we're doing seeing at least some notion of robustness um so in cross-validation you you sort of leave out little bits of your data points uh or little parts of your data points um at a time and use that um to predict given what you've learned on the rest of your data um and an interesting thing about cross validation is that especially for these modern complex models like you just mentioned deep learning but there are plenty of other very um you know complex and often very slow models that we try to fit um that can be really expensive um your you already have this very expensive uh you know computationally expensive data analysis and now i'm asking you to run it you know maybe tens of times if not hundreds or thousands um and there are reasons to think that the more expensive versions of cross validation are actually better um and you you if you're already at the edge of your computational power that's going to be very difficult to do and so one of we've been working on recently is can we approximate cross validation in a way that's you know still accurate still giving us what we want but very fast and equally easy to use and like conceptually easy and things like that um so um so yes i'm really interested in in understanding these uncertainty of robustness and scalability that certainly takes the form of both supervised and unsupervised learning i think a really interesting thing about deep learning that i think you know um for anybody who's who's in the field to really have a good grasp on is that deep learning is super cool and it's able to solve a lot of really neat new problems but i would consider it a tool in the toolbox um so in machine learning i like to think of us as developing tools for people who have really important cool data problems um and making their lives easier so they don't have to worry about those tools they can just get down to the like awesome science or economics or whatever it is they're doing um and so something that's really interesting is that like deep learning is really good at certain things um but there are other tools that are also really good usually complementary things um so you know something that i i there are a lot of things i like about 6036 the the other class um that i taught last year in which you're familiar with um and one of them is also the really awesome interactions with students but another one um is so there's this project component that some people can take for um graduate credit if they if they do a project in addition to the usual coursework um and something that i find really interesting is in this project a lot of people are really surprised to find that things like linear models can be extremely effective and i think that anybody who works in an applied context um you know if they're if they're doing applied work you know at a technology company um or they're really trying to get an answer about some sort of scientific problem there's there's so much um there's so many times that people will turn to linear models and their generalization generalized linear models or generalized linear mix models because they're so interpretable and they're often actually really good in the types of problems that people have in practice another example is actually random forests um if somebody even even if somebody really just cares about predictive power random forests blow a lot of things out of the water and they're very easy to use um and and this is just like a really small sub sample there are so many other models and things that people use and so i guess i think of it as being you know we're all working on this toolbox um and my research works on you know some corner of the toolbox um deep learning is another corner not that i want to say those are totally separate but i don't do quite as much you know that's related to deep learning um and then uh but then there's so many other corners too like there are kernel methods which are super cool you know they're all these different tools that people are developing interesting hearing about your toolbox analogy i've noticed that some of your work has been used to analyze tumors do you have an application area of your work that most excites or intrigues you yeah so um so first of all that's very cool that you know that um and then i'll also say i think you know just as a precursor one thing that i love about working in machine learning and that's like just a huge plus to me is how much you get to work with people in different application areas and also how much you get to kind of flit around like uh it's really fun to me um you know some people describe it as like you get to play in everybody's backyard um like i've just been so lucky to be able to collaborate with people in physics um in social sciences uh in politics and lots of areas um one application that i'm super excited about right now is actually something that i started kind of ages ago but recently have picked up again with an undergraduate at mit we're doing research together um so uh so basically it's this system that lets people who have severe motor impairments so suppose we're imagining people beyond quadriplegic um you know beyond paraplegic and quadriplegic they have something like locked in syndrome where they can only move a single muscle um or cerebral palsy or you know something something where they have a very very limited range of movement and so we're imagining that um again they can like maybe activate a single muscle and that's it like maybe they don't even have the ability to do something like morse code where they can control activation um so they might have like a sip puff switch or they're blinking or you know what have you um and uh and there's some really you know sort of famous examples of this so i don't know if um you would have ever heard of the book or movie diving bell in the butterfly but the french editor of elle very famously was in a car crash and then suffered from lockdown syndrome um of course there's stephen hawking and you know there's other famous examples so anyway okay so when you get in this situation how do you write how do you do anything and something that's really fascinating is that um you know apparently in this this book movie diving bell in the butterfly um what happened was this guy had this person another person who would read the french alphabet to him on in frequency order but still and he would like blink when they got to the part that he wanted and there's just so much that's problematic about this um so first of all a lot of times you want to talk and somebody isn't there and there have been documented cases of people who they'll be in a wheelchair and their hand will be burned or something because they are you know maybe against a radiator or something and they just want to say something um but if nobody's actively prompting you what do you do and so this seems like a case where you really want things to be automated with a computer um and then uh and then second of all this is a super inefficient way of talking so you know what if you want the letter that's at the end you're going to have to wait through all the other letters to get to it like that's not a really good idea now there are other alternatives nowadays but they're not you know at least they're automated i think that's the huge boon but they still have some inefficiency so very common thing that people do is called the grid and it highlights a row then it highlights a row and highlights a row and you sort of have to go through every row and then every column and if you want something down here it takes a long time to get to um and so anyway we have this method it's actually based on vision imprints um that lets people only look at the thing that they want to on the screen so that's very helpful for people who can't easily change where they're looking on the screen it automatically learns how people click um so if you're you know very accurate in the way that you make your selections like if you have a sip puff switch and you're just super accurate about how you activate that switch then it will go very fast and it'll be very you know you'll just go through and type very quickly um but if you're very inaccurate which a lot of people in this population are going to be very noisy and how they they click um or use their switch then it learns that and it gives them more time and more space and also super efficiently uses their clicks and so they don't have to you know do as many um switch activations so that they can you know select things very easily and then something that's really cool is that it basically relies on having these little clock indicators and anything you might want to choose and so you could put those next to letters and you could put them next to word completions but you could also literally fill this screen with them and then draw by just like connecting points and so then that's actually something that we've done um make a drawing program and it's very efficient you don't have to you know do too many clicks even when the whole screen is full um and so anyway so right now uh what's really cool is you know not only is this a project that like we've been developing and put together um but actually we are collecting data on this so we're actively working with um a charity that works with individuals who have these motor impairments but also we're doing a lot of experiments on able-bodied individuals you know to make sure that you know things are working and um you know make sure that it's like a good experience to use the software um and this is all with this uh super cool uh undergrad that i'm working with um nicholas moniker um and uh yeah it's just i think it's really fun it's like the one thing you know usually we're collaborating with people who have the data who are experts in that area um and that's super fun and i really enjoy that as well but i think this is kind of fun because like we're collecting the data um and that's an interesting process in and of itself um and it's interesting to to see that and do that exciting to hear about all the work you've done now let's talk about your trajectory could you talk about what got you into ml yeah so when i when i was getting started everything was so different than it is now like ml wasn't like this cool thing that everybody knew about like i think like nobody knew about it so it's very different um so basically so what happened to me was i had a few different interests when i was in undergrad i was super interested in astrophysics um and so i was actually doing some research in astrophysics um so there were these cool questions like okay you know there's all this um matter that we can detect in the universe in a particular matter that we can see so all the matter that we can see is sort of light or baryonic matter but then you get these sort of gravitational effects that you can sort of see are happening and so you know people posit there must be some other type of matter let's call it dark matter um and then if you even account for all the dark matter and all the light matter in the universe there still seems to be something else going on and people basically are like oh let's call that dark energy um and so my research back in the day was talking about you know can we characterize dark energy what can we learn about it can we characterize dark matter what can we learn about it um and it was super cool and i loved it and i think astrophysics is super fun but something that came up in the course of doing it was we were we were looking at um things like markup chamber carlow and fisher information and basically all these ideas that are essentially like data science machine learning statistics um and at the time i i felt like i didn't understand these and i was like man i really wish i knew like what exactly these tools were doing i mean i understood them enough to use them but not sort of at a deeper level and i really wanted to know that um at the same time i was taking a lot of math classes and i really liked analysis real analysis and complex analysis and measure theory i was super into measure theory and so um i went to the um undergrad advisor and i said hey you know what can i do with this like can i you know get into measure theory somehow he's like no that's dead you can't do that but you should check out probability that's like it uses measure theory and it's cool too and i was like okay i'll check that out um and then at the same time um we actually had this one like very niche course on machine learning that like nobody like very few people took because again it was not a thing back then and i took it i was like oh this seems really cool i'd like to do something like this but the course that i took like wasn't as mathematical as i wanted i think i wanted something a little bit different um and so um i mean it was great i loved it um but i was like you know still trying to figure out exactly what i wanted to do and so i actually went um to england for a couple of years at this point so i graduated from undergrad i still i still was like man i like math i like physics i like computer science i'm not sure what to do with my life um and so i was like okay i'll go to england for a couple of years so i went on one of these like bring americans to england scholarships um and uh and there i discovered basically that um bayesian machine learning was like exactly what i was looking for like it was it was what we had been using um in some of this astrophysics work it was um you know very deep in terms of the math that you could get into um and it was a super cool data analysis discipline and i was like man this is perfect it's totally what i want to do um and so at that point i applied to go to graduate school for my phd and i was like i want to do a phd in this bayesian machine learning stuff and in particular there was this there's this sub area called bayesian non-parametrics that i was and am really interested in sort of very flexible it lets you learn more as you get more from your data as you get more data um and so uh so i applied for that uh and i got in and then basically like that's how i got into machine learning and then and then at some point during my phd i think is when like machine learning took off as a discipline and it was very popular so it was good timing that's a super interesting background and this leads on to our next question can you talk more about how your experience studying in cambridge has shaped your approach in your research yeah yeah so i mean one one big thing that it shaped was just discovering um bayesian machine learning so i i guess this happened in a few different ways um so one was i was looking for somebody to do research with you know i was sort of asking around like hey would there's anybody who'd be willing to do some research with me um and i found a couple of people who are both fantastic um so i was able to do some work with dave mackay and that's actually where i started uh this this project that i was talking about just a moment ago on this assistive technology um so i started that when i was working in his lab and then i also started working with uh bobby gramacy um who was uh at cambridge he's since moved where he's a professor um but um and we were working on uh gaussian processes projects so gaussian processes are these super flexible supervised learning uh methods and like they're super cool um and i love them both and i was like wow this is really neat i want to do something like this um and uh and that really turned me on you know and and made me understand like that this bayesian stuff was was really cool um i also took um so i was doing uh part three of the math tripos the first year i was there and i took a bunch of statistics courses and i was like oh yeah like this is so great to have like way more data analysis courses than i've been used to um and so that was really neat because again like there wasn't much of a there wasn't like so much of a machine learning curriculum anywhere um at the time at least the places i was at um and so it was really cool to see you know more of these things um and and yeah and so i think certainly all of that led me to to apply to the phd programs that i did and to get into the areas that i did um and i think it was yeah it's super formative also cambridge is just really cool like uh it's you know i mean this is very cliche but it's it's very like harry potter you know like everybody dresses in rose to go to dinner um and they're all these it's really interesting they're all these different colleges in cambridge and it's very sort of very different from the us system or what i had experienced of it because they're all kind of autonomous they all have their own like fancy dress rope dinner thing um and so i would like try to go to all of them like get my friends at different colleges to invite me and um it was it was a really fun time it's like a super beautiful campus too and like you'll just run into cows and stuff that sounds like a wonderful time now let's talk about your experience before coming into college you attended the mit women's technology program in high school which i attended too um could you talk a little bit more about your experience there that's so cool that you went too i love that yeah it was amazing so i was so lucky i was actually in the first year they ever had that um which is like very exciting and it's it's weird to think how long it's been now um but it was amazing i mean i just think what was fantastic about it for me was seeing all of these really cool ideas that i had totally not been exposed to before um so certainly like computer science i think just wasn't a thing in high schools back then and so being exposed to like computer science and the math that goes into computer science was just really great um the projects we had were so fun um i loved them like the the counselors were fantastic um i remember very vividly that we had this we had a couple of big projects at the end so one was um developing uh you know building these lego uh systems to biopsy a grape and jello and you had to like just go into the grape and not pass the grape and i think um i think uh my my pair my my partner and i won like this big comfy mit sweatshirt from it which was fun um and then uh and then actually this is really cool so my partner from my motor experiment which i think is still a thing that like they still make the motors um she kept the motor for years and then i think just within the past year or two she sent it to me and so i like still have our our motor which is amazing um it's just really cool and and on that note something that i think is really cool is i really kept in touch with a lot of people that i met from that program even though it's ages ago and it's really cool how a lot of them have stayed in fact i think all of them have stayed broadly in steam so some of them have gone into the medical profession um my partner from my my motor project i think is now sort of investor in tech um but you know all sort of in that space and of course you know some work at like google and microsoft and stuff like that um but uh but it's just really interesting to see you know how far we've come and what we've done also it's just it was really fun i think it was a little bit of a free-for-all because it was the first year so we like went on all kinds of like crazy adventures and you know they they just let us you know hang loose and do all kinds of fun things but you know i remember you know even the activities were just amazing like we would go like whale watching or go to the boston pops and get dim sum and stuff and it was just like super fun um something i love now is that i get to go and talk to the students um at it over the summer and tell them what it's like to sort of be in computer science and be in machine learning um and the most fun part is so i'll be giving this talk and i'll be like oh yes i'm in machine learning and it's fun and then i'll show them the picture of me and my friends from wcp like we have a picture of like all of us at you know wtv back in the day and i'm like no you were in a wtv that's crazy and it's it's very cute um yeah but i like love going back and it's like such a great crowd shifting gears here so for any students interested in learning more about ml in pursuing work in the field do you have any thoughts on whether students should start on the theory side or begin with an applied project and fill in the gaps or some combo of both yeah i mean i guess i personally feel like you really just can't go wrong by following your interests um so you know there's so much work to be done on both sides so if you're if you're very um applications inclined there's almost no application you can think of right now that would not benefit from ml and so you know you can really reach out i think there's a lot of interest you know even just at professors at mit outside of the computer science department or in different areas of computer science or ecs um you know a lot of interest in working with students and you know doing research with students and um applying those techniques and i think it's just a really useful skill set and so um i think that's fantastic and you really learn about why these techniques matter because i think that's something that is so important to keep in mind is that the point of machine learning is that it's in many ways a service discipline like we're providing tools for other people and so you know getting to see those tools being used up close is a really important thing conversely the the foundations of ml what really makes machine learning work are math um you know it's it's just fundamentally um things like probability linear algebra optimization all of these these tools and so you know having that background can just be so impactful for being able to really understand what's going on with machine learning like what it's doing what it's able to accomplish one thing i will say is i think that it is a really good idea whether you're going to focus on theory which is great or you're going to focus on applications which is great to at some point just work on some kind of applied project like it doesn't have to be your focus but just something to get that experience of working with somebody who really cares about a particular application like somebody who is really like focused on that application so not somebody who um is a machine learning person like me but like somebody who really is like the application person because i think it really gives you a sense of what are we trying to accomplish with machine learning like what is the goal um because even i do like i do plenty of super theoretical papers like papers that are totally probability theory for instance um but even in those i really think the the point of good theory is ultimately downstream to enhance the applications to better you know make people understand what's going on with the applications to make people understand um you know the right way to to use machine learning tools and so on and so having experience of applications is really useful because then you understand what's the important theory and like why does this matter um and so i think i think it's just good to like have some experience in that domain that's some great advice finally do you have any advice for younger students interested in this field that's an interesting question i mean i think one really big thing is to just like follow things that you're passionate about right um i mean i think i think in many ways machine learning is this extremely versatile field like i was just saying you know you really can work with anybody like if you like astrophysics if you like biology if you like chemistry if you like engineering like any type of engineering if you like the social sciences like you can totally do something in any of those fields so i think that's that's you know something to take advantage of if there's something you're really interested in you can do it um if you are interested in an applied field i think it's really worthwhile trying to spend some time actually learning about that field you know like if you're interested in applications in a particular area like really learn that area because they're gonna have a lot of insights and it's that it's i think it's what's really impactful in machine learning is the melding of two different ideas so maybe those ideas are machine learning tools with some kind of application area maybe those ideas are some really cool area of math to develop new machine learning tools but i think it helps to have some diversity in the way that you're thinking and your background and things like that um and so i think that's worth going after um and i guess i would just repeat um you know the idea that i was suggesting before is just trying you know something on a real application um you again even if you're a super theoretical person which i totally respect and i think that that's great um but just getting some sense of working at a real application because i think there are a lot of things that we just don't expect and then in doing so we learn like oh this is what really matters to this person like every time i've ever worked with somebody who is applied in an applied field which i've been very lucky to do i always learn you know i'll say like oh you know how about we just do things this way because that's standard and they're like well actually i don't care about that i i care about this thing over here and then that's an opportunity to develop you know new machine learning methods and you know new things that can help people and so i think that um that's really cool also just taking advantage of collaborations is really cool i mean i think that that's uh like a fun thing to do well thank you so much for taking the time to share your story and advice and thank you all for watching the video

Channel: MITxHarvard Women in AI

Views: 1,569

Keywords: Women, Massachusetts Institute of Technology, Artificial Intelligence, Machine Learning, Education, Technology, Academia, Professor, Computer Science, Bayesian, Statistics, Math, Mathematics, Academic, MIT, MIT CSAIL

Id: T3cOoxspe0k

Length: 30min 36sec (1836 seconds)

Published: Wed Feb 10 2021