Finbarr O’Sullivan is a Senior Post-Doctoral researcher and Associate Director in the National Institute for Cellular Biotechnology at Dublin City University. He has research interests in corneal biology and in limbal stem cell culture techniques for corneal epithelial replacement. In conjunction with collaborators in The Royal Victoria Eye & Ear Hospital, Dublin and the Irish Blood Transfusion Service (IBTS) he has developed the technique of using such cultures to treat corneal-limbal epithelial stem deficiency. This technique received regulatory approval in January 2016 and was used on June 2016 in the clinic for the first time.
Douglas Nychka is a statistician who works in applications for the environment. Douglas Nychka is a statistician and data scientist whose areas of research include the theory, computation and application of curve and surface fitting with a focus on geophysical and environmental applications. Currently he is a Professor in the Department of Applied Mathematics and Statistics at the Colorado School of Mines and Senior Scientist Emeritus at the National Center for Atmospheric Research (NCAR), Boulder, Colorado. Before moving to Mines he directed the Institute for Mathematics Applied to Geosciences at NCAR. His current focus in research is the computation of spatial statistics methods for large data sets and the migration of these algorithms into easy to use R packages. He has coauthored more than 100 research articles and with an h-index of 50. He is a Fellow of the American Statistical Association, Fellow of the Institute for Mathematical Statistics and a recipient of the Jerry Sacks Award for interdisciplinary research.
Grace Wahba, 2025 International Prize in Statistics winner
Episode Description
The international prize in statistics is awarded every two years by a collaboration among five leading international statistics organizations: the American Statistical Association, the Institute of Mathematical Statistics, the International Biometric Society, the International Statistical Institute, and the Royal Statistical Society. The prize recognizes a major achievement by an individual or team in the statistics field, particularly an achievement of powerful and original ideas that have led to practical applications and breakthroughs in other disciplines. The International Prize in Statistics for 2025 was announced recently, and the winner is Grace Wahba. This episode of Stats+Stories is all about celebrating her career with her former students, Finbarr O'Sullivan and Douglas Nychka.
+Full Transcript
John Bailer
Foreign the international prize in statistics is awarded every two years by a collaboration among five leading international statistics organizations, the American Statistical Association, Institute of mathematical statistics, international biometric society, International Statistical Institute and the Royal statistical society. The prize recognizes a major achievement by an individual or team in the statistics field, particularly an achievement of powerful and original ideas that have led to practical applications and breakthroughs in other disciplines. The International Prize in statistics for 2025 was announced recently, and the winner is Grace Wahba. Grace Wahba was awarded the 2025 prize in recognition of her groundbreaking work on smoothing splines, which has transformed modern data analysis and machine learning. Her work has seen practical applications in fields ranging from climate science to medical imaging, and has been used to analyze spatial patterns and global temperature data predict disease risk factors and enhance image reconstruction in various medical contexts. Wahba was elected to the United States National Academy of Sciences in 2000 and received an honorary degree, Doctor of Science from the University of Chicago in 2007 Wahba is a fellow of several academic societies, including the American Academy of Arts and Sciences, the American Association for the Advancement of Science, the American Statistical Association and the Institute of mathematical statistics. I'm John Bailer. Stats and stories is a production of the American Statistical Association, as well as Miami University departments of statistics and media, journalism and film. Rosemary Pennington is away International Prize and statistics winner. Grace Wahba also mentored 39 PhD students that have resulted in more than 330 academic descendants, and we're lucky to have two of them joining us today. Our guests today are Dr Doug Nitschke and Dr Finbar O'Sullivan. Nitschke is professor in the Department of Applied Mathematics and Statistics at the Colorado School of Mines, an emeritus senior scientist at the National Center for Atmospheric Research, or NCAR O'Sullivan is Professor and Chair of statistics at University College, cork, UCC, they share an important common denominator receiving their PhDs under the direction of grace, Wahba, Finbar and Doug, it is a delight to welcome you on the podcast.
Douglas Nychka
Thank you. Thank you. So I'm
John Bailer
curious what was your immediate reaction when you heard that grace was named the winner of the 2025, international prize in statistics,
Finbarr O'Sullivan
not surprised. Expect grace to be, you know, all kinds of awards which He has and is well deserved, and, and, and, and, it's great stuff.
Douglas Nychka
Yeah, I thought it was long overdue.
Douglas Nychka
Very appropriate, yeah, and I think you'll hear from our discussion sort of why we think that.
John Bailer
So you heard the as part of the citation she was talking, it was recognition of her, her work and smoothing splines. Could you? Could you give us help paint a picture of what that was and why that was. Was such an impactful proposition and such an impactful development
Finbarr O'Sullivan
when we started out in in statistics, it was at a time when when computers were not used in the way that they're used now, it was at a time when there was no or very little interaction with data on a computer and and so. So a lot of what people were working on at that time were to develop methods that could take advantage of the of the emerging computer power and the emerging ability to interact visually with data, to see signals in data and and to sort of understand patterns that they might be might be expressing. And one of the key insights at that time was that that that you had to have tools that could, that could look at noisy data, plots, histograms, whatever and and be able to discern patterns in those graphics. And it quickly emerged that the procedures that could that sort of could identify a signal, identify some smooth pattern within within a cloud of points, were potentially very useful in that framework. And there were a number of of groups that that were working on this at the time. And Grace, Grace at the University of Wisconsin was, was, was certainly motivated, it seemed like, by by her exposure to to Bayesian. Methods in her in her own graduate work up to that point, or on research up to that point, and identified the role that that prior modeling of a signal could be helpful in in understanding and developing methods for extracting signals from noise data that the data on its own would not be able to recover signals completely. But if you could supplement the information in in a noisy plot with some with some coherent understanding of regularity about the signal you're trying to extract that that could be an important breakthrough in in developing techniques. And I think her development of smoothing splines was very much motivated by regularization and that sort of framework, and she sort of brought that to bear, or brought that out with us in our graduate studies.
Douglas Nychka
Yeah, yeah. Finbar, that was, that was so nice. I want to, I want to give a little bit of pointed remarks about that. So you it's hard to in in envision sort of what fields look like before research developments have happened. And so, you know now smoothing splines no big deal. I mean, it's just a one line command in our but it it's hard to imagine what what the field was like before that, and so I just want to sketch that out for you. So the way, the way people would look at data on a scatter plot was basically run a smoothing operation across it, you know, basically a running average or a kernel. And this worked up to an extent, but as fenbar mentioned, there was no concept of what kind of signal you were trying to extract. There was sort of this feeling that these things sort of worked empirically, and there was some theory as to why they should work, but no real model for the function. And so that's what Grace brought to this. Is she said, Look, we have data. We know that there's sort of uncertainty of the data around this curve or surface that we're going after. And by the way, here's also a model for what we think the curve or surface should look like. And part of the reason this was slow to get adopted, and sort of took a while to merge from sort of more methodological statistics and into practice, was Grace started from the foundational idea that I think this curve or surface lies in a reproducing Colonel Hilbert space. And so that's a big mouthful. And you know, part of Finbar and my apprenticeship under Grace was to become well versed in that. However, the amazing thing about how Grace was able to migrate this is she was able to take these abstract concepts in math and actually apply them to data. And, you know, subsequently, you know, we can say the the rest is history. Once she figured out this foundational insight with reproducing kernels, all these applications were unlocked and and in ways that inspired the statistical community.
Finbarr O'Sullivan
And I think to so to an extent, actually. I mean Wisconsin at the time. I mean the the big personality within the group was George Box and and George box's take on on on robust, on robustness. And which was, which was sort of a kind of a thing that was sort of immediately preceded this wave of interacting graphically with data and so forth. He was very much a kind of a modeler. Had a kind of a modeling Bayesian perspective on on that problem too, and, and I think I kind of like to think I'm not sure whether it's true or not, but I like to think that big Gray's was influenced to some extent by that sort of viewpoint that said that there is a value in sort of trying to put your and trying to put inference into a framework that's that's sort of that explicitly identifies parameters and and sort of tries to grapple with your your understanding of those parameters in some in some kind of probabilistic way, or in some kind of formal way. And I think she incorporated that into her work. But equally, you know, I think she was an environment that also was that was motivated by, by by doing that.
Douglas Nychka
And, yeah, yeah, I think I'm, I'm glad you brought up George, because I think another aspect of in in the background of Grace's career was this idea of talking. Talk to a scientist, find out what they're really excited about in terms of answering questions with data and and sort of let the statistical project, the statistical research, sort of follow what that investigator wants to do with their data. And that's that's a different perspective often than what we do sort of focused within our field, where we basically work on another statistical problem that a statistician has proposed. Great Grace was out there. She was talking to scientists and engineers and figuring out what the emerging data challenges were. That's exactly sort of the way George Box launched his whole professional career and and that was his philosophy.
Finbarr O'Sullivan
And I think we may have come along at a time when Grace started, was starting to do that a little bit more, just trying to think of us there. You know, she had, she had project going on, or in collaboration going on with atmospheric sciences, people, you know, looking at transfer and stuff like that, and looking at the atmosphere and and, and she was all the time, you know, on about, you know, like learning, the Science Learning, the underlying, you know, physics, of those problems exactly, get yourself, letting yourself completely, sort of embedded in the problem to some extent, and and then, sort of, you know, seeing where the statistics may, may, may or may not be important in that, in that field. And so, so that was certainly something that she certainly encouraged in at least the group of students that I remember being like you and I, and people were Miguel and and so forth that were there, that in or around that time, we were all sort of stuck into problems that I think were that were very much applied problems.
Douglas Nychka
Yeah, yeah, I, I, I agree, totally so you were, you were working on atmospheric retrievals, right profiles. I was, I was looking at this funky you could call it a tomography problem, but it was based on pathology samples related to studying liver cancer. Jim wendelberger was going after climate fields, trying to tease climate, you know, climate surfaces, out of out of temperature observations, yeah, and again, coming back to this fascinating transition Grace was coming up through more theoretical multivariate time series methods in numerical analysis, as I said, bringing sort of this functional analytic insight to bear on it, and all of a sudden we have almost like this phase change, where she says, I'm going to look at all these different data problems and find splines in them.
Finbarr O'Sullivan
Yes, yes, yeah, yeah. And at the time, actually, I mean, Grace had just sort of developed together with Svante Walt, you know, an approach to sort of figuring out how to, how to, sort of let the data tell you how much smoothing you should do with price, right? And that that was an important, I think that that sort of was, I think, a real breakthrough for for understanding, for making splines practical at the time, and, and I think that was an innovation with, I think Svante Walt was more of a chemist, and, right? So I think it, you know, it kind of reflected a sort of, kind of a collaboration too. I know she did lots and lots of stuff with cross validation. I think nearly every time I, you know, I would meet with grace, the part of PhD supervision. At some point there would be something about cross validation too, yeah, actually, in fairness, you know that's, that's that ends up being a consistent theme in in a lot of machine learning, too, that people recognize that, you know that you can't, you have to have sort of methods within your your toolbox that can, that's that can sort of force you away from overfitting data and and so, so my cross validation certainly turned up quite important in in long haul.
John Bailer
Yeah, yeah. So, So Doug, you mentioned early when, before we started chatting, is a thing that you did with kids to try to explain to them what smoothing was right. Can you? Can you paint that picture for
Douglas Nychka
us? Sure, sure. And I'm hoping to post sort of a photo of of this on on the splash page of this podcast. So I, I worked at the National Center for Atmospheric Research. We have this, we had this big lobby with a terrace. So flower floor, and there was a, there was a nice grid on it. So I would, I would bring in a class of school kids, and we would say, Okay, this, this floor is actually gigantic graph paper. And what we're going to do is put some axes on it. We're going to use masking tape to mark temperature and annual temperatures for boulder on this on this floor. And so you can, you can imagine all these tape marks there. And I said, Okay, so here's a rope. What I want you to do is put this rope through the points in a way where you feel like it best explains what's going on in terms of temperatures for boulder over the past 100 years. And so you can imagine how this rope, it can it can bend a little bit, but it's not going to fit every point exactly. And basically, the students could play around with saying, Well, should we put a bump in the rope, you know, where the temperatures increase, or should we, you know, pull the rope a little tighter and and sort of ignore that high temperature excursion. It's a it's a great example. The spline part of this is that the rope has some resistance to bend, and you can force it to bend if you want to. Or you can just take the rope at both ends and pull it straight and just make a straight line. And when we were talking about cross validation, that cross validation step is figuring out how much should the rope bend, and it's the school kids get it, and they have a lot of fun. And there's a lot of people sort of wandering into the NCAR lobby that are wondering what the heck we're doing, but
John Bailer
yeah, sounds like a lot of fun.
Douglas Nychka
I also wanted to mention quickly that there was a statistics science interview of grace, and in that we go through a little bit more detail about the whole background of cross validation and how she discovered it, and again, that uncertainty of, how did this actually come about? You know, it didn't, this concept didn't exist before, basically, she and a few other people started proposing it. So you're
John Bailer
listening to stats and stories, and we're talking with Doug nishka and Finbar O'Sullivan about international prize in statistics winner, Grace Wahba these this is, I've been loving these stories about kind of how the application of these smoothing methods have evolved, and the inspiration and the connection to some of the sciences. I'm curious what was, what was Grace like as a teacher and as a research mentor for you.
Finbarr O'Sullivan
So Grace, I think, I think it's true, though, that at the time, Grace used to deliver a course on on smoothing splines basically three or three hours in the week. And it was scheduled, I think, for was it Thursday evening or something like
Douglas Nychka
that? Yeah, it was something brutal.
Finbarr O'Sullivan
And in it we would, we would get a lot of kind of a lot of functional analysis, you know, from a statistical perspective, and and things of that type. But another thing that that sort of, that, that she sort of emphasized to us, was the importance of listening to other kind of mathematical sciences. And one of the things that was that was going on in in Wisconsin at the time was that, you know, splines, and leading light in in splines, there was Schoenberg, and he developed Cardinal splines. And then those guy came after him, Carol DE BOER, who developed B splines, okay, and and Grace was very much encouraging us to, you know, learn about, obviously, what she's teaching about, but also listen to these guys. Go, go, find out about numerical methods from the numerical analysts, because they will have a different perspective on, on on things, and that will help you understand your problem in a different way. And I felt like that was that was a bit unique. You know, a lot of times in in in statistics, you kind of are in graduate school, you kind of tend to think that whatever subject you're actually stuck into at the time is, is your whole world, but there. But there is benefit in in kind of stepping outside that. And she that was one of the things that I really felt like that she encouraged us to do in our graduate education. And yeah, I remember she sent us, I was doing a bit of work, because I worked in a consultancy at the time for a while, and there was a person in there who came in from agricultural sciences, doing measuring hormones in cows and stuff like that and and she said, That's very interesting, David, I told you to tell me, you need to give up that and start concentrating on your what you're meant to be doing somewhere. No, no. She said, You better finish that up. That sounds like it's important to. Have go learn about it. And, and in fairness, you know, it is, it did strike me as being sort of very, sort of, very important to education, to that ability to sort of strike off, follow us, follow us, sort of a connection that you see and and try and learn something from it. And, and and that was, that was the way all the time. I mean, she probably, in my own case, she told me to take courses in atmospheric science too, because she thought that would be, that would be beneficial to the work I was doing, on, on, on temperature profile retrieval, and paid off.
Douglas Nychka
I thank you now. So I just wanted to follow up with a few comments, maybe a little bit more about Grace's personality. So one thing is that grace typically met with us one on one. That was wonderful. And I think part of another model for faculty advising is to have all their students get together in a group each week, and there's sort of a bit of a mosh pit there, but also, also some benefit from the graduate students interacting with each other. However, I don't know, Finbar and I were working on really hard things, and it was nice to have undivided Time of Grace to sort of troubleshoot stuff and and to go go through things. So that's another memory I have. At one point I said, Grace, this must be a lot of work, sort of meeting with each of these students separately. And her reply was, what you're doing is so interesting, it's easy for me just to get excited about all this stuff and get all this energy just to keep things going. So that was and, you know, I find that in in my own advising of PhD students, that the students come in with new stuff, and I just say, Wow. You know, this is so cool, and it's, it's so neat to be able to learn, learn about this, and I can now understand how Grace felt about that. The other thing I wanted to say, I guess this is sort of a little bit outside of the profession. Is Grace was very active as the student advisor for the girls ice skating club at University of Wisconsin. So Grace actually was a talented ice skater, and could do, you know, could could do figure skating. So she would, she would just sort of each, each week, be the faculty advisor. And I think that these women had no idea that, you know, they were being advised by someone who was also going to be a national, a member of the National Academy of Sciences, that all they saw of her was, oh, here's this faculty member who likes to ice skate and is going to teach us how to do jumps and and spins and stuff. So that was sort of amusing. And she had a very vibrant life outside of work too. And occasionally we would see glimpses of that. She once had us out to a dinner for her graduate students. And really nice, you know, we were all all at this fine, fine restaurant. Afterwards, we were all separating to to go back to our to our apartments, Grace, Grace was going on to another party. So she, she had this other net worth, this other social network, and she was her, I think her evening was just starting.
John Bailer
What a marvelous story of an incredibly impactful mentor and model for for kind of your futures, I'm curious if you could, could, could kind of capture the the legacy of her work. What's, what is when you look, when you look at the practice of of the statistics now, and you think about how people are working on our field, what? What are some of the things that you would say, Yeah, this is, this is one of the places where you could find her fingerprint.
Douglas Nychka
I would say, throughout, well, obviously, throughout statistics and machine learning, people mumble things about cross validation, although for machine learners, we would say, Well, you know, there's a training sample and a testing sample. Grace is one of the instigators of that whole idea that you need to withhold data and test out your model on that withheld data so big. The other thing I would say is to think about, think about how your data is related to the model, and then also add a another component that is basically constraining your model in ways that make scientific or engineering sense. And so, you know, the buzzword there would be sort of a penalized likelihood. Or people would say, Well, you know, that's really just a Bayesian model where you're putting some prior on the function. And these are things now that. You know, are basically in statistics DNA.
Finbarr O'Sullivan
Sometimes people think of of sort of that sort of conceptualizing models for signals in a kind of theoretical way. But, you know, Grace has taken it was bit different. I felt that she said she felt that in there are certain circumstances. And I suppose the work in in atmospheric science is a good example where, where would be possible to use historical data to create a base, a valid Bayesian model. And, you know, she, she kind of sort of highlighted that as a sort of a way to sort of rationally think about developing appropriate models for particular circumstances. And that, I think, is that's, that's that I think is fundamental to a lot of of good machine learning algorithms to that, you know, you can't sort of take an off the shelf. You know, scheme modeling, scheme complex. You know way about dealing with complexity in one area, and think that it's going to be the right solution for a different field and and I think that's probably reflected in a lot of a lot of current statistical thinking as well, the idea that if your signal has got has got particular structures in it, you need to somehow try and model and take advantage of those structures, and build those into your complexity, description, into your sort of notion of regularity that you want to sort of bring to bear to help guide you through any kind of Any kind of signal extraction type of process that you might be involved in a I drifted into, into medical imaging, and I think it's, it's certainly within that field too, that that sort of notion, I mean, you know, there is, there are sort of basic things about, about about signals that arise in in medical imaging. So if you're doing something like a CT scan, then you like the attenuation tissue attenuation tissue density must be a physical quantity. It should be positive. And so you need to build in that sort of a constraint into into the process. And then, you know, if you have some kind of contrast going on in a study. Then there are, there are all kinds of constraints that are associated with those that are relevant in in that framework. And I think that's, I think that's kind of a lot of I always feel like that. That's sort of the way Grace would approach. The problem, in some sense, is that that's sort of idea that you sort of need to get down and carefully understand, or try to understand a little bit more of the problem than just a superficial, you know, I'll do this on a Saturday kind of approach, and not so much worse, what the function is about,
Douglas Nychka
yeah, yeah, yeah, that's really well described Finbar and and, you know, for the non technical audience listening to this, I want to contrast that by, say, like a machine learning method that's good at, say, looking at images and determining if there are dogs or cats. And it makes perfect sense that you shouldn't want, wouldn't want to take that application and simply put it into a medical setting and expect it to work or or trust it to work, on trying to do fairly complicated diagnos about about patients, illnesses. So you know, and grace is very, very much on that had had that started that perspective of think about the problem, don't pull something generic off the shelf and expect it to work.
John Bailer
I'm afraid that's off the time we have for this episode of stats and stories. Doug and Finbar. Thank you so much for joining us today.
Douglas Nychka
Thank you. This was really fun.
John Bailer
Yeah, that stats and stories is a partnership between the American Statistical Association and Miami University's departments of statistics and media, journalism and film. You can listen to us on Spotify, Apple podcasts or other places you can find podcasts. If you'd like to share your thoughts on our program. Send your email to stats stories@amstat.org or check us out at stats and stories.net Be sure to listen for future editions of stats and stories where we discuss the statistics behind the stories and the stories behind the statistics you.