5 - Carla Brodley: What role should humans play in machine learning?
Dr. Carla Brodley, Dean of Northeastern’s Khoury College of Computer Sciences, came in to discuss the role human experts should play to create application-driven machine learning. Here's the study Carla did on patients with treatment-resistant epilepsy. Learn more about Northeastern's Experiential Institute for AI or the Align program to increase diversity in computer science.
Read on for the transcript!
Hi, I’m Monte Zweben, CEO of Splice Machine. You’re listening to ML Minutes, where we solve big problems in little time. Every two weeks, we invite a different thought leader to talk about a problem they’re solving with Machine Learning, with an exciting twist: our guest has only one minute to answer each question. Let’s get started!
This episode, our guest is Dr. Carla Brodley, Dean of Northeastern’s Khoury College of Computer Sciences. Carla’s interdisciplinary machine learning research has led to advances not only in computer and information science, but in many other areas including remote sensing, neuroscience, astrophysics, computational biology, and predictive medicine. Welcome, Carla!
Thank you, pleased to be here.
Tell us a little bit about your journey and how you got to where you are now.
So I started out as a freshman at McGill University as an English major, because I love to read. And I loved to read, but I wasn't that good at writing. So I switched to economics. And then one of my friends came home, who was an engineering student, and she threw down a pack of cards, because you know, I'm old. And for those of you who don't know what cards are, they were what you used to write programs on. And she said, you would like computer science. And I thought maybe I would. So I tried Fortran programming, fell in love, and switched my major to math and computer science, and went from a C average to an A average. Then I went and worked for a while, and then I went to grad school thinking I would get a master's in this new thing called AI. And similarly, I took my first machine learning class and that was it. Then I went, became a professor in Electrical and Computer Engineering at Purdue, decided I couldn't live in the Midwest anymore, moved to Boston at Tufts, and then ended up being recruited as dean of Khoury College of Computer Sciences. And here I am.
That's fantastic, what a great journey. But I have one question to ask you as somebody else who's used computer cards: did you ever drop a deck?
I missed them by a year, we got a mainframe.
*laughs* lucky! Okay. As Dean, your work obviously includes research in machine learning. But now you've got an opportunity to shape entire programs in the academic field. So perhaps you can tell us a little bit about why you started the Institute for Experiential AI at Northeastern: what was the problem you were really trying to solve?
So I got the opportunity to pitch an idea for a university-wide Institute to the senior leadership team. And I remember we were all sitting around the table and people were pitching different ideas. And my own research has always been applied machine learning. And also thinking about how do you best use your human expert in the various applications. And so I proposed application-driven human-in-the-loop artificial intelligence. And then this idea really captivated senior leadership. And we renamed it Experiential AI. Northeastern University is really focused on doing what we call use-inspired research, which is solving the problems of today, and working with nonprofits and working with industry to solve important problems. And so this really resonated. And the idea would be not just that we would have AI researchers, but we would have researchers from other disciplines, also learning how to apply AI to their field. And therefore, we could really advance the state of the art across all fields.
That's fantastic. So experiential AI is really all about not just doing machine learning, but having humans and machine learning models interacting with each other making each other better. Can you give an example of human-in-the-loop machine learning that you worked on yourself in your own research?
Okay, so an example comes from a collaboration with neuroradiologists at New York University's epilepsy center within the medical school. So if you have something called treatment resistant epilepsy, that means drugs don't work, and you have a lot of seizures. Well, unbeknownst to, frankly, probably everyone that's listening to this podcast, the treatment for this is to locate the part of the brain that's causing the seizures, and if you don't need it for anything else, basically scooping it out. And the issue here is in locating that part of the brain that gives the seizures. So people get MRIs and an expert neuroradiologist looks at it, and yet, really, they can't see it a lot of the times because treatment resistant epilepsy is often on the cortical surface of the brain and you can't see it even in an MRI. If you don't locate it, then the efficacy of the surgery goes down to 29%. Whereas if you can locate it, it's 66%. Now, these people are desperate, their lives are ruined, they want you to locate it. So what we did is we applied machine learning techniques to figure this out.
That's fantastic. So did the machine learning techniques help the neuroscientists figure out when they should do surgery, and when they shouldn't do surgery?
So the next step on whether you've located the the place that needs to be taken out, or resected, as they say, in the medical world, what they do is they take the patient, and then they do a secondary test where they put them in the ICU, they take their bone off, put it in the fridge, and attach electrodes directly to the brain and induce seizures to get a secondary source. So what we did was we came up with a method to focus that, that then when they put the electrodes on, they could thereby really, really make it more focused. And we were able to see the lesion that was causing the problem in 80% of the patients that had not been able to be seen before. Now, where does the human in the loop come? We don't make the decision about which is the actual lesion, we give the top four to five places that we think it's occurring from, and then the neuroradiologist uses other medical information, such as how the seizure manifests itself, to decide exactly which one of those is the real one.
Oh, thanks a lot. That's a really great example of applying machine learning to healthcare circumstance where the machine learning is being used as an advisor to the human professional, and helping to basically come out with better patient outcomes. Is that right?
Yes, that's correct. And we're also using the expert radiologists to create the training data that we use to train the machine learning method. So we had MRIs of people who had treatment resistant epilepsy, where the surgery had been successful. Remember, 29%, it's successful, even if you can't see it on the on the, on the MRI, and we were comparing their brains, to people who volunteered to have an MRI who did not have epilepsy. And because there's ways now to map this part of my brain onto the same part of your brain, you can see if it's abnormal, so we basically used a technique called anomaly detection, where we look for anomalous parts of people's brains. That's great.
Going back now, from that specific example, of using machine learning to identify portions of the brain and advising a neuroradiologist and looking at the big picture of the Experiential AI Institute. What are you trying to achieve with the institute now, what's driving the new Institute?
Well, we hope to achieve several things. The first is by creating a university that looks at how to apply AI to different disciplines and trains people from other disciplines how to apply AI. And for example, running a postdoc program on experiential AI, where we bring in people who have a PhD in something else. And they spend a two year postdoc with us learning AI and learning how to apply AI to their discipline, it improves the research in that discipline. But even more importantly, or equally importantly, I should say is that whenever you do application-driven artificial intelligence or machine learning, sometimes the stuff that's out there just works. And other times you have to invent something new. And the best research in my personal opinion, and one that I was able to convince my university, is that the best research in machine learning is driven by the fact that there's an application that needs a new machine learning algorithm, and then that's often applicable to other areas as well.
Thank you very much. So the institute is really being driven to do fundamental research, but in the context of real world problems. And in your view, it sounds like that not only helps the world on those problems, but it really helps the research as well. Yes. Okay. So where do you see the institute in five years?
I see us having trained an amazing number of scientists across our university, to really understand how to apply AI, all kinds of new collaborations happening. And one of the really important parts of the Institute is we don't want to do it just at Northeastern, we want to collaborate with hospitals with other nonprofits, and with industry to really get these problems that are so important and hard to solve. And so we're really excited that we're going to have this incredible network of entities that we're working on, both within Northeastern with other universities with other nonprofits and across industry, and it's extremely exciting.
Awesome. That sounds like a great ambition. It sounds like a great program. Outside of the Institute, I love the work that you're doing in bringing diversity to the computer science field. Could you tell us a little bit more about your approach to this?
Honestly, Monte, I loved being a machine learning researcher and I was having the time of my life. And I had always volunteered to do stuff to broaden participation in computer science. And when Northeastern approached me to see if I was interested in being the dean, I was like, This is cool. I'm going to have basically my own experimental testbed to see, can I broaden computer science and make everyone feel welcome. So the mission of Khoury College of Computer Sciences is computer science for everyone. And that's because it touches everyone. And because I want everyone to feel welcome. So the first thing we did was we created a Master's, for people who didn't study computer science. This is the Align Master's in computer science. And people come from English from business. They do two semesters of very tough, undergraduate, ramped up material, and then they join our direct entry master's program. And we currently have 1000 students in this program. And we've scaled it nationally to all of our network campuses. And we've grown a consortium of schools who are offering very similar programs.
That's fantastic. And that's a great program for the master's level. I'm wondering, are you doing similar things to make the undergraduate program more diverse?
Absolutely. So, one of the things that has made our undergraduate program diverse at Northeastern, is combined majors. Some schools call this CS plus x, we actually call them combined. And that's where a student pairs cybersecurity with criminology, or pairs, data science with biology or pairs, computer science with English. And we have, over half of our majors are combined majors, and we have combined majors with 37 different majors on campus still growing. And we have, I believe, this year we have across all of our years, 32% women in this program, and growing.
Thank you very much. That's a great program to diversify further. And I, I think I heard some previous talks you gave where you were really hoping to get more people into computer science of diversity by just getting them to try it. Not having quotas, or lists or percentages or goals or objectives, but just let it happen naturally by getting people to try it.
75% of people who try computer science love it and go on to take a second course. And if you can keep them till the third course, you don't lose them. Mind you, I could put someone on that first course who could kill off anyone's interest in computer science. So the key is putting best teaching talent in the beginning and creating an environment where everyone feels welcome.
Fantastic. Carla, it's been a pleasure. Thank you so much for joining us today.
If you want to hear what Carla's most excited about for the future of machine learning, check out our bonus minutes. They’re linked in the show notes below, and on our website, MLMinutes.com. To stay up-to-date on our upcoming guests and giveaways, you can follow our Twitter and Instagram, @MLMinutes. Our intro music is Funkin' It by the Jazzual Suspects, and our outro music is Last Call by Shiny Objects, both on the Om Records Label. ML Minutes is produced and edited by Morgan Sweeney. I’m your host, Monte Zweben, and this was an ML Minute.