About the talk
Finale Doshi-Velez is a Gordon McKay Professor in Computer Science at the Harvard Paulson School of Engineering and Applied Sciences. She completed her MSc from the University of Cambridge as a Marshall Scholar, her PhD from MIT, and her postdoc at Harvard Medical School. Her interests lie at the intersection of machine learning, healthcare, and interpretability.View the profile
Where are second speaker Q&A this morning and especially in policy and talk? And what do you see, is the role of looking at already cooked in a process? Where, where do you see the, what what do you feed all of this article X Adidas doing for us? I think I'm to me until I can read it as extreme High School. I was really interested, also, a lot of educational distance of uncertainty, is how to use that sort of data to try to make better decisions in the future. And so, I think the fact that we do with previous legislation that was
passed for Hospitals in order to try to figure out are there additional space to do even better for her patients than that, we can do right now by natural variation to figure out what works better pull data. Are we just going to end up talking to the dental field of artificial intelligence reimbursement? Manipulate things. But I think if these cases, I'm pretty optimistic that we can go beyond what is currently state-of-the-art educational games early on that by looking at it
and we had human experts to write down what they thought. This like I'm a really good starting. Gauging decision policy would be, but it was trying to beat special energy into this pot system today for me. Oh wait, you do a lot better sometimes, the people not always, but sometimes To look at doing those in a particular context, for particular individuals, we won't have a lot of data about that one person, so we can extrapolate places to me that you found. It looks like there really is a lot of promise to going
beyond the data. For sure, the first question is is committed and released as well. So you're responsible for creating a lot of papers these days. How do you feel about how these estimator for healthcare setting? Is, if you have a particular, do you cost these type of techniques, one of the things that I shouldn't of the new episode of The Office, where we at, Confounding variable exercise. If I've already eat breakfast and all these other factors and if we're not recording, some of those
factors should be just be like, in commencing all of them or do you feel like they're quite similar? That's a great question. I don't think we know yet. I think we have some answers out there. Some efficiency carrot, how do you envision this happening in practice? Like you steal a process that goes through with experts or clinicians to try to figure out like how much? I think this is it seems like there's different levels of sound in the people. I think it's really tricky and practice some existing.
cake places where I feel it for like the letter is in cases where if we can talk to an expert, like a note because for legal reasons, people have to feel When were optimized, multiple objectives, and share a little bit of what your thoughts are, like, how to think about our problems in the context where there's multiple objectives and how do you extract them from the team of decision-makers involved here? It's really important. I'm really interesting. So we've thought some about
things like making sure things are Equitable between here and they're really powerful but they don't ask a question. Exploration is also Fair. Yeah, uh-huh. That's for all of us. Remove them after they're really the question. Is it many times? The healthcare Seascape doesn't really seem for Kobe in a rather. How, how does it affect your policy evaluation and documentation? Yeah, I think this is between the question of model selection. That is a really important
state representation of using to make decisions are sufficient. Sufficient. But I think one thing that might even more exciting is that in many cases we don't listen What's really good decisions and so there's been a different line of work over the last year that's thinking about. Maybe you might be the markovian model is an imperfect. It's not for victim, all the Dynamics. The optimal policy can be represented as between different models and you look at. What is the probability that if used with fail that test? You pick the Sinclair model with a complicated one, is better. How
much as me. Super exciting? Because if that's our ultimate objective optimizing, for maybe, that's the final thing. We should focus on what we're trying to optimize. And then the model selection, part is only a part of that process. In a related question in that was posted is just how do you do your leg at the end of the day? Because there's not really a test sets or it was just a question to me. I feel like I think it's fine to put Falls with her metric or other options into the set of policies was considering
options as possible. I like things like Macon transit to finally decide what it's good directions on a lot of patients. So I think the more we can make those final evaluators trustworthy Pressure. So there's one more question that's asking about how do we think about it like situations, where they extrapolate a little bit from this question or maybe we're in a low-risk wedding in the accent over a, why don't we really care about you really care about improving?
How do we think about, how does it relate to the amount of available also released from and solicitation for individuals lyrics? On the other hand that sometimes fortunately, they are individual. Could help us understand whether we should be certain, it's optimistic estimates versus the conservative case of Alternatives and figuring out which way is to provide information. Another question that was just posted, if I can. What's the processor that you feel? I should be around
by there? Should there be certain criteria, safety around? Yeah, I think both ideally I think we want you and generally be as Grand as we can about the potential benefits. The basis for evidence that nationally right now, if we can provide action stations, You mentioned in the bible, actions, that any particular point in time to decide that if you're taking actions, that are a respected, how we should be thinking about, limiting the scope and potential action. Yeah, and he isn't, she questioned? So she said I I mentioned your
pessimism and has been several other works. Recently, thinking about why we do that. So fruit, It's really because of statistical cases, you might be able to catch scenario is we are because of your knowledge, a knowledge, even a very small number for additional connect to Bare bowls in, making a mistake, like everyone's temperature goes down in response to Tylenol or things like that. We might be able to inject only knowledge about where we should think about.
That's what I would say there. And I think pessimism is is one way to essentially so to try to get us to only evaluate things that we think we just may not be reliable and safe. This is not very reliable. Here is sort of confidence over. These estimates are leaves, keep Seifer. What is the burden of proof? They would like in order to implement things. Either for personal decisions are for cheap. I'm through it. Where are there? Any directions that you're really passionate about that you
take away points, Badger, offline, learning all the time. I think many of us are areas where if you can get better estimators, find that our policies like you really could make a big difference, in a lot of these domains. And yet, there's a huge amount of And so I think I'm any of us fighting those religious sections is incredibly exciting, assurances, can we provide our YouTube channel, so please check it out. And I'll check out her work a little bit of housekeeping.
So, as I mentioned, there's our YouTube channel, has all of the talks out there, the link is just been posted in the announcement, go there. If you can look up there, you can see the schedule for today and the next thing coming up is our workout session. Just going to be gathered up, secondly. I'm in together. Town, in this experiment, is going to get his fun Avatar, peace world, where you can go around from 11:30 till about 11:57 breakout sessions, which are all posted at Agenda
and then it's 1:30 eastern time. It's just a poster that option where you can wander around. Look at posters. And then after that, we'll come back here to go to love all of the speakers and we love all of our Papers and posters. And we also really love all of you are Community Schools with the founder of our success is really measured by the collaboration to come out of us talking together. So when you do it together. Respect each other's ideas and help each other, you know, on our way to improving machine learning for healthcare, I'll see you there.
Buy this talk
Buy this video
Our other topics
With ConferenceCast.tv, you get access to our library of the world's best conference talks.