Hi everyone, thanks for joining the master class. It's a pleasure to be here with you to continue the discussion. I'll start by sharing a little bit more information where I left it off in the talk right before, and I assume most of you were with us, so it's going to be sort of jumping right in and continuing the discussion on that. And after that, I'll talk for about 10, 12 minutes or so, and after that I'll be happy to take your questions, so please share them in the chat, and hopefully we have a bit of a lively discussion around this topic today. And so, in terms of where I left it off, really, you know, I was discussing this interaction between the theory and the practice of applying AI, and more specifically reinforcement learning, towards healthcare, and highlighting that a lot of different levels of responsibility that we need to carry on. And in many ways, as researchers, we tend to show up with these problems with sort of our toolkit of engineering and scientific solutions, but there's really a need to think much more broadly. So to engage in that conversation, specifically for the case of healthcare systems, I think it's important to keep in mind the different levels of autonomy at which we inject our AI system. The first level is perhaps the really lighter touch version of this, where we may have an AI system that is used to monitor and enhance human perception, so really just feeding in information that can complement the human's ability to understand the situation. Level two, a little bit more involved, we rely on the AI systems to make predictions and diagnosis, so the system is using that information, digesting it, and using that to draw some conclusions and recommendations. Step number three, the system is going that additional length of providing treatment recommendations, so we assume the AI system has digested the information, come up with a prediction or diagnosis, is now using that to recommend the treatment. And level number four, the AI system is going ahead directly and applying the treatment. And so the example I shared with you in terms of the neurostimulation device is really at level four. In this case, the system is applying the treatment directly. There's no human in the loop deciding whether or not the electrical stimulation will be applied. And there's very good reason in a system like that to let the AI system do the intervention, in particular because the timing requirements of such a system are such that a human can't be in the loop when the decision is happening on a very, very tight time scale. And so really what's important to keep in mind is that the level of responsibility that we consider has to match the autonomy in the system. At level one, where the human maintains a lot of the autonomy, a lot of the responsibility lies on the human. But at level four, when the AI system carries most of the decisions and holds most of the autonomy, then the level of responsibility has to be equally high. Thinking a little bit more broadly, right? There's really this sense that, of course, we want AI to improve people's lives, but we have to also be very sensitive to how these benefits are realized for the population. And so really this notion that for who and at what cost are we going to achieve all of these promises of AI. I take this opportunity to draw your attention to some really important literature of the last few years on fairness and bias that are injected into AI systems. We hear every day of one of these new systems promising great, really great achievements, but often some follow-up study, in this case a study showing that though we can identify skin cancer from photographs, in many cases the proportion of the representation of our population in this data is really, really lacking. And so in this case, this is a very well-known case of trying to do automated dermatology. You would think that from a set of nearly 130,000 images you have a very, very large range from which to learn, but it turns out in the end, after some scrutiny, that only five percent, even less than five percent, of these images are from people who have darker skin tones. And so that means obviously that the quality of the recognition in terms of identifying the skin cancer for these people is going to be much less good because there was less data to train the system. And so that's really an important thing to keep in mind when you develop these systems, right? What is the representation in the data that you're using to train your algorithm, and who is going to be using this system over time? Another salient result on the literature on fairness in AI has shown, in this case it's not for the dermatology application but more for facial recognition systems, has shown that a system that was widely used, in fact, and touted to have excellent performance in terms of facial analysis, had an error rate of only 0.8 percent for light-skinned men, but had an error rate upward of 34 for dark-skinned women. So as these systems are used in practice, whether it's in healthcare, whether it's other spheres of our lives in our society, if we cannot show that we have equitable performance across different populations, we're really not living up to the full responsibility of building these systems. Going back to the medical side, one very common data set that researchers are using, including for COVID-19 patients, is this UK Biobank. UK Biobank was opened in 2012. Researchers around the world can actually go in and get information. There's data from hundreds of thousands of individuals, very complete data. It's perhaps the largest biobank, and so people as a result are using this to develop all sorts of AI-based systems. The reality when you dig into this particular data set is that 94% of participants are of white ethnicity, which is not representative of the general population of the UK, let alone the general population of the world. Yet researchers across the world are using this data set on which to build their AI system. The data set also tends to have a population that is older, more towards the female gender compared to the standard population. And so if you build your system based on this data set, though it's very conveniently available, but then use those systems on the general population, you're really going to have a gap in terms of the performance of the system. And so I'm pulling out these as examples really to raise the awareness of bias in developing these systems. It is important to understand that in many ways this notion of bias is not uniquely an AI issue. The notion of bias in healthcare and in developing technology solutions in other fields as well has a long history. And so we have, you know, our responsibility as we develop AI systems to really interact in a much more thoughtful way with these broader issues. It's not always easy as engineers, as computer scientists, as people who love tech to dive into these deeper social and individual issues, but as you move from theory to practice, there is no way to get around it, and we really have to step up to this broader responsibility, often do it in partnership with people who have the right toolset to do it. And so why I'm going to, you know, really close this before opening up the conversation and the Q&A portion for today is really on the left side I've put in a set of characteristics which, you know, as engineers, as tech developers, we often think about. Most of these have to do with properties of a system. We want to have a system that has high performance, that is efficient. Specifically for machine learning, we often really look at the trade-off between these two things, right? How much can we drive up the performance in terms of having very few errors, having a really good prediction, and how can we do that in a way that's very efficient, so with relatively little data, little computation? Trade-off between performance and efficiency. Sometimes we broaden up, and especially for those of us who build AI systems that are deployed in the real world, we will of course think of notions of robustness, of stability, of speed, so on. All of that is really couched in engineering and scientific terms. What I'm calling on us to do in a much more systematic way is to incorporate considerations based on social responsibility in the design of these systems. And so it's important to be very mindful of the safety aspect. What are the safety guarantees of these systems? What are the security aspects of these systems? Can someone target it for misuse, or actually do we have high guarantees in terms of the security of the system? Fairness, I've talked quite a bit about. I hope I have provided some useful information in terms of making people more aware of how large these discrepancies can be. Privacy is one that we talk about in an increasing way, and especially as we target healthcare applications, of course, the integrity of one's own personal data is top of mind as we develop these systems. Inclusion, in terms of bringing in everyone into the development of these systems, is absolutely crucial, and doing so will ensure that different populations have equal access to the development of the technology. The next, transparency and control, often go hand in hand, right? Transparency is about being really thoughtful about how can humans understand the decisions and the predictions of a system, and how do we allow humans to have some level of control over those decisions. And so without offering the transparency, it's very hard for people to have that control over the system. And so to be really thoughtful about enabling these two properties is something that I think we have to be much more, have much more intent from the early design phases on than we have typically in AI systems. Last but not least, sustainability, I think, is top of mind in terms of ensuring the long-term viability of this approach. And we know that machine learning algorithms are incredibly guilty of using large amounts of computation, and so being thoughtful about developing systems with a mindset of sustainability, in particular with respect to energy consumption, I think is really quite important and falls on all of us. And so just to conclude, I want to say thank you very much for joining me, and we have some time together. I will be taking some questions directly from the chat stream. If some of you want to start putting them in, I have a couple people helping me out which will pin questions sort of one at a time, and we'll try to get through as many of them as possible. But thank you very much for joining us. Question number one is whether there has been any progress made in using generative adversarial networks, so these GANs, to generate simulated MRI imaging that includes not only anatomical but also functional aspects like physiology and stimulus. And so for those who may be a little bit less familiar, these GAN systems are the ones that are being used. In my main talk, I showed these images of the two people which were completely synthetic. This is the type of images that we produce nowadays with these GAN systems. I'll be honest, I'm not aware of any work that uses GANs specifically for the MRI imaging. In our work, we use the U-Net, and I shouldn't say our work, this is colleagues of mine at Facebook. I was not directly involved in the technical work, only as really a sponsor of the project. But the type of architecture that was used for that is a U-Net, which sort of allows to process the data at different levels of resolution. And we used a similar U-Net for the COVID-19 X-ray. So that's really the type of network that is used more than GANs nowadays. It will be very interesting to see if we can inject functional aspects into GANs. The place where I've seen a little bit more of that is on the analysis of video data rather than on the analysis of imaging. And so stay tuned to see where the literature goes in that direction. Question number two, on the issue of representativeness, how much variability is enough? Do we really have a good handle on what we should be targeting to ensure that there isn't bias? And do you have any recommendations? There's the issue here, it's interesting because this touches both on the issue of fairness, it also touches on the issue of inclusion, which is maybe I'll go back to this, to the slides for those of you who are following along, right? The issue of fairness is to avoid bias, but the issue of inclusion is to provide equal access. And the reason I tease them apart is because we have different commitments in terms of doing that, but also different strategies to ensure this. Specifically on fairness and avoiding bias, I think the most important principle is to assume that the representation of the data that is used to train the system matches the representation of the data that is used to use the system, and that there's a notion of equity in terms of who uses the system. So it doesn't mean that the system has to make the same prediction for everyone, but that the impact of those predictions on the different people, on the different communities, is equitable. And so there's sort of different ways to think of how to build that in our system. One of them is to ensure that the representation on the data itself is consistent. One of them is in terms of ensuring that the impact of the prediction is going to be equitable between the different populations. And so specifically with respect to inclusion, now we're really broadening the question and saying, is there an equal access to this type of technology by different populations? We are certainly very preoccupied with that with the deployment of AI-based systems in healthcare. It is not the case that, you know, every healthcare system has the same level of accessibility to the newest technology. And so that's really a question that we have to tackle from a really much, much more broad partnership, not only, of course, the people who develop the technology, but the people who offer the services and who reach out to the populations. There's a question that says, in the previous slide, I mentioned that the biased algorithm was considered very efficient. How was it measured? Why was bias ignored initially? And so on. There's a few cases where that is, you know, the notion of efficiency takes on different meanings. But more specifically, efficiency often deals with data efficiency, meaning can we learn from a relatively small amount of data? And so what happens in practice is if your data is very homogeneous, so your data all looks very similar, it's easier to learn, and you can learn from less data to do that, right? And so the efficiency cost comes trying to really broaden the task that you're trying to do. So if I'm trying to make some predictions, for example, in terms of my epilepsy population, with respect to just one particular model, for example, you know, I was working in an in vitro model with these slices of rat brains. All the seizures look really, very, very similar from one slice to another, one subject to another. I don't need a lot of data. I can learn that from, in this case, you know, eight slices. As I move to the general population, try to do this with humans, the type of seizures we observe from people is much bigger variance. So from one individual to another, the seizures look different in the EEG. Even from one individual, from one seizure to the other, the signature of the EEG changes drastically. And so really there's a sense that you're going to need a lot more data to have good performance over the full variability. So it's really in that sense that as we want to remove bias, have more fairness, have more inclusion, we're going to have to accept an efficiency cost for doing that. And that's where we have to really take on this broader set of considerations when designing our systems. And that is true when you're dealing with images. That's also true when you're dealing with EEG data. It's true when you're dealing with all sorts of other clinical data. And it is also true outside of healthcare. We've seen all sorts of cases where the representation of the system was not adequate for all populations, and there was really this trade-off between the efficiency and the fairness. Thanks for all the good questions. I'm going to move to the next one. We have a question from Chris Brucker, in this case, is asking: since health data is hard to get to train AI models, what would be a good incentive environment for people to share their health data? I think one of the things that makes health data hard to get is because we do actually worry about some of these elements of social responsibility from day one, and in particular we worry about privacy, and sometimes we worry about safety and security. And so once you want to ensure a certain level of privacy in your data, it means that you have to add friction in terms of how the data is shared. So if I'm going to build this system that just does prediction based on, you know, like detecting images of cats and dogs and so on, my privacy constraints on that are going to be pretty low. Most people are happy to share pictures of their pets on social media and every platform they can get without too much concern for the privacy of their animals. When it comes to your healthcare data, and the more detailed we get about healthcare data, you know, it's one thing to have a picture of perhaps your, you know, dermatology, your skin patch, that's one thing, but when it comes to sharing all of your lab tests, when it comes to sharing your genetic information, it gets really much, much more personal. So we have a higher standard in terms of privacy. One of the things that's really starting to be put in place is to have some networks through which we can share the data of a particular type. Through COVID-19, we've seen a number of these networks come together quite quickly, and that allow common sharing of the data. It is true that, you know, my data by itself isn't worth a whole lot, but my data as part of a larger data set is worth a lot more for me and for everyone else. And so building up the right structures that ensure high privacy but also allow us to get the knowledge that comes from joining all of this data together is something that we're going to see more and more through consortiums. There's a really interesting notion of data trusts that is coming through, and that's coming much more from the legal and the policy side. The idea would be that you entrust your data to a data trust, and the trust makes these decisions for the group. And that's one of the models that people are thinking about in terms of building bigger coalitions of data, but doing it in a way that's respectful of individuals. So as an individual, I can decide which trust I share my data with, and maybe different ones of them are taking on different strategies, but I still get the power of the group in terms of knowledge sharing and so on. What's great about these questions is they touch profoundly on some, you know, scientific engineering issues, but they also touch really profoundly on some social, legal, ethical questions. So for researchers, we're brought into this really multidisciplinary world, and we have to partner with experts across disciplines in order to make progress. Moving to the next question, Jean-Jacques is with us today, asking whether the scientific and the social responsibilities are in tension or are they mutually reinforcing? And I think that's an interesting one. I think we can see it either way, to be totally honest. I think from a very engineering mindset, often the more constraints or design property I add into my system, I'm adding some tension because if I think of, you know, like my solution space being a blank canvas, every time I add a particular constraint, such as I want my data to be private, I want to be fair with respect to some group, I want to allow transparency of my solution, it's like adding a constraint on my solution space, off of parts of my canvas every time I do this. So in that sense, you can think that there's really a tension between them. At the same time, honestly, as a researcher, sometimes constraints are creative fuel. By adding these constraints, it forces us to expand our solution space and break out of the old mold. So I don't see these, I don't want to see these social responsibilities only as constraints and tension on my ability to deploy technology solutions. I think we have to welcome them as sort of fuel to force us to find better solutions and break out of a very narrow mindset. And in doing that, I think we can embrace the full potential of these technologies, but it's not easy to do on a daily basis. Moving on to the next question, we have Lena Maria Daramilo, who is asking us, how do you do QA on a system that has high responsibility regarding people's health? So I'm reading a little bit into this question here in terms of QA meaning quality assurance. So we could go into like question and answer, but I think you mean how do we do quality assurance on these systems? Really, it's about, for most of us, it's about having a high standard of the performance of the system. And on my slide here, which is a little bit misleading, I've really identified performance as maximizing outcome and measuring errors, but I think we have to broaden to all the other criteria that I've shared on the slides and really embrace all of them within this framework of quality assurance. Of the properties that I've listed here, some of them we know how to test for, we know how to give formal properties about. So if I talk about properties about privacy, for example, I have in my toolbox a mathematical language that can be transposed to conceptual and eventually perhaps the legal concepts of how to ensure privacy. We have a whole set of privacy-preserving machine learning techniques. So for some of them, we're able to give formal guarantees about the properties of the system. For other ones, we're not there yet. We don't really have a language in terms of fairness, inclusion, transparency that gives us the ability to give the right level of guarantees right now on a system. Even on safety, we're starting to have a little bit more of a formal language with respect to safety, but it's really not that common. So we have to keep on working both on developing the right formalism, like the standards by which we're going to offer these guarantees, and then meeting these guarantees. And this is work that we do hand in hand, often in research, but in doing that, it's useful to have a large set of partners to help us do that and to kind of, you know, give us some feedback on whether we're asking for the right criteria and whether we're meeting that criteria. You shouldn't necessarily have, you know, both of these things done by the same party. I will use this as an opportunity to open up one door which I haven't mentioned before, which is the fact that a lot of the AI research today is done in an open science model. So that means, you know, all of the work that I've shared with you today, but essentially all the work that I do in my own lab, both at McGill but also at Facebook, is work that is shared openly with the community. We publish papers, we open source the code, we talk about the research that we do in a very, very open manner. It is part of the culture of the AI research community at this time, very, very strongly so. And I think that's one of the elements that's going to enable us to make progress on all the properties I've listed here. Really, this commitment to very strong transparency is part of the ingredient to allow us to make progress together on these questions. So I'm happy to continue to support that movement towards open science, and I'm really quite proud that our community is leading the way in that respect. I think we have time maybe for one or two more questions before we unfortunately have to close the session today. I'm going to go to Nina Amiri's question: How do you address the differences between the data from different centers, hospitals, and machines? How do you make sure one model works well in different settings? That's a really challenging thing to do still today. I would say I've seen two ways to handle this problem. One of them is by early standardization. So this is something, for example, in the UK Biobank, right? There's a group of people that decided together what was going to be the format of the data to participate in this biobank, and then any hospital center who wants to contribute data has to commit to that format. So that format's decided up front, and there's buy-in from all the different hospitals and centers to adhere to that format. That's one way to do it. The other way I've seen it done, and this is what we did with our COVID-19 project, which I talked about, is that you may have different data sets from different hospitals. You use one hospital to develop the data from one hospital to develop the system to build the prediction, and then you take that and you validate it on data from another hospital. So in that case, you don't need to really mix the data. It has stronger properties in terms of privacy. What you share between the two is really just the code, and it allows you to do a reproducibility study of your results. So develop in one hospital setting, test in one or many other settings. The advantage of this is that you have a much higher confidence in terms of the robustness of generalization of your system. So in some cases, and usually we do it because of privacy constraints, but that's one way we've done it. So, and you know, on each end, you can sort of massage the data set in the format that you need for your model, but you can do that independently rather than really, really merge the data. Going forward, you know, I think we're going to see more consortium of the first type where there's a prescribed format and everyone can feed data into it, and we are seeing some of that even on COVID-19 across international agreements. In our case, we wanted to get the project off the ground very, very quickly, so we adopted the second technique of really working with each hospital as a separate partner and treating the data independently that way. I think we are out of time already. Thank you so much for joining me today and sharing your questions. It's been a pleasure to have this discussion with you, and I hope you have a good rest of your day and enjoy the rest of the event. Thank you.
