Scientist Stories: Sotirios Tsaftaris, Building Blocks of AI in Healthcare

Good morning, everyone.

Welcome to the Alice Live NCT Data Science Seminar. A warm welcome to all the representatives, to our auditorium, and especially to Sotirios Zapataris, our speaker today. We are really happy to have you here. But before we start, I have to announce that the seminar today will be recorded and the video record will be published later on our YouTube channel. If you have questions, please use the chat function or raise your hand. This can also be done after the talk or during the talk. Now I hand over to Lena Meyer-Hein, the host of today's session. Lena, the stage is yours.

Yeah, it's my great pleasure to have Sotos here today. I, I take the short name, that's easier. So Sotos received a diploma in electrical and computer engineering in Greece and then a PhD and MSc from the US in 2006. '06 and '03 respectively. He's now the Canon Medical Royal Academy of Engineering Research Chair in Healthcare AI and the Chair in Machine Learning and Computer Vision at the University of Edinburgh. And his contributions have been recognized in many ways, actually. He's Turing Fellow with the Alan Turing Institute. He's Ellis Fellow. He is Associate Editor for PAMI. He has served at major medical machine learning conferences, such as CVPI, ICCV, ECCV, MICAI in various roles. I will not list them all, but I want to mention that I know him because from my perspective, he is the go-to expert for disentangled representations in the medical domain. And this is also one reason why we invited him here. And yeah, I'm really looking forward to your talk.

Thank you, Lena, for the kind introduction. I must say that I'm very happy to also I coincidentally have received my Ellis hoodie. Very nice. I love it. I don't wear it today because it's actually really hot in Edinburgh, and that's a very rare thing. It's almost 17 degrees outside. That usually happens in July. So anyway, thank you very much for the kind invitation and Doreen for the help and Lena and your team for setting this up. I hope to do justice to what I'm going to try to describe today. About how I believe we can start thinking about building blocks towards building a little bit of a bigger vision in healthcare and how we can do things with AI. Of course, they will be biased because it's all the things of what I do and what interests me, but I hope at least this will instigate several questions from the audience or yourself and of course follow-up. So without further ado, let me try to do what technology supposed to work and share the screen, and hopefully this will work. Okay, feel free to interrupt me as I speak, uh, if, uh, I, you know, if I'm not clear or if you have any follow-up questions, I'm more than happy to interrupt, be interrupted. So yeah, I, you know, Lena did a fantastic great introduction about my CV. I only had some very nice pictures to accompany places that I lived and worked in the past, such as Greece, Chicago, Italy, and now in Edinburgh. So I think it's good to always think about what really motivates us in life, and particularly in our research. So one thing that maybe, maybe not all of us appreciate, that we're definitely getting older. I certainly am. My hair is already white. And despite our efforts in a lot of major diseases, still, you know, the number of people affected by them continues to remain at a very high level. But there is a hope that with the resurrection of AI, perhaps we can alleviate the pressure introduced to our healthcare systems, either because we are aging or because some of the diseases are still at a high incidence and a high morbidity. So this actually motivates us in part of the group to work particularly in the area of healthcare, but it's not the only areas that we work on. So what we do at Vios— Vios means life in Greek— is that we try to do what we call interdisciplinary AI. So how we can look at the AI as problem, as a solution to several problems, and we try to think of new AI on how potentially also solve some of our problems. So today's talk will center about how is the AI today in healthcare, at least from my point of view, and introduce 4 building blocks that I think will allow us to think towards better representations. How do we develop an AI that can work with multiple inputs, be multimodal, work with multiple tasks, how we can develop an AI that can generalize, and also have to do this with less supervision and less data. Time permitting, I will also introduce a very complex task called prediction, and I will say why it's a hard task. And then probably we'll run out of time, but I will discuss about the need to have systems that we can trust. Now, to— let's talk about the motivation of why we need Epicure AI in healthcare. So this is a typical clinical workflow in the cardiovascular domain. And these are very 3 common cardiac diseases: arrhythmia, heart failure, and ischemia. The only reason why I have that there is to show you the mix of modalities that are being used as a way to diagnose or to ascertain what's happening to the patient. So we can have, for example, the EKG, or we can have, let's say, imaging exams being taken. On top of that, we have the clinical history of the patient, which is, in some sense, a free text record. So you would hope that you can think, you know, the huge burden that clinicians will have is a way to understand and make sense of all this data. And you would hope that AI can actually help reduce this explosion. In fact, we already seen some success. So we already see success of the AI in towards like, how do we can do segmentation or even classify disease? But we can also be more exotic to think about, can we predict what's happening at the evolution of a patient's path through the disease, or even think of prescribing treatments or interventions? And this is Definitely possible, but I'm here to convince you that in order to be able to do this thing well, we do need better representations. So now, how are we developing AI today? This is the classical picture in cartoonish ways of how we see AI today. We see several specialized narrow AI models, all learned via supervision. I do believe there is a future, a future where we can think of a bigger AI that can learn across various inputs, do many tasks, and with less supervision. Now, you will think that's crazy, that would never work. That's not true. It's a reality actually elsewhere. Already Tesla is developing systems where they develop one big model that does all parts, major parts of the autonomous driving aspect from the very beginning. But this is largely relying on supervision. It's a paradigm actually embraced across AI and vision. But here is an example of the Pathways model from Google and the Meta Data2Vec, which are both new paradigms actually less reliant on supervision. So already we see several, I would say, paths that this is not impossible. So now, how can we make this possible into healthcare? So I envision, and probably you will all agree with me, that there are several building blocks we will need to be able to do. In order to be able to deliver this idea, we will need to learn better concepts. Here I use the word representations. So we would like these representations to actually help us build across solving problems across multiple tasks and with multiple inputs. And this is actually how babies also learn. And they actually learn without— with less supervision. You're not giving them any examples of what— how a car looks like. They learn to extrapolate the concepts of what the car is. And of course, we would like to extrapolate to new data and generalize as we try to see new things that we haven't been exposed before and go beyond of the data available in the training So let me start with why it's important to learn better representations. And here, as you can imagine, I would probably advocate for the syntangular presentations, but you will see later how I come to the conclusion of how I raise it up. So why, first of all, why do we need representations? It might be boring, but let me rehash that. To our eyes, an image looks like an image. To an expert, an EKG, they know exactly what to look for. For. For a text record, we know exactly what are the keywords, but to a computer, all of this look like numbers. So we somehow need to develop methods that we can extract the relevant information. Let's take images, for example, and ask the question, what makes a good representation to identify, let's say, the brand of a car? Okay, and you will think about, we would like this system to extract variables that will be sensitive, for example, to the shape of the car, but insensitive to whether the car is rotated in the picture and will be insensitive to the location of the camera. It will also be insensitive to the color because the brand is not directly correlated with the color, unless you think of some exotic brands and very, very expensive supercars. The current approach in machine learning is to think about basically what are all these variables that control the process, and we call these variables the generating factors. So you might think that the factors that control, let's say, the location of the camera, the location of the sun, the texture, the material, the color of the car, or any— how the complexity of the background looks. I'm oversimplifying here using an example of Alan Wills' paper. But this idea of the generating factors, you can start building them up and see how many possible combinations exist. And the current supervised machine learning paradigm says, in order to solve for this problem, we need to accommodate for as many generating factors as possible and will require lots of data. Why? Because modern supervised learning throws away information and requires supervision our work to learn what to throw. This is clearly a paradigm that we cannot continue. We need to think, we need to start thinking in a different way. We need to think of a way that how can we learn what are the right generating factors. And one of the most, I would say, easy approaches is to think of this entangled representations, representations that informs independence between the factors. And we can take it even a step up. We can think of some variables that, for example, look for differences between patients and separate some other variables that look for difference with respect to the sensor. Let's say the different type of scanner, different type of Tesla magnetic strength, and so on. This type of representations essentially break apart things that are related to patient specifics and versus related to sensor characteristics.. Now, this is a paradigm that can work across images, time series, text, whatever. And this is perhaps— it explains the explosion that we see now in work related out of this entangled representations. Now, if you are interested to learn more, we gave a detailed tutorial on the subject in last year's MICAI, and now all the videos are on YouTube. As well. We have a paper that is now under revision and the archive is already available. We have a repository that collects all the code out of this, but is also this year in collaboration with some colleagues from Germany, and thank you for the invitation, we have now a new workshop on the QI in disentanglement, and here is the link if you're interested. We would really hope to invite your participants. Participation and your papers. So now that we've done this, let's start to think about how can we use these representations to actually start learning from multiple inputs and multiple tasks whilst not using a lot of supervision. I'll start with a very simple model just to explain a bit the concept and not going to go into many details, but at least show you how we have used the same idea for several other problems. So let's say we have an image of a patient and we're going to use a network that's going to extract binary spatial variables. So each channel, I will call it, its output will have multiple outputs. It will be a binary variable that will say I'm white if I belong to the same thing, but we don't know yet what that thing is. We will also have another network that will extract some sort of sensing characteristics out of this image. Like statistics about the image of the distribution of intensity and so on, the characteristics of the noise, etc., etc., etc. If I were to put them together, they can actually reconstruct the input. So this is a perfect self-supervised task here with a simple reconstruction. However, I still need to know some information about what is the thing that I extract. So here I can have some annotations from segmentation, or I can actually use predictors. And regress to other tasks or other information. This simple paradigm is actually very effective. It takes about 6% of the annotated data to come up with a good quality of segmentation output, as we measure here with DICE in the how accurate we can delineate the myocardium, the left ventricle, and the right ventricle. But more excitingly, it's if you think about how much extra knowledge I can gain from electronic healthcare records. So imagine that I have another dataset where I record some information about the electronic healthcare records as they come out from the PACS. I can actually create a regression task, and if I add this additional dataset, I improve my performance to 83.2%. This is actually very close to have full annotation. It's actually equivalent to adding 8 times more per-pixel annotation. Now, there is more things that this model can do. So it can actually work— the same exact model can work together with MRI and CT data. The same model can also be used to exploit temporal relationships and learn better representations by predicting future frames in the cardiac cycle or in a temporal sequence.. It can also be used to expand to pathology and do what we call triple disentanglement. You can use for multiple inputs simultaneously in a very nice paper we did in TMI to kind of simulate how clinicians, radiologists look at images. And several other colleagues have extended the same concepts around other organs, modalities, imaging settings, and problems. And the paper also describes some of these applications as well, the tutorial and survey paper. But I think, as you all probably believe and agree with me, generalization remains a very difficult task. So let's— I will try to put some more emphasis here because I think it's a very exciting area, and I think our approach is a little bit different. I would say. So the classical ML paradigm is that, you know, if you want to solve a generalization problem, collect more data. So the more data variation you have, the better performance. So you can think of, let's say, try to collect data from all possible scanners or all possible imaging devices that you will use. But, you know, ideally we would like to think of that this is probably not going to be easy to scale up. We would like to create an AI that can generalize and separate the variation that comes from the sensor, the device that we use, versus from the population. And dealing with that problem is not very easy because we don't have gold standards. And the gold standard to think about how we can isolate the influence of these two variations or factors in the domain shift, you can think of it as having the same patient traveling from one vendor to the other. So you can think of going to vendor 1, vendor 2, vendor 3, and so on, and getting the same patient being imaged by all possible vendors in a classical medical imaging setting. This is impossible, right? But we can actually simulate this process. If we create technology like the one I showed you before, we can separate the effects of what's happening with respect to anatomy or pathology versus the effect of what's happening of the sensor. We can separate them and then we can actually virtually start mixing coefficients between what's coming from information of the anatomy variables versus what's coming from the sensor characteristics and artificially create new data as if a patient was traveling to different sensors, to different settings. And we can use this as a way to augment and regularize the data that we have. This, of course, relies that we have done this in a perfect way, that we have separated this perfectly, the characteristics of the sensors from the characteristics of the anatomy. But we can actually do better by saying, well, if you haven't done the separation very well, improve the learning of the representations. And how do you do that? Basically, we do it with what we call— what probably you've heard of also before— is called meta-learning. So we essentially hide— we have, let's say, variable 3 sites, 3 different sensors with 3 different populations, We hide one of the sources of information, so one of the sides from the model. We train on the other two. We test the performance on the one that we've hidden, and then we see what doesn't work, and then we penalize the representations. We came up with this, I would say, first model that tries to solve this domain generalization problem in a semi-supervised setting. As you can see, it does well, particularly also when you have a fraction of the annotated data available, and it does much better than just random mixing of the variables that I showed you. So it means that there is still room for improving how we disentangle the representation. Now, meta-learning is pretty heavy in terms of training, so we start thinking, can we go backwards and say, do we need all the components? Do we need to be so complex in how we design the models? This is a typical issue with this entanglement. And just a recent paper out of my team of Xiao following up on this work where we are thinking of better bottlenecks as a way to kind of like quantize the coefficients, as a way to quantize where do we learn and how do we disentangle and separate information. And he's using here the concept of compositionality, which is, without going into too many details, basically this is the feature map as it comes from an encoder And then we learn how to quantize this by doing clustering together with learning good representation kernels. So, and then we learn how to put it back together as a way to reconstruct. So this, by doing this, we can also learn in an unsupervised fashion or semi-supervised fashion. But the biggest benefit is that this network is actually very lightweight and it's much faster to train. There is no real need for meta-learning. And so on. And it works efficiently also in very, very low short settings, even below the 20% annotation level. Even 5%, it beats the competition by far. Now, I've talked to you about sensor variation, but I haven't talked to you much about what happens with population variation. So, and here I would like us to think about we will always have not the full picture of the population. We will always have parts of the population that we under-explore. Okay, so we might have somebody who's normal, somebody who has, let's say, some disease, and another one, the different type of disease, but we will never be able to explore everything. The question is, how would the patient look like if they were somewhere here? Or how would a patient look like if I could imagine an interpolant patient going from, let's say, here to there and somewhere being able to sample somewhere in the middle, how would the patient look like? And if I will be able to create of how this patient would look like, then I can actually start transferring some characteristics and use that synthetic image as a way to throw it back into the generated data and improve the learning of our models. The key here is to try to do this in a very high quality and also be able to carry over learned segmentation outputs or learn facts that can take from the two inputs. So the way we address this is we call it anatomy arithmetic. So we actually do the arithmetic on the binary spaces that I showed you before, and we do this with a network that we called noise injection model module where we try to solve the discontinuity problems by allowing for some stochasticity. So there is some relaxation of the problem involved in how we approximate this. But I, I think without boring you from the details that we can find on the paper, and there's all our source code is available for everything that we do, I would like to show you at least why I'm excited about this, because I think the way I see it is We don't want to over-augment everything, right? We want to augment where we have, let's say, rare pathologies, or we have underrepresented sites and hospitals, or we have perhaps gender-race imbalance. And the only thing we were able to do all of this is to have appropriate classifiers that will separate, for example, let's say, for the gender-race imbalance, and then use that mechanism to drive our synthesis. So we can tune where we want the synthesis to happen. But also exciting is that we can generate data by design by actually doing image editing. So here is an example where we take, let's say, the left image here where we have the factor. It's binary now, so it's like an accompanying segmentation output. We can actually go and say shrink this and make it smaller, and we can get an image by design that makes this look smaller. So this is a part of the input to the model, and then we can generate a synthetic output, and we can overlay this in other images as well. So you can create bespoke, I would say, data should you wish to do so. Now, another thing that we try to approach this is not by synthesis, but looking at things that we tend to actually ignore. So maybe some of you are familiar with the idea of test-time training. So test-time training is the concept of, can I adapt the model on the fly at test time without using annotations? So what we observed is a lot of the things that we do, but also a lot of the things that others do, capture priors. So for example, if I were to use like this this model that I showed you before, that it has an autoencoder here. This encoder captures a prior. This decoder here captures a prior that we use and that we learned during training. But at test time, we actually throw everything away. We only keep, let's say, the segmentor or the predictor, but all of these other things that we've learned, we throw them away. We do the same thing if we learn a conditional GAN, for example, where we have a segmenter and here we use a discriminator, for example, to learn a shape prior. We throw this away. So with a paper, with a series of two papers with Gabriele, we said, well, let's stop throwing them away and reuse them. So we designed discriminators that are actually able to work during inference and not collapse. So they still provide you informative gradients backwards that we can use to actually backpropagate gradients and update part of the segmenter. But we also used it in the reconstruction models where we actually use the reconstruction module, and that gives a causal interpretation to test-time training. And you can find this on the second paper. Excitingly, you can see, and we did it on several datasets, we trained several models on several datasets, and you can see before test-time training, we make lots of mistakes, such as here, the model makes mistakes. But by allowing it to do some test-time training, it adapts and it comes much closer to the true ground truth. Now, it doesn't observe We observed that it doesn't always help to bring up the average a lot, but it really shrinks your outliers. And I think, why not? It comes in some sense for free. Now, over the next few minutes, I would like to speak about a little bit of a more complicated task, a task that— of prediction. Or even prescription if you want to think about it even more long term. And there is— we have a large body of work here, but I want to highlight only a couple. And I will try to motivate it initially with a problem of longitudinal data. So how can we think about learning spatiotemporal manifolds? Without longitudinal observation. So if you think about in our— it's easy to collect data, for example, of faces for how people age, okay? But it's very hard to collect longitudinal data of the variety of the diversity of the population, of how their brain ages, for example, or their heart ages, or other organs, how they get affected by the progression of age. So in some sense, we really don't have longitudinal observations. We have sparse samples of diverse parts of the population as they progress in life. So in some sense, how can we learn the effects of age on an organ if we have sparse cross-sectional data? So can we separate the aging process from individual characteristics and somehow learn bespoke curves, but somehow all of them get learned at the same time, and we learn the effect of age, for example, on the brain. So we designed actually a pretty simple model that is a conditional GAN, but the key here is that the way we formulated this conditional GAN to regularize it such that it doesn't deviate and create, let's say, the brain, how the brain will be aged of another patient. We would like to maintain what we call the identity. But another key here is how to learn these models here that will do joint learning between images and other factors, other numbers. And this is a theme that is occurring in a lot of machine learning now of how do you integrate different sources of information. Information. Now, this was a very simple model, and the name that we picked at that point was terrible. It was called Transformer because then BERT came in and had another name, and everybody gets confused that this is a transformer. It's not a transformer, it's a simple convolutional neural network. But I would like to show you mostly the results, particularly visually, in another morphometry study that we did to show you how we can take a single patient. And we can age them assuming that they have no disease. As you can see, there is a nice progression of the disease here— of the, sorry, of the aging brain. Or if they have some form of mild or advanced Alzheimer's. Now overlaid here, this is the ground truth age that we asked the model to synthesize, 77. But overlaid here, we have an image-based age prediction. So what is the apparent age of that brain? And what you see is that the, the, this image of somebody who has Alzheimer's appears to be progressively aged. So they're higher in apparent age of 80, like 3-year difference almost, versus if somebody is healthy, they are exactly— we ask them to be 77, and they are— the model predicts that this image looks like 77 years old. So the model, without understanding anything about that the patient of somebody who has Alzheimer's, the image should look his age, it actually has learned to advance and accelerate aging for people with Alzheimer's. This is something that we see also when we do a more, I would say, classical analysis. Where we do morphometry. So we take the images, we stack them in the volume, we segment all the major parts of the brain, we co-register images between, let's say, baseline and follow-up. Again, this is from ground truth because we have some ground truth available in the AD dataset. And we look at the relative change in middle temporal gyrus, and we do this between baseline and follow-up. Across the patients that have Alzheimer's and they don't have it. We see that there is actually a shrinkage, a change. But what will happen if we take the same baseline and rather than use the actual ground truth follow-up, we put it through our model and we predict an output? If we do that, as you can see, we still find there is a change, and we find that there is— the F-statistic is actually very close from the one-way ANOVA.. So it is kind of, I would say, statistical proof at least that the model does learn disease effects. And the second key learning out of this is that this means that we can actually use cross-sectional data to learn longitudinal effects. And we can perhaps then augment new studies, or we can actually think the potential of pooling studies across cross-sectional, longitudinal altogether rather than treating them bespoke. Because I truly believe that if we want to start solving some of the bigger problems in disease, we have to start pulling them together, lots and lots and lots of different studies. And very recently, this just came very quickly in my email box last week, we are now in collaboration with colleagues from Pompeu Fabra in Spain, we are repeating the same work. In the UK Biobank. And this is just early images of where we're trying to do the same type of cardiac aging in the heart and see what is the effect of aging in the heart. But I think starting with a picture that I showed you before, it's useful to think of where do we really like to envision AI helping us in the future. Yes, I'm pretty positive we'll achieve very high quality reconstruction, image processing, segmentation, characterization of the disease. But I think it's worthwhile to think, can we think of AI helping us in prognosis or understanding what is the optimal therapy to assign or when do we do follow-up? I mean, in the cardiovascular setting, for those who are not familiar, One thing that might happen after myocardial infarction, it could be that you get arrhythmia or pericarditis or even a stroke. So wouldn't it be nice to know what was the most likely, you know, complication for a patient? Maybe that will allow us to optimize treatment, optimize when we do the follow-up and how we monitor the patient. Now, in order to be able to identify what is the optimum treatment or the optimum path for a patient, you have to start thinking of measuring the effects of doing nothing or doing something. And the promise of this lies in causality. So because with causality we can think of like, you know, we have a patient, we can have, let's say, potential outcomes like treatment, uh, when we give a treatment or when we don't give a treatment.. And this is what we call the individual treatment effect. This will be the ideal gold standard that, say, randomized controlled trials would do. But that's, as you can guess, randomized controlled trials are very, very difficult to execute and costly. But there are advances now in machine learning where we start thinking of, can we start thinking of obtaining counterfactuals? That given, let's say, something observed, we can hallucinate a potential counterfactual for that patient. Now, the analysis that I showed you before on the brain stuff, we can take the image of a brain of a patient, let's control and hallucinate the counterfactual of that patient, how they will look older and with Alzheimer's. So you're already thinking of like imaging-based counterfactuals. Counterfactuals here, but they were not founded in the causality domain. And very recently, we completed a paper with one of my students, Pedro, on thinking about the concept of counterfactuals but embedded within a causal framework. And in order to do that, we started working on a new class of generative models called generative diffusion models. Where you think that you have, I would say, a causal graph and you relax the connections of this causal graph. So if these are the variables, what we call the causal variables X, and U are the exogenous ones, you do a diffusion process where you relax in time these relationships and then you put them back together. So, and the way we do the counterfactual estimation, we actually We first infer the latent variables with the forward diffusion, and then we intervene and we cut things on the basis of gradients coming back from anticausal predictors, anticausal classifiers. Now, it sounds pretty complicated, and trust me, I cannot do it justice in one slide, but essentially what this allows us to do is that get the factual do an intervention, and obtain a counterfactual. But one of the most— and this means that we can estimate treatment effects, I would hope. But as of now, we demonstrate that we can manipulate images. So we can say, for example, if we have an image of a plate with a carbonara pasta here, you can actually say, do espresso maker and plop an espresso maker on the top. I know it's ridiculous, but The key concept here is that how do you modify this image without deviating a lot from the original? How do you maintain identity? And this is something that we also worked on to come up with new metrics that measure counterfactuals in the process of preserving identity that has not been done before. Now, how can we use counterfactuals? So you remember this caricature that I showed you before, where we can create synthetic data. But we found out that not all synthetic data are useful. So we wanted to ask the question, can we synthesize only what is useful? So in some sense, if we have some data and they are spanned in a manifold of, let's say, a feature space, where we have some control patients and some of them that have Alzheimer's Alzheimer's disease. If this is a classifier between the two, try to find the data that make the classification hard, and then use this data as the origin that we can create counterfactuals on, that we can somehow then use as a way to augment the data. And this is a paper that we now just uploaded to arXiv, where we used a fixed generator here. It doesn't matter, it's fixed. But this fixed generator is conditional on, let's say, an image and some other attributes. And we have a classifier here. So we identify what are the right attributes that optimize the classifier and the classifier's performance that actually make their job harder. So we go and identify examples where they get very either misclassified or they have a the loss is very high, we identify and we search around them good attributes such that we can modify these examples and augment them. And you can see that for some of the cases where a classical, let's say, classifier for an age group will have a high difficulty with a naive classifier, just doing random synthesis helps. Just training twice also helps on the hard examples.. But using our augmentation helps even further. But critically also, we identified that the same mechanism can be used to correct for spurious correlations. So we can use augmentation as a way to guide where we would like to correct for spurious correlations once it has been identified. Now, in the last few minutes before I conclude, since I still have a little bit of time, I would like to talk about another final ingredient. It's not a building block, it's actually, I think, an ingredient of trust. And, um, and although it doesn't work direct, it doesn't highlight immediately this entanglement, there is this entanglement behind the scenes, um, and there is definitely motivations for better representations to solve that problem. But I would like to highlight some of the risks of trust, losing trust, and easy things of how we can actually create models that can break trust. So in a very, I would say, recent study of perception of the value and risk in machine learning applications, patients found that using AI for health diagnosis is something of very high social value and very low social risk. So they are with us. So we have to ensure that we maintain their trust in us. Okay, but here is some interesting facts that we've recently started digging up. So we had a— we have a very large survey on the effects of leakage and privacy during inference time. So what happens if I train a dataset on the training data, and then how can I use several attacks to identify something about the training data that I trained with? So whether I can infer whether the training data was used as part of the training set, or I can use, for example, to infer some unique property about, let's say, a patient, about the data, the gender, and so on. Or actually reconstruct training data directly with what we know as a reconstruction attack. So, and even in a more recent paper, we found something very shocking that let's say that this is a caricature here. So let's say we have an X-ray image of the lungs, of the chest here, and there is by mistake the name written of a patient on the top. But let's assume that this is in a single image in all the data sets happening, just in a single one, because sanitization processes are not perfect, particularly if you're learning on vast data sets behind, let's say, hospitals, as one of the projects that we have with the NHS. So what will happen? What is the probability that a machine learning model actually learned this thing about the patient, like learned their name and somehow memorized it. So we proposed a metric, and we would like to do this in a black box, so we don't know anything about the model. We only have its outputs, and we don't even know the training data. So, but we do know that this will be a name. So we proposed a leakage metric where it uses only out-of-domain data to show that we can actually detect leakage happening even when one datum on the whole training set contains some unique feature, unique private feature. And classical ways of preventing this, that is regularization, does not help. And even more important is that hard-to-learn data, data that confuse the classifiers, like the ones I showed you before that are hard to learned that we need to augment actually are even more prone to leakage. Now, of course, you can think of, well, can I prevent this thing? Can I help it? We've done some work on if I develop, let's say, models that have been trained from different data sources, and then some of the data sources, can you make sure that you forget some of the data and can you verify that you have forgotten? We came up with a model that we can measure whether a dataset has been truly forgotten, given a classifier, whilst discounting against data similarity because patients' data look similar. So if you think of images of the face, for example, we all have eyes and ears. So we cannot assume that the model will forget how to detect eyes and ears. We just would like to forget my face. We also then said, well, how can I forget? In an area where we call machine learning, we came up with a new model that tries to say that, can we forget certain pieces? Can we unlearn information from models? And we found out that easy data have a different way to be forgotten and should be unlearned. But edge cases, like rare cases, they need a different process. And this is something that is still work in progress. Now, in the last few minutes, I would like to kind of summarize over these ingredients and kind of tell you where I think still there's still work in progress that we should focus on. So we should definitely be— think of doing things with less supervision because we will never have enough data, I think, particularly in healthcare. Rarity, some data will always be rare, will always be a problem. So we need to think of how we can take advantage of rare data and devise methods that work well also with rare data. Annotations will never be plenty. So we have to think of how we can come up with zero-shot, low-shot, few-shot, continual learning, and so on. Now, I always believe that the data will always be hard to acquire. I think that's a helpful way to think of coming up with new solutions. There is some progress of using, let's say, multiple tasks, multiple multimodal inputs, different inputs, and so on. But there's still lots to be done. And I think it's a very fertile area. And causality, the proper way of using predictions and doing predictions, will all still remain, I think still remains a hard task, particularly for imaging due to the high number of variables involved. Now, of course, it wouldn't be my presentation if I didn't talk about representation learning, and I think representations will remain important. And I think disentanglement is a good start. It's an easy prior of independence, but it's not the best prior. We have cases where the data are correlated. We have cases that our data are rare. Is disentanglement the right prior for that? And of course, I talk here about concepts, but really I didn't tell you about concepts. I talked to you about representations. So how representations come together to devise concepts is still a very large work in progress and still a lot to be done. And finally, generalization. Is, I would think, the problem in AI. I don't see a single solution to this problem. And yes, whether you can use synthesis, test-time training, augmentation, you know, continual learning, few-shot, all of this will help. But the key question is, they don't help for all the right problems, for the same problem, and always. So we need to find mechanisms that find the right solution for the right ask for the right data. So to kind of close visually with the same things that I started, yes, we are getting older. Disease is still an important thing that we have to worry, particularly if we want to think about affordable healthcare systems. But I'm still optimistic that AI will be very helpful in trying to achieve and resolve some of the bottlenecks. So without further ado, I would like to thank you, my team. We have always openings if you're interested. You can find us on our website. We have all the code available. And I'm not— I try to be on Twitter once in a while as well. And I also lurk the Reddit forums. And I would like to also thank my collaborators. Not all of them are here. Funders and yourselves. Thank you so much for your time, and I hope we still have good time for questions.

Thanks a lot, Sotos. Very nice talk as always. Some clapping in the audience. Thank you, Len. So, um, yeah, so as always, just either raise your hand or paste your question in the chat. Maybe I start. I'm always particularly interested in what doesn't work. So for example, we have started looking into disentangled representations, and we were always very happy about how the latent space looked, but it didn't really give us a benefit in the downstream task.

So yeah, what's your view? So I get this— let me stop the share. I get this question very often. And I think, you know, if I'm going to learn, let's say, a deep model, let's say a deep ResNet or without any restriction on the latent space. And we compare, let's say, the performance of this model with a disentangled model where you actually induce priors, right? You induce, you hard force certain restrictions on the latent space. First, to have a latent space, and second, that it is disentangled. So you are bottlenecking, right? So it means that on par, maybe performance will not be the same. So if you're comparing performance to the downstream task, I think it's not going to help. But I think the opportunity lies in the fact that if you have disentangled, let's say, spaces and you have multiple variables, then you can start thinking about how can I properly use, for example, multitask learning. If the tasks compete in a general way of learning representations without any sort of disentanglement, introducing extra tasks may actually hurt your performance on another task. So I think this is something that will advocate for the use of disentangled representations. And I think the other thing that complicates disentangled representations, particularly in our domain, is that a lot of the work that you see now, particularly in vision, they trained, they tested, and they use datasets that have a more or less ideal distribution of the factors that are involved. So you might think a lot of possible combinations, let's say, of, you know, faces in a nice background where, you know, the, the, this classical CelebA dataset, right, where you have people of celebrity, faces of celebrities, not mine, uh, where they're centered around the nose and they all have a very similar pose and so on. So this entanglement works very well where your input data have some sense of a uniform distribution in the factors that you observe. If that's not the case with the data that you use, disentanglement may lead to not learning a factor that you really need or learning it with some skew, which hurts performance. And basically, essentially, I don't know if you recall the last slide where I said correlated data, If your data are correlated, so there are correlations in the factors, disentanglement will not help. So there is need— we need to think of how do we deal with correlation in the context of learning better representations. So separate the uncorrelated stuff, keep the— and learn the correlations at the same time of the factors. Otherwise you reach collapse, you collapse the model and you actually don't learn what you I think this is perhaps the better explanation of why you haven't seen the performance that you would expect.

So when you get a new student working on a new problem, what do you start with?

What would be your baseline starting point for a new question? So I think we, that's an interesting question. So I have new students, they suffer, but I think the first question that we start with is what happens if we don't have lots of data and we don't observe all possible combinations of the factors? That's the question I start them with. I think this is a problem of how do you learn good variables that can solve for many tasks when subject to correlation or even worse, causal relationships. So this is why I introduced the concept of causality as well, because the concept of causality also is embedded into the problem of causal representation. So how do you learn representations that are causally related? I think it's a fantastic problem. And, and there is fantastic people work within LS in that area as well.

Okay, any questions from the audience? Everybody's so silent today.

We had after-work drinks last night. Ah, okay. Maybe that's the reason.

Yeah, I have a question. Yeah, go ahead.

Thank you, Professor, for the great talk. Great talk.

So I saw the example that you gave for your adversarial classification training in the paper Adversarial Counterfactual Augmentations.

Yes.

So if I understand correctly, the generator in that work was conditioned on age, and brain images were generated— synthetic images of the brain were generated. Based on predicted age. Do you see anything more, let's say, general than age for that conditional task? Maybe it can be applied to, I don't know, other problems other than classification of the classification task in that problem.

So I think the— yes, the answer is yes. You're going to say, why did we use the model and in Modda Generator because that's the one we had available in-house and we can easily test the idea. I would suspect that anytime you have a conditional model where you can control some of the attributes of generation, that you can do basically what we did. You can backpropagate the gradients and identify what are the perturbations that you need to do on the attributes that will generate the data that you want. You want as a way to create the counterfactual. So the short answer is yes. The long answer is we haven't tried yet another model because the paper that you saw finished last week or two weeks ago. I see. So, um, we did have a discussion with Jan if we can try it on other, uh, on, let's say, a StyleGAN general problem, uh, as well, uh, or other conditional guns, but we haven't gotten there yet.

All right, thank you. You're welcome.

Other questions from the audience? Yeah, hi, Tim here speaking. I have another question. Thank you, first of all, for the really nice talk. A lot of interesting food for thought for afterwards. My question would be, I have oftentimes the understanding or the feeling that I think— what's the correct English word— that oftentimes the models that people are trying to build are like jack of all trades, like they should work on young people, old people. You have obviously different genders, different skin colors and so on. But if I think about it in reality, I'm— does the model, if we later would deploy it, for example, in a local hospital here in Germany, the distribution of patients that would come in would not feature this jack of all trades, like it would be very specific. Just geographic locations and so on. So my question would be, does it even make sense to have every single factor that we know about have resembled in these representations? Because later the patients who would use it, for example, in Heidelberg, would be different users than, for example, in Nairobi, for example. So does this even make sense?

And I ask myself, yeah. Yeah. So that's an interesting question because I think you're basically advocating against generalization. Generalization but specialization. And I think I would always think that a general model can be easily specialized, but a specialized model has the risk of learning very, very lots of spurious stuff. And I don't think that's spurious correlation stuff is something to ignore. So when you're pulling a lot of data together, the likelihood of learning something serious reduces. Okay, so learning an artifact or learning something weird characteristic that's happening on the local data will be less likely to happen when you are creating bespoke, like, say, specialized models. I do see the value though of getting a generalized model and adapting it, and I think this is the— I would say the hope with what I told you before, what I presented on test-time adaptation. So I think there is definitely the idea of can I adapt to the local population, and you know, even guidelines in healthcare, like, you know, when do you, what is the lower, the high value of cholesterol upon which you prescribe, let's say Lipitor, right? And, you know, they use different values in the American population, different values in the British, different in the UK, different in Germany, different in Greece. If a Greek, as a Greek myself, when I was living in Chicago, every time they tested me for cholesterol, they wanted to put me in drugs when I was 21 years old. But I'm Greek, we basically are bottle-fed olive oil from when we are born. So that leads to higher cholesterol, but also leads to higher good cholesterol. So I definitely believe there is an interest in specializing to populations, but I will always advocate that it's better to generalize and then specialize than to create a lot of bespoke specialized models. In fact, Information sharing in the models helps.

They learn better.

Thank you.

Definitely good food for thought again.

Yeah.

Yeah, so maybe one last question, Sotos. So I wouldn't know how to answer this, but you're an expert. You showed us a lot of nice public— but can you, as an expert, can you point us to something really exciting that you saw outside of your lab?

Publication that you thought, okay, this is, this is really— Yeah, uh, sorry, I, I actually had one open in my reading list and now I— hold on, let me pull it up. No, that's not that. That's not it. So there is a paper, sorry, I'm trying to find the window because my memory usually is not the greatest. And there was a paper that I popped yesterday about NeRFs. I don't know if you're following the whole NeRF explosion in computer vision about the neural radiance fields. I think it's a term introduced a year and a half ago. And suddenly I think there are like hundreds of papers, one improving on top of the other. It's mind-boggling what is happening on the ability to create 3D scenes out of basically nothing. It's unbelievable. I'm not an expert, but I follow it like crazy because the results look— every day on Twitter, a new picture appears of new results better than the other. So NERF, I think it's very interesting. I don't know how we can use it in the medical field, but I think it's very, very exciting to listen to this from a computer vision perspective.

Yeah, I agree. We also started discussing. Okay.

Yeah.

Many, many thanks for coming. Yeah. Have a good rest of the day and hope to see you soon, maybe in person at some of the future meetings.

I hope so too. I hope so too. And thank you for the invitation again. And thank you to the audience for bearing with me and to Dorin for the help.

Thank you very much.

Bye-bye. Thank you. Bye. Goodbye, everyone.