 Good afternoon everyone or afternoon in Vienna time. My name is Melissa Danica and I am the director of the division of physical and chemical sciences and the department of nuclear sciences and applications here at the IAEA. Artificial intelligence and machine learning are topics presently permeating all four sections of my vision so that I am especially honored to be co moderating today. So, hello everyone and welcome and this is a May of the will have on the director of the division of human health at the IAEA and delighted to jointly organized the side event on this very important topic. We're looking forward to some exciting talks and to hearing some of your questions. Yeah, welcome to the side event. The future of Adam artificial intelligence for nuclear applications. Today, we will showcase ways in which artificial intelligence based approaches can be used in nuclear science and application. For example, in nuclear fusion research and in efforts to protect global water resources from over exploitation and contamination. The goal we have is to give you a flavor of future applications of AI in nuclear science. And to nurture ideas of possible collaborative and coordinated research that can be conducted under the stewardship of the IAEA. So, additionally, you will find out about how artificial intelligence based technologies can be applied in medicine and enhance diagnostic accuracy to enable quicker detection and support of radiation treatment planning and treatment delivery and radiation oncology and radiology. We'll also take a look at the challenges and ethical considerations surrounding big clinical data. And why medical images require special attention in the QA process. We want to understand the pitfalls and what can happen and how to avoid them. We'll also look at the role of the education and possible frameworks that could be established in relation to education. So many exciting topics. And now I'd like to give the floor to the head of the IAEA Department of Nuclear Sciences and Applications, Ms. Najana Khattar, for her opening remarks. Ladies and gentlemen, it's my pleasure to welcome you to this live streamed side event on the future of ATOMS, Artificial Intelligence for Nuclear Applications. This is an exciting topic that has real potential to benefit humanity. Artificial intelligence, or AI, is advancing exponentially. AI codes can already sort and interpret massive amount of data from various sources to carry out a wide range of tasks and help tackle many of the worst, most urgent challenges. For example, artificial intelligence ability to recognize data patterns and analyze high resolution images from satellites, drones or medical scans can improve responses to humanitarian emergencies, help doctors identify cancers and other diseases, increase agricultural productivity and track animal and marine migrations. In fact, artificial intelligence will be an integral part of the agency's new Zodiac projects, helping to identify and contain future zoonotic disease outbreaks. This side event is bringing together leading minds in artificial intelligence and establishing a historic milestone as the first event to launch an inclusive global dialogue on the potential of AI in nuclear science and applications. I'm grateful to our distinguished speakers today who will offer us insights into three key areas of AI, applications in nuclear science, namely human health, water resources and nuclear fusion research. The transformative power of artificial intelligence also comes with challenges ranging from issues of transparency, trust and security to ethical concerns. So we will also hear a brief overview of the ethical aspects and challenges of AI today, the implication for education and its role in shaping up the future. The IEA is making artificial intelligence. We expect that today's discussion will help build a common understanding of AI's relevance to the peaceful use of nuclear technology. And I hope this event will foster practical idea and proposals for collaboration which can be followed up in the months and years ahead to ensure that member states can take full advantage of this essential emerging technology. I thank you all for your active participation. I wish this event every success and very much looking forward to hearing much more on this topic in the future. Thank you DDGNA very much for that introduction. As you gathered from her message, we will hear four presentations from distinguished experts. And as you gather from Mateo, we will be taking a couple of live questions at the end of each of the individual talks, but we will have another Q&A at the close. Before we move on to those keynotes, you will now see a short introductory video about artificial intelligence published by the United Kingdom Royal Society. These are exciting times for artificial intelligence, or AI. But what is AI? AI is the science of making machines smart, using algorithms to allow computers to solve problems which used to be solved only by humans. AI already powers search engines, online shopping recommendations and digital assistants. Radiologists can use AI to calculate the exact shape and volume of tumours, revolutionising treatment. Astronomers use AI to find and evaluate exoplanets in distant solar systems. And unhappy motorists are using it to discover if they can overturn parking fines. The possibilities of AI are endless, from fraud prevention to developing new strategies to address climate change. AI technologies today are brilliant at analysing vast amounts of data to learn to complete a particular task, a technique called machine learning. But AI is not good at transferring what it has learned from one type of task to another. Learning abstract concepts and one-shot learning, where a general rule is learned from a single experience, the sort of broad intelligence humans excel at. But even without achieving this broad intelligence, AI is raising questions. How might AI affect work? People will still be needed in the workplace, but jobs and roles are likely to change as some tasks become automated and new jobs are created. There is a risk that, at least in the short term, this could increase inequality if some people are disproportionately affected and benefits aren't widely shared across society. And what about social equality and fairness? Bias coming from AI systems, even if unintentional, could disproportionately affect some social groups, influencing anything from job prospects to treatment by the justice system. Technology can play a role in helping to manage bias, but people need to address big questions about how society wants to use AI. Technology with AI at its heart has the power to change the world. The more of us that engage with shaping its development, the more chance we have to ensure a better, fairer future with AI. So such an interesting video from the United Kingdom Royal Society, and it's a lot of food for thought. So now we'd like to get started with the key notes. So dear guests, let me introduce our first speaker, Professor Sanjens, who is a radiation medical physicist. Full professor at James McGill Research Chair, Director of Medical Physics at McGill University, senior scientist at the Research Institute of the McGill University Health Center. He's actively involved with committees in AAPM, IAA, ICRU. He's co-authored over 270 journal papers, technical reports and books, and his involvement with machine learning and AI has been through project learning from medical images, clinical parameters for prediction of cancer outcomes in radiation medicine. So he will highlight the potential for AI to support physicians in the diagnosis and treatment of cancer through better image interpretation, improved treatment plans and contouring, as well as allowing efficient adaptive radiotherapy. So the generation and analysis of big data can be aided by AI for rapid analysis and interpretation and will allow action and will improve effectiveness and efficiency of medical processes. Professor Sanjens, the floor is yours. Good morning everyone. My name is Jan Sanjens and I'm a professor and medical physicist at McGill University in Montreal, Canada. In the healthcare environment, artificial intelligence or AI is already present or making its entrance in a variety of subfields, including detection of disease and that includes both from medical imaging as well as from digital pathology. The management of chronic conditions, the optimization of healthcare services, logistics, optimization of treatments, drug discovery and so on. In this presentation, I will concentrate on diagnosis and treatment of cancer. Modern oncology, although very advanced, is still largely a one fit for all endeavor. With this I mean that the treatments prescribed for patients with a particular disease are still very much generic, or in other words the same for from one patient to the other patient. AI gives us the opportunity to make treatments much more personalized, taking into account all the information or data that we have on that particular patient, and this is known as precision medicine. Within AI, one hears often the term machine learning. In simple terms, machine learning is the study of computer algorithms that improve automatically through experience. The computer algorithm is shown structured data, for example images and labels or annotations, for example a classification of damage versus healthy. The more variable the data algorithm can be given, the better it can learn all the characteristics in the image that lead to a decision of the outcome being healthy versus damaged. We call this the learning phase and in this particular case, since the images were annotated, we speak about supervised learning. At some point the performance of the prediction is tested using a part of the annotated data set that had been previously unseen. This is the testing phase and in the testing phase we can estimate the accuracy of the prediction. Only when these tests are completed satisfactorily can the algorithm be used on a new, not previously seen image and make a prediction. Deep learning is a form of machine learning, where the algorithm can learn on a complex and variable data set. Deep learning has been very successful in applications such as for example recognizing cats and dogs on photos. There is a large amount of information floating around in a medical environment. As soon as the patient enters a medical center, data is being generated and stored. For example a patient registers at the hospital and basic personal data including address, age, sex and so on is entered into a database. The patient then continues to see a physician who enters a doctor's note in the database and perhaps some tests are ordered. This could be an x-ray image, a blood test or a CT scan. The doctor's note could also contain some glimpse of a diagnosis based on clinical examination. As the patient proceeds through the system all these data elements are being collected. With all of these data it is becoming increasingly difficult for medical professionals to take into account all the information they have access to. It has been shown for example that humans can process only about 5-8 variables at a given time. In addition, the participation rate in clinical trials which is the gold standard of medical hypothesis testing is perhaps less than 3% for a variety of reasons. This means that with not learning from all the data that present itself an opportunity is missed. We can say that from the perspective of AI healthcare is a big machine that generates real world data. With the capability of AI to ingest large amount of data we could use real world data to help us predict using AI how the patient will do. We could optimize intervention based on the learned experience from previous patients combined with data for that particular patient. We will now concentrate more on cancer and on medical images as data for AI. Patients undergo different types of imaging including x-ray based imaging including computer tomography, imaging in nuclear medicine such as positron emission tomography or could be ultrasound or MRI with different sequences and scanning protocols. Different types of images represent different biological features, anatomical features or metabolic processes. In radiation therapy the goal is to treat tumors and spare normal tissues as much as possible. This involves the annotation of what constitutes the so-called target volume and what constitutes normal tissues or normal structures and those are established by using segmentation or delineation tools on three-dimensional patient images that define where the target and the different structures that need to be spared are. Computer algorithms are used to solve the optimization problem of maximizing radiation dose to the target while minimizing radiation dose to critical structures. Nowadays AI is implicated with making this planning process including the segmentation step and the optimization much more efficient. Radiotherapy is delivered in more than one session and on the right we show how imaging is used to verify and correct the position of the patient for that particular session. Here we bring together some of the main areas of application of AI in cancer radiation medicine and radiology. These include the segmentation already mentioned, quantitative image analysis known as radiomics and we'll come back to that a bit later, outcome prediction where we use data about patients and try to learn outcome, radiation dose quantification for fast dose calculation algorithms, dose response studies and generation of images, for example, synthetic computed tomography from MRI. In radiology aside from segmentation, applications can be found in classification for disease staging, point of care detection of disease, image registration where we align images of different types with each other and other applications. Here we focus on medical images used in precision oncology. On the left we show a number of different types of data that might be available for a given patient including medical images, pathology images, genomics profiles, proteomics, demographics, doctors notes and the like. Focusing on medical images, radiomics refers to the use of quantitative information from those images. For example, in quantifying aggressiveness of a tumor, the underlying hypothesis is that the heterogeneity at the genomic scale is translating into heterogeneity at the anatomical scale. The idea being that that type of information can not be qualitatively discerned but can be captured using quantitative image analysis, sometimes called higher order features. So we extract these features or descriptors of the image, such as parameters from the image intensity, textures. Features could also be as simple as volumes or shape parameters, but in general, texture features are mostly meaningless for a human, but they carry some very specific information present in the image. Radiomics is typically applied to subregions and quantified by segmentation, where it may be representative of or associated with necrosis, oxygenation levels, vasculature and the like. Here we summarize the process of patient outcome prediction using radiomics. At the top, we have the imaging step, the segmentation step, followed by feature extraction and followed by feature analysis. Features are then used to generate a model using machine learning methods. Once trained and validated, the model can be used, new patients can be fed into the model and an outcome prediction can be made. In any outcome study, it is important to control the way we arrive at the features, which are subsequently used in a prediction model, and standardization in these steps is important. Radiomics analysis can also be done using deep learning. Deep learning is a very powerful machine learning technique that can work with the entire image and that can, if enough training images are available, be used to recognize patterns. We apply deep learning to predict outcome based on images and clinical parameters. On the left, an example of a deep learning convolutional neural network is shown. That takes image information from CT scans, PET scans and clinical inputs such as clinical tumor staging of patients with in this case, head and neck cancer. All of this information can be used to generate a score that allows a classification of an outcome of interest. For example, disease recurrence, distant metastasis or overall survival. The bar plot on the right shows the metric AUC that quantifies how sensitive and how specific the prediction of, in this case, overall survival is. We can see that if the model is trained on patients with all three types of input variables, CT, PET, as well as clinical information, then we can predict to high accuracy the overall survival. When one of the modalities is not present, the prediction capability decreases. More in the radiology space, texture-based analysis plays an important role in diagnosis. Here, I highlight a study on the development of a tool called Cut for Kids, which is a collaboration between the diagnostic imaging analysis group in Delft and Netherlands and a group in Johannesburg, South Africa. The goal is to develop a completely automated computer-assisted diagnosis tool for fast diagnosis of pneumonitis in children based on classification of imaging characteristics of the lesion as assessed using texture analysis. The algorithm also indicates the region where suspected texture abnormalities occur. On the left, the algorithm successfully classifies the severity of the lesion. A color representation of the area where the lesion is detected allows physicians to fast review the images and validate the prediction model. The curve on the right-hand side shows a measure of the combined specificity and sensitivity of the method. An AUC of 0.85 is a respectable performance and certainly helps in speeding up diagnosis in situations where a large number of images need to be reviewed. Ensuring reproducibility and reliability of an AI tool based on medical imaging data is paramount. There are many steps involved in the construction of the AI model that hinge on the quality of the image acquisition parameters, the feature extraction, the processing steps and so on. The current quality assurance tools that we have in radiology imaging may not be sufficient. Quality assurance is a comprehensive process that requires performance metrics and feedback. With the development of new AI algorithms, the quality assurance cannot be an afterthought and needs to be built into the design of the tool. I highlight a paper on the right, discussing this in more detail. I have tried to briefly highlight a few areas of application of AI in medicine, focusing on cancer diagnosis and treatment using radiation. The benefits of AI in those areas are numerous and include the reduction in workload of the healthcare workforce, enable quicker detection, treatment planning and treatment delivery, therefore allowing for higher patient throughput. In radiology, AI can enhance diagnostic infrastructure in resource constrained settings and alleviate the burns related to pathology tests or quantify disease progress and recovery and informed adaptive treatments. But there are also warning flags that we should raise and these include, first of all, the fact that AI deep learning models are not always easy to interpret and have a black box nature. Medical decision taking is still the responsibility of the medical professional, not that of an algorithm. We might see adaptations in activities of healthcare professionals, but the overall result should be better healthcare. Secondly, trained machine learning models are prone to quality of the data used and require a quality assurance step before they can be used in a particular setting. The design of this QA program must be thought out at the time the model is developed. And thirdly, I have not discussed the ethical aspects of collection and systematic storage and access to data, but this discussion will be highlighted by another speaker later in this session. There is a need for educational materials and guidance in this field and we are pleased to see that the IEA will undertake a role in this. Thank you for your attention. So thank you very much for that. So the next question is which AI tools are or have already entered the Radiation Archeology Clinic and which will take much longer to become available. Given the quality and quantity that we were listening to that's needed to get a good output, what are the main challenges that you see? So there are a number of commercial tools that have already entered the Radiation Medicine Clinic, known under the broad term of knowledge-based planning or automatic planning. That's in the space of Radiation Oncology and this has been happening over the past three to four years. Essentially these are tools that utilize some form of statistical learning to pre-select beam direction in external beam radiotherapy and that generate optimization criteria that are used in the determination of the correct beamlet, what we call beamlet weight, so the optimization process of the planning. More recently another time consuming aspect of planning is being tackled and that's known under contouring or delineation and that is being tackled using deep learning and these are currently being tested by different research labs and by manufacturers. One can expect that these tools will enter into clinical treatment planning very soon. Other algorithms that are being tested that perform logistical tasks rather such as scheduling patients on the machines, on the chemotherapy chairs, make physicians and physicists cover schedules and the like, but those are tools that are also known from other fields of application machine learning tools. What will likely take more time to materialize is predicting patient outcome using machine learning and AI and the inhibiting factor here is the data. So the data is not conglomerated or is not federated sufficiently enough at the moment for these solutions to come quickly to fruition. But in principle one could think of real model predictions that can inform and generate adaptive treatment plans based on outcomes rather than on radiation dose as a physical quantity. So optimize the dosage based on outcome rather than rather than clinical knowledge of the door that is required and that feel that feel we call outcome based treatment planning. But that is further in the future and hinges on the data, the availability of the data and the availability of the data is a complex process is a complex aspect. It's not just technical, it's also ethical. That's that's a very important point and the the IA has international coordinate research projects and now is developing a database and this may be something that could feed into that. But you mentioned about education and you mentioned the need for the support from from the IA for that. In your view, what role can we play in a safe and responsible dissemination of AI technology specifically in radiation medicine or radiology departments and what specific actions you think would really be helpful other than the education that you mentioned. Right. So as you mentioned, so I think one of the prime roles that the IE could play is to provide educational documentation information and training with the emphasis on developing guidelines and standard operating protocols for commissioning and safe operation of AI algorithms in radiotherapy and radiology. So that's not an activity that the IEA will be alone in there with this other organizations also working on this and but the IEA has always been played a consolidating role in generating protocols for for example for radiation dosimetry and and thereby playing a leadership role in particular. Currently, I realize that this might not be a specialization of the IEA, the AI applications, but for example, medical physicists which work very actively within the AI are very well placed to play an important role in this and the AI is a track record of working with within the international community including the WAPM and other organizations to standardize processes and protocols and play a role in quality assurance and auditing. And this brings me to a second point. The IEA is generally very well connected to the cancer clinics worldwide. It has a reasonable idea of the reasonable pulse on the needs of the radiotherapy community through its extensive databases and this will help in disseminate information with full awareness of local needs. And as I mentioned before, I see the role of the IEA not necessarily to be on the front of the new development but on the consolidation of new science when it comes to when the rubber hits the road and when it's actually going to play a heavy role in a clinical environment. So that's more or less what I wanted to say to this. Thank you very much. Talk to Melissa. Okay. Thanks very much. It was very interesting. Before I introduce the second speaker, I've been told that some participants are having audio issues. So try to use a headset or reset your audio or you might try joining your mobile. Note, however, not all is lost. The slides themselves have a huge amount of information and we also will be publishing the event later on this week. Okay. So moving forward, I would. It's my pleasure to introduce our second speaker today. Mr. Clement Batalla. Mr. Batalla is assistant professor in the Department of Earth and Environmental Sciences at the University of Ottawa in Canada, where he established the SAE group. This group works on developing isotope geochemical tools and numerical methods for provenance tracing and ecology and forensic application for understanding weathering processes across regional to global scale and for reconstructing paleo environments in particular greenhouse gases. Mr. Batalla will talk about the potential of artificial intelligence supply to pick data in isotope geochemistry. Isotope sciences entering the big data around thanks to global collection networks and repositories, such as the GNIP database in my own division, which is gathering global precipitation data for over 60 years. Using AI for such big data is becoming critical to obtain interpretable results and to enhance our understanding of the environmental and hydrological implications and processes involved. This emerging field offers huge potential to solve large scale issues ranging from groundwater sustainability to atmospheric pollution. So, Mr. Batalla, I'm looking forward to your talk. Hello everyone. So I'm Clement Batalla from the University of Ottawa. And so I will be talking today about artificial intelligence for large scale mapping of hydrological environmental processes. So when I study environmental and hydrological processes, I use a tool that's called stable isotopes. And to explain you what is a stable isotope, I've put that there at the periodic table. And if you look at it with a little bit of kind of an artistic eye, it does look like a piano keyboard, right? So if I always think about it as a piano keyboard myself, if you were to be able to play this piano keyboard and understand how all those atoms were interacting with each other, then you will be able to understand this full symphony of the universe. What's fantastic about being an isotope geochemist is that we don't see this piano keyboard as just a piano keyboard, but as an organ keyboard, like more of a 3D piano keyboard. Because each of these different elements has multiple isotopes or multiple forms that can tell us something about environmental processes. Why is because each of those isotopes actually distribute in the environment and have a certain abundance in the environment as a function of physical, biological or chemical processes. So if we start to analyze different substrates in hydrology environmental science, each of the substrates is going to have a different composition or abundance of isotopes. That is going to tell us something about either the source or the process that led to the formation of the particular molecule we are interested in studying. So now what's really useful in isotope science when we start to collect data at a very large scale is to start to map how isotopes vary in the environment. Because it's time to tell you some key answer about environmental and hydrological processes. And here I'm going to give you a few examples with some important studies that happen in the field to give you an idea of how we can use isotopes to understand processes. So here is a study that was looking at hydrogen and oxygen isotope variation in tap water across the western US. And as you can see here on the left map, we mapped the variation of the stable isotopes on the landscape. And what we figured out is that this pattern is telling us something very important about water management practices in the western US and how actually this water management practices are impacting water losses. A second study here is a time showing some people went and collected a bunch of plants all over the savannahs and started to look at nitrogen isotope in those plants. And they found very different patterns of nitrogen isotopes across the savannah. What they figured out is both nitrogen fixation and termites have a huge impact on how organic matter is cycling in the savannah and actually on the sustainability of ecosystem in the savannah. Here's another study, completely different. Here's a study where let's use transium isotope to start to look at how we could track the migration of Pacific salmon across watersheds. And what was really interesting about this work was not only we could trace salmon migration, but we could start to figure out something about how to manage their stock. Because transium isotope could help us figure out both those production sites of the Pacific salmon on the watersheds, spatially, but also how those production sites will vary year over year. And here's the last study that looked at hydrogen and oxygen isotopes across a storm through time. And this storm was the sandy, big storm that happened in 2012 in the United States. And what we could figure out by looking at the pattern of isotopes in this storm was something about how the moisture sources of that storm changed through time and how actually this storm evolved and managed to go through the continents. And so that could really help us kind of forecast some idea of the strength of that storm in the future and could be used in hurricanes or other applications. So as you can see, a huge range of possible application of mapping stable isotope on the landscape for environmental or hydrological knowledge. So now, why do we want to use artificial intelligence and machine learning and isotope data? Well, the reason is that isotope data are kind of entering this new realm of big data science. Why am I saying this is that if you start to compare here on the y-axis a number of publication that happened in stable isotopes and DNA and genomics, you can see kind of similar pattern. You all know that DNA and genomics kind of had an explosion in the 1970s, and this led to all kind of the new genomics revolution that we are seeing now. All the technology advances we're seeing now, like DNA sequencing, massive database of human DNA, advancing medicine, advancing agronomy. So what you can see is stable isotopes following kind of similar patterns delayed by 20 years, but will have a really massive sort of impact on society, both on the hydrology and environmental science, but also in all sort of other fields like forensics, like paleo-ecology, like history and anything. So what's happening is we're doing a little bit what happened with DNA and sequencing. They in the 1970 constructed this database that's called NextGen, and that kind of led to all those advances in genomics. And so we're doing the same thing with stable isotope. The community came together, and I'm involved in this project. It's called the IsoBank project, and it's a massive database that will gather all the stable isotope data that have been produced in the last 50 years and will be produced in the future. So now we have this massive unified repository. We can start to apply to this some big data technique, which is artificial intelligence and machinery. So in addition to the explosion of data in isotope science, they've also been an explosion of data in general in hydrology, environmental science. And so here is kind of a schematic showing that basically the volume of data that we are receiving in hydrology and environmental science has increased dramatically. The variety of data also is increasing dramatically, and those data are totally different scales, right? You have data at the local scale, at the plot scale, you have data that you obtain at the catchment scale, but you also have global remote sensing data that you can integrate in your analysis. All those data have different uncertainty, and so how do you get knowledge from those data, right? So you need some of this algorithm of machine learning and artificial intelligence to help you integrate all those data together, to try to get small and digestible information from those data, something integrated across disciplines, something that you have some confidence in because you have an uncertainty analysis associated with it, and that's how you gather knowledge from those data. So this is where, in my opinion, artificial intelligence and machine learning are coming into the picture. Really in the past we did not have very good algorithm to integrate those data that were coming from different sources, that were different type, that were different scale together with isotope data. So let's say you had an isotope data set, a point data set of isotope across a regional scale, and a bunch of covariate that you wanted to use to predict those isotopes and try to explain those isotopic variations in terms of processes. What you needed to do is to do a lot of preliminary work to prepare your covariates to not be redundant, to prepare your target variables to follow normal distribution, not have too many outliers, and then only then you could just integrate all this into some sort of regression framework using linear or nonlinear models, and then make some prediction. Now what's nice is that this process is really simplified with machine learning. You have your target variable and a series of data points with isotope data, you have a bunch of covariate that can be of any form and any type, but you can integrate all this together relatively simply into a machine learning framework and provide prediction. So really machine learning, in my opinion, has been very, very useful in integrative data from multiple sources and scales and limiting the amount of work to do when you do data science. So I don't have the time in this talk to just cover all the type of machine learning and artificial intelligence. Algorithm we can apply to isotope data and hydrology and environmental science, I'll stick to one random forest regression, which is kind of my favorite algorithm those days. And so you probably all know what is a regression to start with. So regression is just the idea how you try to predict the value of target variables with some sort of predictor. And so the most famous sort of regression is a linear regression, which is just trying to predict this target variables using some sort of linear relationship with a predictor. Let's say you're trying to predict the grades of students, likely if you plot this as a function of the number of hours those students have worked, it's going to plot in some sort of line. The problem with this is in the natural words, a lot of the relationship are definitely not linear. They tend to be nonlinear or they tend to be tracial-based, they can be periodic. And so all those sort of relationship are really hard to integrate when you start to use kind of typical regression algorithm. That's where machine learning became really useful. Because what machine learning does and particularly random forest regression in this case, it takes your data sets, all your entire data set with all the covariance that you might want to integrate that we've spoken about before, and it has one target variable, the isotope data. And so what it's going to do here, it's going to take all your possible predictor and split them into what we call trees. And so it's going to test basically the value of your target variables and try to narrow it down as a function of the value of a series of those predictors or covariate that you're trying to integrate. And it's going to do that in a series of trees in parallel. And then the power of random forest is that it's going to take all those trees and make them vote or make them give you an answer on a prediction and aggregate all those predictions together by calculating the mean of this prediction. And that's really helped you to model in particular those kind of nonlinear, tracial-based or special-based or periodic sort of relationship between your target variables and some other environmental predictors. So here's an example of using isotope machine learning and all sorts of other environmental covariates to try to advance the field of environmental science in this case and global geochemical cycles. So here's a study where I worked on mapping global strontium isotope in plants. So I use global database of isotope value in plants. I use a bunch of geospatial covariates that I knew probably had an influence on this particular isotopic system. And I use a random forest regression to try to predict how this particular isotopic system was varying on the landscape. So now what's interesting about this model is that it not only provides you with a model that's extremely precise, like an order of magnitude more precise than previous algorithm that we could use, but also it provides you with a somewhat interpretable framework. It really helps you to understand what are the predictor of this isotopic variability and knowing those predictors really helps you to understand the process that's behind this isotopic variability. For example, here we knew that strontium come primarily from geology, so we knew that the primary source of strontium to ecosystem at the global scale would be rocks. And you can see that here on figure C when you look at the strontium being influenced very strongly by geological variables. But what we did not know, at least we didn't know how to parameterize, was that strontium isotope was also influenced by the inputs of aerosols, atmospheric aerosols either from sea salt or from dust deposition that you can see in B and D here. And those are processes that not only influence strontium isotope, but influence any sort of metal over ecosystem. For example, you all know that the Amazon forest would not survive without the depositions of dust aerosol from the Sahara Desert. And so really machine learning here at the global scale can help us quantify those processes, quantify the amount of this metal that are coming from different sources in ecosystems. So here's a second example of using machine learning and isotope data to answer big scale hydrological question in this case. So here this group mapped isotope, hydrogen isotope variation in groundwater across the United States using a random forest framework very similar to what I showed earlier for the strontium. And they compared this map predicting hydrogen isotope across the US with a map of hydrogen isotope in precipitation monthly. And what they could do is they could get the map that you get now on the right showing the timing of recharge of these groundwater across space. Why could they do that? They could show that if the groundwater isotope value looked more like different months of the precipitation, then it was likely that the recharge was coming from this particular month. So you can see, for example, that the Rocky Mountains get most of their recharge during the winter rather than the midwives gets most of their recharge during the summer. So again, something very useful, this sort of map, is extremely useful in terms of figuring out water vulnerability, water sustainability, how are we going to manage this groundwater resource over the long term. So I'll conclude this talk on just what I see as the future of artificial intelligence and machine learning in isotope science. So I think this is really an emerging field and something that's going to grow in the future tremendously. There's a lot to do with isotope data in hydrology and environmental science and using machine learning. I will compare the process to drinking a good wine, really. I'm a French person originally, so my parents used to be a winemaker, so drinking good wine is something important to me and making good wine is even more important. And so the process of artificial intelligence is a bit the same because like making a good wine, you need to know your grapes well. Well, in machine learning, you need to know your data well. And so doing machine learning doesn't prevent you to really know your data very well and to know what you're putting into your model. But once you know what you're putting into your model, it goes through this kind of complex process and complex maturation through the machine learning algorithm. But what you get out of it is fantastic. You get information that you never knew that were present in your data. You get information that are really interpretable, very easy to visualize. You get, you know, very enhanced prediction precision in your models. So it really leaves you with a good taste in your mouth when you actually do the process well. And what I'm thinking is, you know, as physical scientists, we tend to always want to understand the word in terms of process and to improve our mechanistic models. And that's where I think artificial intelligence is coming to help us. Using the big data, it can give us new knowledge about processes, but it can also complement those mechanistic models and fill kind of the blanks when we don't know a process. For example, when we try to predict the future of the planet due to increasing CO2 level, well, there's a lot of fine-scale processes we don't know how to model on the physical world. And that's where artificial intelligence can help us parameterize things much better. Thank you very much for listening. I hope you have some questions for me. Yes, thank you very much. That was very, very interesting. I enjoyed your talk immensely. And due to time, I'm going to summarize into one question, a couple of questions that came and were put forward. And so I'm going to ask you to listen very carefully, because I want you to summarize a specific project or application in your research that you think best exemplifies the utility of artificial intelligence and machine learning to actually reduce the potential for bias specifically. So if you could give an example. It would be wonderful examples to show the breadth of applications you pointed out the ability of AI and machine learning to integrate different scales, which is very good. You've demonstrated isoscale data for strontium and also for water. But the questions coming up is, is there an example to reduce potential for bias that you mentioned in your talk? Is there something to answer that question? Yeah, I mean, I'm not really a mathematician. So I felt that's more a question for someone who actually he's developing machine learning algorithm there to just try to improve or reduce biases. And I think Random Forest is a really good example of this in some ways, like it was a machine learning algorithm that was tree based algorithm, but that we're creating a lot of bias and kind of over prediction on this algorithm. And so machine learning was developed then after that to kind of remove some of those biases by just like doing some bootstrapping of the trees and preventing some of those biases. So I feel like it's more on this, on the mathematical side of the algorithm that's kind of reducing the bias comes from, I think also, and that's something that I mentioned in my conclusion that knowing your data well is something that's going to reduce the bias also dramatically. So I mentioned this isobank database knowing the data quality that you are putting in your algorithm, knowing how it's distributed, for example, spatially, and what's a data quality process is also something that's going to, I think, tremendously help to reduce any of those biases that you might find at the end, that's how I can answer to this, I think. So thank you very much. You're mentioning the isobank project raises another thing, the challenges of accessibility of data and applications. So what other limitations or challenges in machine learning applied to environmental sciences do you see? Well, I see, like, as I said, I mean, again, that's a huge growth of data that we have in environmental science. I think the biggest challenges these days is just, you have all this data, all the skills, and it's becoming enormous. And then you're putting that into some sort of black box machine learning, artificial intelligence algorithm. So I kind of, as I said, knowing your data well, so you can get very quickly overwhelmed and you get very good prediction at the end. But really, you get a model that's absolutely uninterpretable because you're putting so much things into your model that it becomes very difficult to explain anything. And I think so, really, I see the power of machine learning and artificial intelligence in environmental science in the interpretability of those processes. Otherwise, it's just, there's no points of not using just a regular multivariate linear regression to explain some processes. So it needs to be so much interpretable. It needs to give you an answer of the processes that are behind to just help you then do management of resources. So that's what I see, just like kind of integrating all the skills, all those data into something that helps you understand natural world processes. Yes, thank you very, very much. We'll have an opportunity for some more questions at the Q&A at the close. Thank you. And we'll move on to our third speaker, which is Mr. David Humphries. Thank you. Thank you. Mr. David Humphries is our next distinguished speaker. He is a principal scientist and head of the plasma control group at the D3D National Fusion Facility in San Diego. The D3D control group specializes in tokamak control physics and real-time systems, including applications of control mathematics and machine learning to solving critical problems in fusion energy and beyond. He is the recipient of the 2017 IEEE NPSS Fusion Technology Award for contributions to plasma control and coordinate several international efforts to advance solutions for the international thermonuclear experimental reactor, ETER, and other superconducting tokamak. Mr. Humphries will highlight the potential of artificial intelligence for enabling prediction and control solutions necessary for sustained, safe, and efficient fusion power plant operation in the future, as well as the opportunities and associated needs in AI areas that would help address challenges in fusion research to target collaboration. Thank you, David. I'm looking forward to your comments. I'm Dave Humphries from General Atomics in San Diego. Today I'm going to be talking to you about advancing fusion research with artificial intelligence. I'd like to start by reminding you of a few features of nuclear fusion. Fusion produces energy from isotopes of hydrogen, in particular the heavy isotopes of deuterium and tritium. This is favorable because the efficiency is about a thousand times that experienced in terms of energy per mass for a fission reactants. But the gas that you create out of the deuterium and tritium must be very hot in order to overcome the Coulomb repulsion so that the nuclei can fuse together, producing helium and a neutron. In fact, the temperatures tend to be greater than 100 million degrees Celsius. This means that we produce a high-temperature ionized gas, which we call a plasma, and the challenge is to confine this plasma with temperatures really higher than 10 times the temperature of the core of the sun. Fortunately, plasmas at those temperatures are highly conductive and they can be confined by magnetic fields. As you see on the right in the cylindrical configuration with the magnetic field, the particles will orbit around the field lines and thus minimizing loss in this direction and contact with the wall. Of course, in a cylindrical configuration, they will lose particles and heat from the ends. So the solution is to create a toroidal configuration. The leading configuration in the world of fusion energy today is the Tokamak in which the magnetic field confines the plasma and stabilizes instabilities and indeed is self-heated by fusion reactions. You can see here on the right, the simulation of a particle orbiting around those field lines and indeed staying away from the wall as it needs to. Now, key features of a Tokamak include a high degree of active control, the need to drive large currents in the plasma, active regulation of the plasma state instabilities and indeed prevention of particular instabilities that could disrupt the plasma, terminating it suddenly. You have to actively regulate the fueling and the burn state and the pump out of the waste helium. ETER is a fusion reactor with net power gain that will address the remaining challenges or many of the remaining challenges for fusion energy. It nowadays used to be an acronym once in a time, nowadays it stands for the way. It is the Latin word ETER, the way to fusion energy. But ETER is indeed under construction in France and first plasma is planned for 2025. Its mission is to produce more than 10 times the power that's put in so it's quite a challenge. There are some key research challenges on the way to an economic power plant. High confinement and fusion power output being able to sustain that high power for something on the order of a year to extract and tolerate high heat fluxes, control those instabilities I mentioned before with very high reliability and produce a high breeding rate for the tritium fuel which does not occur naturally. Now machine learning and AI can provide some very powerful methods for extracting knowledge from large scale databases and this is attractive for addressing those challenges that I just mentioned. I'll briefly mention a few of the well known approaches in machine learning just to set the stage for more details in a moment. Transfer learning refers to a set of approaches that allows one to bridge gaps in existing theories and in particular create hybrid models between theory and data driven knowledge. Also one can extend an initial model with incremental data. Classic example is recognition system for cats based on a database of cat pictures which one fuses with a data set of dog pictures to create an extended performance system that can identify both cats and dogs or just dogs in fact. Unsupervised learning refers to methods that allow one to discover trends and dependencies in a large data set in particular identifying classifications within the data. Supervised learning and surrogate models allow us to encapsulate physical behavior with often compact models. Structured equation parameter identification allows one to identify coefficients and parameters in analytic descriptions. These are just a few of the approaches but let me get more specific about how fusion can be helped by machine learning in its science discovery mission. So by science discovery we really mean extracting maximum knowledge from data bridging gaps in the knowledge using data and creating models from data. So when you're doing that it's important to keep a focus on interpretability. That is the ability to identify the mappings in particular between certain input parameters and certain output parameters. So this for example would allow you to map a control path for regulation of this output parameter and identify these two quantities as the key knobs that one must regulate. Uncertainty quantification is another area that's critical really for all results but in particular for predictors and knowledge discovery that will lead to control. Imagine that you have a growth rate that you are estimating in real time with a machine learning algorithm and you need to avoid, you need to keep that growth rate from getting close to this death zone that this red dashed line implies but if you're gonna do that you need to know what the error bars are on your calculated value so you can keep the error bars indeed away from that death zone. So for control uncertainty quantification is critical. Also validation measures and the ability to quantify the space of validity and extrapolability are important. Another area that machine learning can help in is for fusion prediction and control. So for example, predicting the likelihood of a disruptive instability. So on the right I show some examples from an actual experiment and a machine learning trained algorithm. The discharge is represented by the current in this black curve and the blue curve shows a probability of disruption which you can see at the end rises up above the threshold and indeed the plasma does disrupt. And here you can see the result of interpretability features extracted for control in these two quantities are the largest contributors to the disruptivity. A brief comment on reinforcement learning which is a popular approach to developing controllers for fusion. It's very successful in domains that can be completely characterized. Well known are the successes it's had in the games of chess and go and a variety of video games by basically characterizing your environment perfectly and feeding back a reward and state information to a dynamic agent that then takes actions in your dynamically evolving space. One can make very, very effective controllers. But again, only in completely characterized domains and that's a limited solution set without detailed performance guarantees uncertainty, quantification, et cetera. So it's important to keep aware of that. So the US Department of Energy has been assessing the potential for machine learning to contribute to this research over the last several years. And there are several workshops over the last three years and the most recent is one on advancing fusion with machine learning just last year. And that one identified several research opportunities including science discovery, control and prediction that I've just mentioned as well as the need for connections among domain experts. More recently, an IEA technical meeting in July of this year focused on plasma disruptions and mitigation and identified the status of the field and opportunities for IEA-eater coordination. So it focused on disruptions that I mentioned before and reviewed predictive capabilities including machine learning approaches, the effects on Tokamax of disruptions and the status of mitigation methods. So this was a virtual meeting with MP4 format for the recorded talks which were reviewed separately and then discussion sessions were held. So this workshop had several results. I'm going to highlight a couple. Machine learning disruption predictors are now seen as increasingly accurate and able to guide control strategies and there's a strong potential for gains from increased collaboration and coordination of researchers. In the planning stages is an IEA coordinated research project or CRP on control oriented solutions for plasma disruption prediction and avoidance in IEA-eater. It focused on the need to limit the number of disruptions in IEA-eater and therefore the need for accurate prediction of those instabilities that can cause disruptions. So the focus is going to be ensuring reliable stability metrics for plasma control by identifying controllable scenarios. Some of the activities will include developing an event database from experiments that lead to disruptions and qualifying these stability metrics for control hopefully extrapolable to Eater scenarios. You can see the role that such knowledge will play in the Eater forecasting system which will of course embed real time assessment of the stability and therefore can make use of these metrics and will also have to of course project the system health overall and map that to the likelihood of disruptions. So in summary the growth and success of machine learning and motivated international assessments of their potential for advancing fusion energy science and I highlighted the U.S. Department of Energy workshop on this topic. A recent IEA technical meeting co-organized by IEA and Eater and in particular Mateo Barbarino and Mikhail Lennon which identified the potential for gains from coordination of cross-domain researchers applying these methods and in the planning stages IEA coordinated research project being led by Mateo Barbarino on control oriented solutions for disruption, prediction and avoidance in Eater. And with that, thank you very much for your attention. Thank you very much David for your interesting talk. The end sustainable plasma is the alpha and omega of any fusion future and I'm very hopeful that these activities in machine learning and artificial intelligence will help sustaining that goal. So thanks for highlighting though. But I do have a couple questions I have to keep them rather short because we are actually going over time but we saw in the previous speaker how they're creating like a super database of some of their stable isotope data and so in your opinion what modality is there in an access among our member states would best enable use of artificial intelligence to address fusion problems? Yeah, indeed one of the most useful mechanisms we think to facilitate cross-member collaboration has been identified as a common fusion data platform perhaps a distributed or federated network of access points really to enable a larger set of collaborators to gain access to curated data that would really be in the right format to enable AI machine learning applications and that would also include tools that would facilitate interface and visualization and of course to assist in the ultimate machine learning analysis and model building. So we think this is one of the most important steps to take to reach a wider audience and wider collaborative community. Now there's been a more general question posed about are there optimal combinations of traditional machine learning approaches to solve fusion problems or the research you're doing you've identified in your talk as a stand alone or is this more generic? Yeah, this question has arisen many times I think it's a super important one I see it in the chat as well people are asking about does machine learning really eclipse the time honored and very powerful multivariable regression methods and other statistical inferences and it really doesn't and one of the our challenges is in fact to keep in mind the optimal merging of these kinds of approaches those that are more or less traditional and not highly computationally intensive and then the machine learning tools which exploit greater computational ability and such things as GPU hardware that can accelerate computations and indeed ways that are new for fusing data because of these new capabilities but it's really important to have an optimal sense of where all of our mathematical tools fall in our bag of tricks Thank you very much I now hand over to May to introduce our next speaker So our next speaker is Emma Redkamp-Lew who is professor of philosophy in the department of philosophy faculty of humanities at the University of Pretoria she's a leader in ethics of artificial intelligence research group at the center for artificial intelligence research in South Africa and a member of the UNESCO world commission for ethics of scientific knowledge and technology commenced for the period 2020 to 2023 she's also chairperson of the bureau of UNESCO ad hoc expert group on the ethics of artificial intelligence test to prepare recommendations for a global instrument on ethics and at this point she will discuss the ethical aspects and challenges of AI and the implications for education and the role of AI in shaping human futures we see in many times the technological advances and models move faster than the discussions that should be had beforehand in ethics and others and we're looking forward to hearing from you Good afternoon The potential of AI to contribute positively to the protection and improvement of the lives of all living beings and after the environment and ecosystems is almost impossible to exaggerate we have just seen three examples of the positive potential of AI in the health environmental and energy domains but there is a darker flip AI systems raise new types of ethical issues that include but are not limited to the impact on decision making, employment and labour, social interaction healthcare, education, media freedom of expression access to information, privacy democracy, discrimination and weaponisation furthermore new ethical challenges are created by the potential of AI algorithms to reproduce biases for instance regarding gender, ethnicity and age and thus to exacerbate already existing forms of discrimination identity prejudice and stereotyping AI systems play a profound role in a human practices and society as well as in their relationship with the environment and ecosystems creating a new context for children and young people to grow up in, to develop an understanding of the world and themselves to critically understand media and information and to learn to make decisions in the long term, AI systems could challenge human special sense of experience and agency raising additional concerns about human self understanding social, cultural and environmental interaction, autonomy agency, worth and dignity so what can we do almost 100 documents on AI ethics have been generated over the past decade or so including intergovernmental policies such as the recommendation on the ethics of AI that I have just referred to the new UNESCO document national policies many countries are now working on delta protection frameworks for instance professional policies such as those generated by the IEEE and other professional bodies policies generated in the private sector for instance Microsoft's open AI project but at present it is clear that this collection of documents has not yet had any real practical impact on the ground on the contrary the current situation is leaning itself to the exact opposite of what is intended for instance ethics shopping, shirking blue washing, lobbying and dumping I'll give you an example of some of these, ethics shopping simply means you pick and choose from the list of 100 documents which one will cause you the least trouble shirking and blue washing have to do with again doing as least as possible instead of creating public accountable evidence based transparency on best practice lobbying simply means companies lobby for self-regulation as opposed to regulation from outside of governmental regulation and ethics dumping has to do with exporting unethical research practices to companies with weaker laws than one's own reasons for the non-impact, what are those? They include the lack of mechanisms AI ethics has to reinforce its own normative claims, the V of AI ethics guidelines is coming from outside the technical community so alienation from the ideals of AI ethics the lack of distributed responsibility the lack of knowledge about long-term or broader societal technological consequences causing software developers to lack of feeling of accountability or review of the model significance of their work which leads also to feelings of alienation from AI ethics so simply creating a plethora of documents is not the answer I will argue in the coming slides that a participative model for actionable AI ethics may have some kind of answer for us to move forward first in order to approach such a model we need to cultivate an atmosphere of shared responsibility among all AI actors where AI actors can be defined as any actor involved in at least one stage of the AI system lifecycle and can refer both to natural and legal persons such as researchers programmers, data scientists so end-users, large technology companies start-ups, universities public entities and so on so to cultivate this atmosphere of shared responsibility one way to do this is to involve civil society by focusing on what is called intrinsic values intrinsic values are values that have worth that is inherent to them for instance the value of peace the value of honesty, the value of justice promoting AI fundamentals, AI ethics and media and information literacy I will come back to this point a little bit later when we talk about education and setting up multidisciplinary research as a necessary counter to a culture driven by commercial values and feelings of alienation from AI ethics among members of tech communities how can we do this recognising that the multidisciplinary explosion of the scope of the discipline of the ethics of AI firstly reflects the potential impact of AI technologies on human societies and political stability in a manner to which the technical community is more open as it is more concrete to them because it includes them in various roles and secondly this kind of multidisciplinary attitude should inform AI ethics practices so the multidisciplinary explosion of the scope of the discipline of the ethics of AI refers to the crystallisation of this discipline into many sub-disciplines such as data ethics, machine ethics, robot ethics neuro ethics, information ethics and so on and in each of these sub-disciplines we have members of the tech community involved together with philosophers, anthropologists, sociologists political scientists, lawyers and so on and in other words the work done in these sub-disciplines is more familiar to the tech community than the abstract ideals so we want this kind of multidisciplinary work to inform AI ethics then also I argue that we need an addition to cultivating this atmosphere of shared responsibility a comprehensive participative model of AI ethics that is built on responsible interconnected participation of all AI actors that is adaptable to advances of AI technology in its own political context and that allows every AI actor to manage their own model sensitivity on a continuous basis this can be done within a virtue ethics approach virtue ethics as opposed to the ontology which means that the right thing to do is prescribe to us by some universal law or to consequentialism which prescribes that the right thing to do is the consequence of our actions virtue ethics actually points us towards a lifelong journey of striving to be the most virtuous person that we can be for Aristotle, he is the father of virtue ethics virtue is the result of training and habit so every person is actively involved in their own and society's moral growth a virtue ethics approach to AI ethics is thus sorry, typo not focused on universal codes of conduct or abstract guidelines individual level at which everyone in society has a duty to ensure that they themselves as well as everyone else are able to be or to become the best possible moral version of themselves moreover, virtue's actions involve rational deliberation rather than some form of external justification such as universal laws or the consequence of one's actions such as commercial gain and individual members of the tech community are addressed not technology itself and it's not just individual members of the tech community but also all AI actors that are addressed and not the technology and this means that the tech community can maybe thus be motivated to steer technology in positive directions because it's them that are on the line as well as end users and all the other AI actors their companies and so on in addition the focus is on socio-technical virtues such as honesty and justice and thus approaching AI ethics as a virtue ethics brings together the focus on technical discourses as well as the genuinely social and personality related aspects of adhering to AI ethics guidelines but here is an added warning it all sounds very good but no one should simply accept based on false sense of hubris or misguided sense that humans are good and technology is bad that humans and their interest will always be protected actually humans have to ensure this in a responsible manner for one humans will have to lead much more technologically aware and responsible lives and for another every person needs to rethink what it means to be human in societies where AI technologies have increasing control over what we can or may do and achieve and how we interact with each other this responsibility on every human to become involved and to participate in debates on AI ethics echoes in the educational, scientific, cultural and communication and information context of our lives and we will just stand still a moment in the educational context education is somehow caught in this whole process of AI technology advances caught in the web of interaction and adjustment among society, the workplace governments and industry and there are huge challenges and partnerships are crucial, cannot stress that enough we cannot be in silos anymore academia on this side industry on that side and that's not how it can work in the future so what is the role of education here willingness to take up the responsibility to become involved in AI ethics is crucial to the success of any instrument of AI ethics and it should be made possible via rethinking the core elements of education this implies a lifelong learning and transferable skills model where AI fundamentals are taught so coding and digital skills computational thinking is taught and this is often overlooked computational thinking relates to being able to formulate problems in such a way that they can be solved of a computer and other tools so this implies logical thinking, algorithmic step-by-step thinking, the ability to represent data mathematically and being willing to find different solutions to the same problem these two of AI should be taught at the moment the focus is on data driven AI but before the 2000s we were mostly focused on logic based AI and maybe in the future we will move towards a merger of the two again AI ethics skills it is necessary to sensitize everyone to the possible ethical and social and economical consequences of AI technologies and future of work skills, transferable skills, critical thinking, collaboration, communication creative thinking, the foresees and then lastly media and information literacy if we think of how our social narratives have been changed already by AI algorithms we really have to make sure that children from a very young age know how to go about access to information and processing information so in a world in which everything has become fluid we have to ensure that the ethics underpinning it is dynamic and actionable and that the education system driving it is adaptive living in harmony and peace with AI technologies and ensuring she benefits of AI for all imply at least firstly recognizing that the impact of the interconnectedness of AI technologies with human lives should be considered as a continuum of social political, environmental, educational scientific and economic concerns and never only through commercial or political lenses secondly working with the participatory model for AI ethics which is inclusive of all AI actors based on a virtue ethics approach and underpin by multi-disciplinary research and collaboration and in the last place rethinking education and training models on all levels ideally every AI actor should hold every other AI actor ethically accountable and this needs to happen within overarching legal frameworks guided by international law and human rights law principles and standards such that enforcement and reviews become possible thank you so thank you very much for that overview and of the activities and efforts that are behind all of these ethics of technology but when we're thinking about that we understand that we have the responsibility to be involved but also how can we ensure that the general public cares about AI technologies and ethics and understand what can actually occur because you listed many of the issues that we see how can we increase the awareness of the general public about this to be able to galvanize that political will thank you May well the most important thing is to is awareness is to raise awareness whether it is through different programs in civil society whether it is in terms of education we need to make sure that we reach every person in civil society and that civil society is basically our secret weapon because the members of social society the end users are very powerful in terms of commercial impact but for them to be able to make decisions about whether they would use certain AI technologies and others not they need to be informed so what is important is that we should understand that AI technologies impact on every human being's life it's not only in terms of science and it's not only perfect also in terms of science but at the moment for instance serious concerns have to do with decision making in financial risk analysis bail hearings, healthcare and diagnosis law enforcement in terms of privacy Jan referred to this in his talk a lot we're basically taking a lifetime of selected personal data we have the right to know whether our data the shelf life of our data would be whether it could be depersonalised can it be repersonalised in the health sector it's very important in terms of the right to dignity and the right to privacy but also in the broad sense for instance of surveillance and the right to privacy and freedom of expression in that context. Media information and communication self understanding and new social narratives are being shaped in terms of by AI algorithms and people lose touch with reality the future of work employment, the use of AI in recruitment does that impact on people's right to dignity again to privacy freedom of expression human relationships our interaction with robots, care robots what does that say about human dignity and work self understanding or social interaction or cultural interaction warfare and health do we want AI to take life and death decisions whether in the context of health which is really positive or in the context of warfare which is really negative but there is huge potential for both and huge debate about both political stability and social justice do we want reproduction of existing structural bias what do we do about this and misinformation what do we do about surveillance the new context in which our children grow up we don't know we have not had a generation that has grown up from birth with a very very very intense presence of AI technologies culture fine arts performing arts and literature who owns the artwork in terms of languages natural language processes a lot of human expression nuances loss damage to the environment and ecosystems and challenge then broadly to human unique experiences we should fight to keep what is intensely human human but we should not think that this is simply a right that we have we should make sure that we actually deserve this right so we should take care that we use these technologies in response and then we should never forget the positive impact all that we have heard already in information connectivity agriculture restoration of the environment and ecosystems have a negative impact on the environment but also a positive impact in healthcare and energy law enforcement and so on so we need discussions about AI at all levels thank you thank you very much so there is a question about the role of governments in promoting ethical AI and whether they can influence the private sector and researchers and whether AI can even affect the scientific method as well right so in terms of governments yes governments should play a very important role but there are the usual obstacles in terms of politics we think about the discussion in the UN on the use of autonomous weapons systems in warfare did not go very well we are now working on a convention on the ban of autonomous weapons systems but these are very sensitive political issues in terms of data protection frameworks that governments should set up does not exist in all countries at all not even in all first world countries so these things are important but what we have to understand is the part of your question referring to private sectors that is one of the biggest concerns that we have on the side of AI ethics about the effectiveness of the role that governments say and also in those terms not all legislation in all parts of the world are equally fast but if we think in terms of what is this big obstacle that I refer to it has to do with multinational corporations so at the moment there is this big standoff in terms of digital sovereignty on the one side we have governments that should protect their citizens and that have the power of legislation but on the other hand we have the private sector and these big corporations these big corporations who have the power of creation the governments want what the big tech companies can give them but they also have to protect their citizens and data moves across borders and data is not owned by anything and it's a big debate in ethics but whether data is a good or whether it is a tool if it's a tool it doesn't belong to anyone but then how to be regulated so these are really important questions but they don't necessarily have very simple answers in terms of science we have all the wonderful opportunities that AI offers us there are also some concerns in terms of data management AI methods and capabilities and the scientific method itself I'll give you some examples just briefly because we're going over time but for instance the whole concept of open data in terms of data management what does it mean in practice with data sets are just too large or too complex for anyone to actually act as an understanding in their entirety so here we really need to work towards creating new data standards encouraging researchers to publish data from metadata the current journals and other data holders to make the data available where it doesn't cost ethical lines in terms of AI methods and capabilities how can researchers reuse data which they have already used to inform theory development while maintaining the rigor of their work traditionally in philosophy of science the classic experimental method is to make observations then come up with a theory against new with new experiments so one is not supposed to adapt the theory to fit original observations theories are supposed to be tested on new and fresh data machine learning we make the distinction between training and testing data but it does not really expensive to obtain or requires different experiments to be scheduled at an uncertain future date is there a way that we can reuse all data so that we all have scientifically valid methods in terms of integrating scientific knowledge maybe a last example how can AI be used to actually discover and create new scientific knowledge and understanding and not just classification and detection of statistical patterns would it be possible that at some point computational methods would have enough to main knowledge built into them that they can then start to make new knowledge and scientific breakthroughs and scientific methods for the rigor and for all of experiments in science which we all hold very dear to our hearts thank you thank you very much very thought provoking and many questions and things that we should answer the society so I think we can move to the general questions and answer session and for this one there are a few questions from that we've gotten on the chat some of them I'll start with Jan and the question is contouring and planning models are built on specific cohorts of patients and equipment as we know so what impacts could there be in terms of like let's say over dosage delivered to patients misdiagnosis etc if the models were applied without adequate training of professionals and the same thing with equipment if we use different cohorts of patients so for example you can use Canadian patient models and cohorts applied to African patient models and what could happen or different machines like 128 slides CT versus 16 slides and you know how do you bridge that gap and how do you make sure that we can make sure that our clinicians, technologists, metaphysicists are well trained so this draws the attention to what we sometimes call the commissioning process so let's say there is a tool that is developed and trained in one institution and it needs to be applied in another institution typically there is a process called commissioning which adjusts the tool or which subjects the tool to quote unquote a local data and a process needs to be in place to retrain the model on the local data test it validate it and then roll it out clinically so that's sort of the general approach to this the person also introduced one of the questions was also regarding different machine parameters which can influence performance of contouring models or the like such as 16 slice CTs versus 128 slice CTs the performance of image based radiomics models has been investigated quite extensively in terms of these type of parameters and also parameters that are part of the machine learning model itself so or reconstruction parameters of the CT router so all of these all of these variables need to be tested in a commissioning process and the commissioning process will also lay out the limitations of the tool and so hence to do the commissioning process properly one needs trained professionals who are able to who have gone through training on these AI models and can work themselves through that commissioning process. Wonderful thank you for that answer. Melissa would you like to shake it? Yes I would like to ask a question I'm going to modify it some more and this is actually for David so we heard a lot about the benefits and importance of using machine learning deep learning, artificial intelligence and all of the technical talks underlined the benefits for extracting knowledge, bridging gaps, creating models of high complexity the question specifically posed is are there any possible limitations of using machine learning except bad data quality and I'm going to pick up that except bad data quality because in two cases as mentioned creating I call them super databases consolidated databases what happens if you have non curated curated data in your databases for the activities you're looking at this is an extremely important state of question and then maybe you could do comment on the possible limitations. It's a big scope but indeed it's a huge issue for data quality and data quality tagging is a field that needs constant attention and a lot more advancement. There are of course standards in various areas of science and industry and commercial application for what quality tags look like on data streams and sometimes we can exploit those and turn them into a machine learning setting so we take those standards of tags and turn them into what's called metadata the data about the data and quality is one of the metrics and there's kinds of quality of course there's statistical relevance, noise contamination whether we know there's an issue in a sample or whether we can infer it and in fact there are machine learning algorithms which can analyze data retrospectively and put metadata tags and information based on statistical inference so there are a variety of tools available and I think the main thing you're highlighting is the need to make large data systems that have the curated data already so that a large number of people can come and we can exploit larger communities to make use of that. I think we need to pay attention and having these large networks of distributed access points will enable that to be done on a large scale in an efficient way and then your second question What do you see as the biggest limitations of using machine learning? Yeah and so I kind of obliquely referenced a few of these issues for example machine learning methods are extremely good at creating models that map to what is embedded in your data whatever trends or classifiers or dynamic time series predictors but they're still very limited in how well we can assess the validity the quality the uncertainties embedded in that and to guarantee performance of something that we base on our data and this is a super huge challenge especially in my field of control in control the entire field you might say of control theory for the last 150 years has been dedicated to quantifying the performance you get from a controller when we fly on commercial aircraft we want confidence that our control systems are going to work taking off landing and flying and this is a requirement worldwide in so many real-time control applications that have them used in the public domain especially as they become mission critical maybe a thermostat is not so much regulated but transportation solutions are hugely and machine learning methods still have a long way to go to provide similar confidence in the performance of the solutions we get from them so there's a big limitation and there are corollaries to that in virtually every kind of machine learning approach knowledge discovery you already referenced and Jan discussed the limitations in that and Emma referred to how well we can make new science from machine learning I think this is a great opportunity and a very exciting potential but we have a ways to go Thank you for that I'm hearing maybe we ought to think about having a research project on control sounds very timely from that next question wonderful so this one is for Jan and also for Emma so what ethical issues open the use of chat bots in the field of medicine applications that replace a doctor quote-unquote or suggest methods of treatment and diagnosis and how do you classify do they classify patient needs and the direction towards specific specialists or specific medical departments I mean the gist of it is basically who controls the situation is it just a tool or is there more control on the practice of medicine and what are the things we need to watch out for so we don't we don't get blindsided so who would like to take this Jan okay To be honest I think this might perhaps be more an ethical question although in my talk I've always I guess the goal of AI in medicine is to provide tools to help the medical professional carry out its responsibilities more faster and more efficiently now not the goal of AI is not to take over decision taking or over medical decisions and there is a real there is a real challenge here because many of the systems that are especially deep learning systems for example the way they come to a decision is not always clear from from during the design or let's say a deep learning system needs to recognize a certain area that on an image or so that is suspicious or that is linked to a certain outcome we don't always know how it came to that conclusion so that the black box nature of some of the tools is a real problem so the end of the in the end the medical profession should just use it as a helping tool to come to a decision with regards to chatbots it shouldn't be no more than channeling certain requests not as to provide a replacement of a medical professional but perhaps Emma can comment on this further Emma please thank you I think what is important firstly is that patients have the right to know if AI is used anyway in their treatment in their diagnosis and in decision making if we come in the end to life and death decisions patients have the right to know whether they can understand their exact processes due to the problems with explainability interpretability also transparency issues with AI is another problem but that makes it even more important to focus on transparency and accountability to make sure that the patient knows from the beginning what is what the tools are that will be used in the treatment and diagnosis and that they know what they right to dignity and privacy would be whether in what way the data on them would be shared and we all know that there are huge issues in terms of data in their health sixer and that makes actually innovation and progress in health science is really difficult where it can really contribute tremendously to how we think about how but then also it is important and if I think of the UNESCO recommendation that we have just finished we have made it very clear that patients should be a part also of the entire cycle so they should be this kind of communication that I spoke about but also that in the final instance any decision should be made by humans and we feel very strongly about that human sentence how do we make sure that what the impact afterwards down the line it reminds me sometimes of something very simple it sounds like it's very simple we say it's like the guidelines medical guidelines and we understand many companies will just take them and say this is all recovering based on this guideline or that but we know no human being is the same and they're inevitably for anyone who's practice medicine realize is that you cannot fit a human being into these guidelines all the time there are always special aspects of it so for using let's say AI to develop to decide what happens to a person who makes certain decisions how are we going to know the impact later on without going through the process getting data and finding out that you could have done something better the outcome maybe you can answer that better than I can but I think that we should just that speaks to transparency but Jan maybe as you stand in the field maybe you can answer yeah so May is talking about personalization so again AI could be a great tool in personalized medicine because we have access to a much broader set of data than can be interpreted by a single human being guidelines are distillation or could be are based on clinical experience over given number of years of a certain field and are then cast into a certain rigid pattern the data gives us the opportunity to end the machine learning gives us the opportunity to open the door to more personalized a more personalized view on on medicine but of course I think all the ethics comments that Emma gave about this transparency and information needs to be needs to be respected here as well if the patient will be treated in a more personalized fashion general accepted guidelines and the patient should know about it there should be a clinical trial that guides this process and the like it is a process though and until we reach a point where this is actually acceptable for example the early days of AI where you could have pattern recognition and sometimes the patterns would be non-recognizable because you don't have enough data but you might recognize it as a human being example is the face of a cat and you ask for the cat face of another round kind of animal the early days and in medicine it may be the same situation where early on we don't detect these differences that could lead to much larger problems or a big mistake versus a smaller mistake if you're a human being looking at I don't know if my point is getting crossed but thank you for that okay yeah I think that this might be the last question because we are I've been made aware of the time and we have a question that is for Clemont and this is quite a specific question you showed a lot of applications and one example was you could identify seasonal variations across the United States with water so they got to thinking if artificial intelligence machine learning can enhance our capacity to model environmental transport especially the challenge of reactive transport on long-time scales and so this specifically comes from colleagues dealing with safety of nuclear waste disposal and they are faced with these sort of challenges so are you interested in your expert opinion of this being an avenue that is fruitful for exploration? yeah I think machine learning can be applied to any sort of science so as far as what I see for example for reactive transport modeling that would be useful would be that when you do this sort of application again you have data that are at very different scale for very local microscopy data you might have tomography data you might have seismic line if you want to do your kind of very large scale like reservoir contamination transport modeling so all this could be integrated relatively easily to try to understand like trends or patterns but one thing that I haven't talked about is kind of the reverse way you have also through the last 40 years a lot of mechanistic models particularly in reactive transport modeling that you can apply generally at relatively small scale because it's very computationally intensive to apply this model at kind of large scale so what you can do is use those modeling like the physical modeling that you have good validation for as kind of the training set for applying machine learning then and kind of extrapolating your data from your process model at kind of larger scale so I think that works in both ways and that's where I see like you could use it for this sort of application yeah yeah this is I'm thank you for this answer I think so too and then we get into all the issues of your data quality going into it how your interface interpretability parameters are dependencies that you generate but yes thank you very much for your answer I do not want to give kind of a full view that machine learning is extremely interpretable and kind of give us all answer about processes it's not what I'm saying I'm saying that I think the algorithm are making a lot of progress in helping us interpreting processes but it's still very much of a black box sometimes you can put a lot of stuff in this machine learning algorithm that you know don't relate to your actual process and still get a very precise model at the end and that's not what we want we want to just have a good user a good informed user that understand the data that comes in and then also good mechanistic models and kind of work that in tandem to understand the process better yeah yeah so you mentioned in the chat about supporting our learning in modeling in a non black box model bringing up words usage of cause and effect and what I learned from Emma that Sir Francis Beacon might roll over in his grave in terms of how this might impact our scientific message on the longer term so maybe these are things we should think about more as researchers as we go forward so I think I'm going to wrap up the question and answer I'm supposed to turn on my video so I have someone I have someone in the background coaching me we're going to have some conclusions we sum up going forward so if I could have the summary slide see yes so I think May go ahead so well first I just want to thank the last comment that what you said would fit into the personalization as well so that really solved that issue so first we know that AI machine learning and deep learning are powerful tools supporting healthcare in different contexts but we also learned that we have to be aware of interpretability of AI algorithm behavior and black box nature we had many discussions about that and that was a significant concern medical decision making is still the responsibility of medical professionals on algorithms so everything kind of makes sense and fits in with that final summary go ahead yes combining big data and isotope science with artificial intelligence machine learning deep learning framework to extract new information from what often we're looking for small variations and has a great potential in a multitude of fields including hydrology ecology friends and food security as well AI tools they provide methods for extracting knowledge from large scale data sets be it images high resolution images or database external data spaces also knowledge from combined data sets of different scale and applied diffusion science specifically can be used for diffusion prediction control and more generally for useful model generation and so living with AI technologies implies recognizing the impact of these AI technologies on human lives and the need for participation in AI ethics and rethinking education and training models on all levels and the application of AI of course to address many novel opportunities and challenges requires new efforts but they have to be multidisciplinary and they include training and education as we said well-designed workshops technical meetings knowledge sharing platforms for coordination amongst cross-domain researchers so it has to be an effort by a large group to be able to see the different aspects that can be quite limiting if we only come to data one aspect so with that I think we came to our final discussion I think we could go ahead to closing remark well first of all I want to thank all of our distinguished speakers today for their insight and sharing of their expertise they gave us much information and much food for thought and the one thing that blew me most away is I have to rethink what it means to be human and this is I'm going to be thinking about this thank you very much all of you and of course I also want to thank my IAEA colleagues for organizing, coordinating and participating especially Mateo Babarito the man behind the scenes and last but not least I want to extend a very warm felt thank you to all of the participants both for your interest and for your pertinent questions we did not get to them all but we tried to consolidate some questions into one and I hope it's helped and I sincerely hope this event has either instilled a broader interest in AI based approaches nuclear science and I also hope we can continue discussion to find pathways to work together in the future to effectively provide machine learning, deep learning and efficient delivery of the UN Sustainable Development Goals now over to you Mate to close so again, Excellencies delegates, colleagues, everybody online in closing the event I just would like to express my appreciation to all the speakers and participants for attending the IAEA side event and we've heard exciting and interesting talks today that really gave us food for thought as I said but I also also would like to acknowledge our colleagues behind the scenes also Yaroslav Pinda and Debbie Vandenewa were working hard to make this event smooth and successful along with Mateo so thank you, thank you all for your attention have a nice day and hopefully we work together towards a better future thank you