EP5: MBZUAI, CMU : Causal AI, Answering The “Why“ and “What if“ Questions | AWS for AI Podcast

Welcome to the AWS for AI podcast. Welcome to AWS for AI, the podcast where we explore cuttingedge AI solutions and the innovators behind them. I'm your host, Hamza Mimi, AWS solutions architect. And today we are diving deep into the fascinating world of coal AI and its far-reaching implications across various fields. I'm honored to welcome our distinguished guest, Professor Kung Tang. Professor Zang is the acting department chair and director of the center for integrative artificial intelligence at Muhammad bin Zed University for AI where he also serves as a visiting professor for machine learning. Professor Zang is also a professor in the Carnegie Milan University philosophy department and an affiliate faculty member in the machine learning department. He is widely recognized as one of the most cited researchers worldwide in the field of causal AI with over 200 published papers. Professor Zang's research interests span machine learning and artificial intelligence with a particular focus on causal discovery and inference. causal representation learning and machine learning. His work aims to make causal learning and reasoning transparent in science, AI systems, and human society. Professor Zang expertise extends beyond theory as he applies his knowledge to diverse fields including biology, neuroscience, computer vision, computational finance, and climate analysis. In this episode, we'll explore Professor Zang's groundbreaking work in coal AI, its potential application, and how it's shaping the future of artificial intelligence and scientific research. Professor, it's great to have you with us today. Welcome to the show. Thank you for having me, Hamza. Very nice to meet you here. Thank you. So before we get into the deep dive uh research uh and we have a lot of things to discuss uh can you tell us a bit about your personal story? You're born in China. You're uh a professor in Carneg University. You came in uh from the US now to the UEE. Uh what brought you to what bring brought this interest in AI and what's your personal story getting in where you are today? Yes. Wonderful. Yes. actually uh you know during my PhD study I really got interested in causality and machine learning. Why? Because I believed we have to understand why why things are happening how to make a desired difference and how and answer questions like uh what if something had been different right and at the same time you know nowaday we have more and more data. So as a consequence we can learn from data which is essentially what machine learning aim to do. So I have been uh interested in causality and machine learning and that's why and then uh after my study I was a researcher uh I was in Germany for a couple years and then uh basically KJ Milan is a very well known um for its contribution in the field of uh artificial intelligence and specifically um I would say we have kind of pioneers um in the field of causal discovery at Kimla essentially they made this field or this particular line of research possible. So that's why later I joined them to be a faculty member at the Kimla and um in the past years we identified a unique opportunity right here in the UAE or Abu Dhabi. So and uh uh basically with this university MBZI we try to make try to develop the right type of AI and try to transform the world in the right way with AI. That's why you can see uh you can see the trajectory. So you're you're really making the research universal in multiple countries, multiple teams. Um so you're part of the faculty members for Carnegie Mill University and uh also now in uh Muhammad bin Zed University for AI. uh can you tell us a bit about what are the key research area you're working on in in both universities and where do you think the vision for uh about are actually going uh in the future? Sure. Um great question. So I personally um care very much about why and what if questions. So uh from my perspective even if you just want to make a prediction whenever you deploy the system then you can see the users and those who design the uh software right will bas basically see the difference right it has an impact on the users and on society so that's why causality is a really fundamental problem and of course in many cases say uh we see the symptoms of Alzheimer disease but we don't know the true cause that can be easily changed otherwise the situation will be totally different. Actually you can see um in history whenever we can discover the new hidden clauses of kind of important phenomena uh or hidden laws like f= ma or e e= mc² then the situation will become to totally different humans will become more powerful and resourceful and human society can become much better so that's why I care very much about causality of course at the same time causality longstanding problem in philosophy and a lot of fields right and at the same time you can see this is the ripe time to really uh go to next stage in causality research why because now we have basic tools and we have so much data so that we can really um try to learn things from data and that's why I could see the potential and opportunities um and I believe uh I personally would like to make the causal process including hidden things like a true cause of the disease like virus, right? Um completely transparent to humans so that we can really just change what we should change in order to uh achieve a better future across many disciplines. Yeah, coausality is a very deep human uh research field since Aristotle uh where we we had multiple u sitations from different philosophical uh uh figures across the history about cause and effects which is still very very relevant today. Um before going maybe into the causal uh effect can you talk about uh maybe Muhammad bin Zed University for AI? This is a recent university that was created in Abu Dhabi. Um, and it's very particular in the sense that's only focused for AI. Uh, can can you talk a bit about what's Muhammad Zane University? Yes, great. It's a university kind of that completely unique, right? So, it's the first university uh dedicated to AI all over the world and uh and you can see why it's a ripe time to have such an institution. uh you know a lot of universities have AI programs right at the same time you can see right now in order to really educate um qualified AI researchers the paradigms or even the education model might be different from before. So at this university you can see we aim to do uh the powerful and trustworthy AI and we aim to really transform uh each discipline of science and each industry by making use of AI. So this is the right time to consider this problem and make a contribution and at the same time you can see the impact and I believe almost all disciplines will become totally different in the very near future as you can see from the uh the Nobel Prize winners right who um in physics chemistry and so on you can see basically they try to discover the information behind the data and then they can really make a lot of new things happen so And as a university yeah we aim to do transform AI so aim to address important problem across many disciplines. That's why we have uh departments machine learning department, computer vision departments, natural language processing departments but in addition we also have we are going to have decision science, biology and public health and so on because you know when when you try to really uh deal with real problems it's so important to really understand what the problems are and how to really find the best way to deal with the problem. That's why so we are doing AI but we are doing AI to really address real problems in human society. Absolutely. And I I can argue that all the physical world research and science is about finding causes to understand how the physical world basically behaves. Um before going into the use cases and I know you're working on different use cases. Uh I want to go into the foundational uh concepts. uh and let's start with causal AI and causality. Can you explain in simple word what's causality and why it's important and maybe causality in AI? Uh what is it? Wonderful. Yes. You know uh in machine learning right in recent development machine learning usually we care about prediction and very often time we use a blackbox models right to make prediction into the future. Um that's already kind of powerful. However, clearly that's not what we want in many scenarios. For instance, for disease treatment, we care about not only the diagnosis given by AI, but also uh what kind of mutation cause the disease and how to treat the disease, right? So, in order to answer this question, we have to understand the cause effect relationship, right? For instance, if you consider rain and a wet ground, now you can see they are not symmetric. If you have a have a way to make it rain then you can see with very high chance then uh you see wet ground but you can find a way to make it to make run weight without really affecting uh rain at all. So now you can see the asymmetry relationship. If you want to change or achieve some consequence and if you find some causes that can be easily changed just go there right and this is how we can treat diseases and this is also why nowadays you can see in biology right even medical studies we always care about caused relations and how to find the right cause that can be changed. So this is what I mean by uh causality. You have to find the right thing entity uh so that you can really change um the uh target thing you want to uh you want to change. M so causality is the um the the search for the causes of the effects we are looking at and there are multiple ways we can look at this and I know you're you are very expert in this domain but what what's the difference how can we uh recognize causality versus correlation versus association as as humans we have developed that intuition of okay this is a cause but we also still have that experimental way of getting uh the confirmation that this is actually the cause of something. Uh how can we mathematically and based on the data differentiate between associations correlations and causality? Wonderful question. This is a core problem in causality. So um first of all the classical view is that from only the observed data without without doing any experiments usually we cannot really identify causes right uh so because in the data you can only see uh say correlation or dependence but if you really want to find a causality generally speaking we can still do something uh one typical way to discover cause effect relations is to use interventions or randomized control experiments. Like for instance, we try to see whether this particular way of treatment is effective. What can we do? We have um we just put people or patients into two groups, right? And suppose the two groups of patients have kind of almost the same features meaning that they um they have the same distribution for all the features. And then for one group of the patients, we don't do anything. For the other we just apply the treatment and then we can compare and if the consequence is very different right we know oh this treatment actually makes the difference right so very often times we can do such experiments for instance if you try to see whether rain called the wet ground or the other then we can try to find a way to just change rain and see whether the chance of wet ground will be different okay right clearly generally speaking that's the case and then you can also find a way to change only the wet ground and then whether it rains would be irrelevant right and in this case you can see I know rain cause the wet ground but not the other way around so this is a classic way to discover cause if relations but this is almost too expensive and sometimes it's impossible or it's unethical for inance right if you try to see whether divorce has some effect on something how can you really force people to divorce right so that's why um in the field of coder discovery we try to just identify cause effect relation and even hidden causes just from the data we can observe. So this is this field is particularly uh known as a causal discovery. Yes. So there are multiple ways and it depends on the data we have obviously um and how we we will treat time series data for example is going to be different from like general data. Yeah. uh because causes are generally happening before the effects I guess. Yeah. Um but to go to the example of um wet ground um it can happen from all different sources. So and we and you mentioned some some experimentation where we cannot actually do the experimentation. Um, how can we mathematically just based on data without doing the regular process of hypothesis, experimentation, manipulation and getting the data? How can we just based on massive amounts of data get that uh intuition and discovery of uh uh of cause? Wonderful. So basically the principle is um uh known as the fundamental principle is known as a kind of modularity property of a causal system. What does that mean? U we have a system involving rain, wet ground and so on, right? You can see all those things can be dependent right with high correlation. However, if we have the system in which rain causes the wet ground, we know actually the system can be decomposed into irrelevant modules. One module or one process generates rain right there and the other generates weather ground from rain right here. They are not related. M the two things are not related although variables are related. Now we can see we have very clearly defined the modules involved in the whole process and this is very nice. This is how we can do divine and conquer and this is how we can build aircrafts. We divide the job into different modules. You take care of this part I take a different part and then align each part um basically certified. then we can put things together right and then we have the um a good system big system. So this is what I mean by modularity basically all methods that can help us discover kind of causality or causal information from observational data are different kinds of instantiations of this property. We just try to see um how to find the uh a model behind the data that can satisfy the modularity property. So in other words if rain cause wet ground then when you first model rain and then model wet ground given rain then the system is very simple otherwise I mean if you consider a problem in a different direction the system would be very complex so that's the basic idea I know yeah this might be too technical essentially it's it's very interesting so you're div you're basically using the divide and conquer um are you dividing also the features of the data you're Yes. So usually we talk about the cer relation among the features, right? Okay. Of course in many cases uh it makes no sense to really talk about the cer relation among features. Let's consider one problem, right? Suppose we try to see the personality and then uh you can design questionnaires. Then we have so many answer scores right to the questions. Then generally speaking although the answer scores right are clearly related or correlated they there are no direct color influence in between instead generally speaking they are caused by the underlying hidden mental conditions. So in this case you want to analyze the dependency patterns of the measured variables right aiming to discover the hidden variables and relations and this is also kind of part of the problem of causal discovery or causal representation learning. So that is a very interesting point where sometimes we don't have that cause in the data we have and it's a hidden variable maybe it's a composition of different features that we do have. Um are we able using AI to find those hidden variables and how can we do that? Wonderful. Yeah, that's a great question. The short answer is in most cases yes and we aim to produce very reliable method to do so. Why? Because you know scientific discovery is mainly about how to really understand the phenomena so that you can see what's going on behind the data. Right? As I just mentioned, if you discover the hidden causes of what you observed, then you can try to find a way to measure the new things and you can even try to manipulate them, then you become more powerful, right? So that's exactly um the progress you saw after say um that discover virus, right? If you know actually those diseases or those symptoms are caused by virus and you try to understand how to manipulate the virus then clearly you can treat the disease in different ways. Yeah. Yes. So um right nowadays I would say in across many different settings or different problems yes um we have specific methods which can directly recover such hidden causes and even the relations among hidden causes like can different uh we have different mental conditions right we have different dimensions of the mental conditions and the different dimensions may be related for instance right Um if you talk about personality probably the underlying dimension of personality known as say agurableness would be a cause of extraversion right so those relations among the latent variables or latent representation could also be discovered from from uh the measured answer scores okay and so depending on the data and the use case we have sometimes it's easier to get to that uh causality relationships Um, are we able to quickly categorize these areas where it's feasible and the areas where we don't simply don't have the data? Wonderful question. Wow, that's a really cutting edge uh uh question to ask. So, I would say yes. Generally speaking, if you have enough data that are diverse enough to cover the system, then you can uh relatively easily discover the hidden variables and relations. But in some fields, in some particular domains, maybe you can only measure um a very small number of different things. As a consequence, you don't have enough information to recover the whole picture. then that in that domain it would be very hard to discover causal relations. For instance, we talk about um psychometric studies, right? So we can measure so many variable we can try to think of different conditions and we design questionnaires and we can manage me measure so many things relevant to the mental condition. Then this field is not difficult to really discover the underlying reasons of the answer scores and uh try to give some interpretation of the reasons. And in some other fields for instance even climate science it might be pretty hard because usually we don't really measure too many things and the system is very large right there could be a lot of kind of relations and hidden things so it' be pretty hard we are doing research in this field but um compared to those say uh a number of health healthcare problems biology problems I would say climate is still um a pretty uh particular domain uh that is very challenging. So climate climate science and climate analysis is basically one of the largest domains where we have a lot of calculations, a lot of data and it's hard to modelize everything. Uh and there is also the butterfly effects and all sorts of effects that can go into into play. Um is it fair to say that it's depending on the size of our modalization of the problem that we are able to get the causality or not? I would say first of all complexity could be one factor but it may not be the crucial factor. The crucial factor is something related to the information you can measure versus how complex a problem is. So if the ratio of the information you can measure right to the capacitive model is high meaning that you have a lot of information then it's easier to solve the problem from the causal perspective. However in some other scenarios although you have you measure so many things the problem is so complex as a consequence you still don't have um enough data or enough information to have a recover the complete picture behind the data. Okay. So there is the computational power, the algorithms and how we are actually treating this problem and the complexity of the problem and the data we have. Exactly. And it's a combination of all of these three. Exactly. Okay. Um maybe um can you can you maybe speak about what are some of the ways we do the causal discovery? Um I know this is a fast research domain of yours uh but just as a high level how are we treating that data in order to define let's say for a basic example where we are able to define that causality uh what are the ways we are doing this wonderful yes uh that's a technical question but I'm very happy you ask the question so um it's not actually not mysterious so um a classical way to discover causality to make use of how variables or entities right or events are dependent on each other. Let me give one example. Suppose now we have say uh now we have two persons. Suppose we have the third one right here right. Um so Hamza uh whenever you do something the third person may follow you with some probability right and similarly if I do something the third person right here may follow me with some probability. Now suppose there are no influence in between between us. Now I can see clearly the cord relationship would be going from you and me to the third person right because both of us can affect the third person. Yes. Although we don't uh influence each other right now we can see from data we can easily discover this pattern from data can see since there are no relationship in between our behavior would be independent right would be independent there are no dependence between us but between you and the third person and between me and the third person there'll be dependence okay and then this there only one explanation in this case both of us would be the causes of the third person okay so we we basically look at the data define find the features and variables that are independent and the ones that are dependent and basically on that we are able to start inferring the causality between the different Exactly. So it's a basically a way to make use of dependence and a conditional independence or independent patterns to infer the informationational causality. Exactly. That's a classical way and modern ways make use of kind of more advanced machine learning pro uh settings and algorithms to discover um deeper information. So for instance even if you have only two things or two variables say uh suppose let's consider you and me suppose uh whenever I do something with some probability I follow it but there's some noise right so I try to follow you but there might be some noise then generally speaking you can see the following phenomena um so when you try to explain my behavior from your perspective right by making use of your behavior then the system is very simple why you can first explain your behavior And then when you try to blame my behavior on top of your behavior, you just introduce some kind of independent um factor you cannot control, right? Because I try to follow you but there's some noise. However, you can show that generally speaking, if you try to explain your behavior by making use of my perspective, meaning that you consider you as effect while me as a cause, then the system would be the mathematical model would be very complex. Generally speaking, you cannot just first explain my behavior and then explain your behavior by making use of my behavior and the independent noise. You cannot find independent noise anymore. So it's a kind of asymmetric uh between the two variables Hamza and absolutely I can see how that can be a very computationally hard problem to solve as well because of uh finding the relationships but also the ratio from signal to noise and uh developing the highest most performant algorithms in order to find that is basically a a big domain of of of research. Um now that we've talked about the causality and how to get to that I want maybe to push on okay now we define that relationship um how can we make use of that relationship and there are the applications but from a theoretical perspective um maybe I can touch on the explanability in AI and also uh maybe a little bit about the counterfactual uh reasoning Okay. Uh so maybe let's start with uh the explanability. How can we use the causality definition in order to have more explainable AI? Wonderful. So there are different way to think of explanability, right? In AI. Uh sometime we just try to have a mathematical model. We just try to see how the input features affect the outcome. Right? It's just a mathematical model. But very often time we try to use a column model to explain it. We try to answer why questions. For instance, right uh in AI diagnosis, maybe you want to see, oh, why basically is the diagn diagnosis result going this way, right? Why for this person particular person, right? Uh uh the diagnosis result is say um is positive. You may want to see answer this question sometimes with similar symptoms as well. Exactly. In that case you may want to see uh the um basically the effect of the disease and then this is because when you do diagnosis you try to find a relevant feature but you want to give a clear explanation so that the doctors can really see what's going on right so you may want to go to a particular region say right this region basically has this function and for this person this region basically um has some issue right so this right here we try to explain or answer why by making use of the underlying causal relations. So right we um and very often time we try to achieve explanability for the purpose of making a difference right this is a basically a causal uh problem. So for intent we try to say oh why is this person so poor compared to other persons and when you try to answer this question automatically or very often times you try to see what this person could do in order to make a difference right so in this case clearly you may want to make a use of a column model to explain uh why okay let's let's go back then to the counterfactual uh reasoning um once we have that uh uh root cause uh that we have we have identified um we can infer uh other scenarios. Can you maybe explain what's counterfactual reasoning and uh how can we use that? Wonderful. So basically counterfactual reasoning is about imagination based on what you see. So from my perspective counterfactual reasoning might be the best way to achieve um personalization. So actually it's a way to really make sure you can see the hidden information and make use of it. For inance one kind of a couple right says oh you are unique to me. What does it mean? Basically they can see the unique features or hidden features that others cannot see and then they can understand each other very well across different scenarios even if the scenarios are just imaginary. Right? So basically kind of factual um question is about the following question um say today is not raining what if it had rain today right u what would uh have happened to me if it had rain today now you can say by doing this kind of question we can try to reflect on what happened what we did and so on and then given that reflection we know um what we should do in a similar scenarios right and at the same time confession reasoning is very different from the uh traditional kind of prediction model involving uh a lot of features because in counting we have to see what's what happened and then in for the hidden thing by the way here it's very interesting to see that actually in the world there are no noise right if you have a cold model say rain cause weather ground we know relative to rain there could be some noise in weather ground why uh because of a lot of other conditions however the noise is only the omitted factors you did the measure. They are just regular factors, right? And as a consequence with a kind of factor and and model you can really try to infer the noise so that I can have a better understanding of the whole system to provide a more informative answer. So if you like I can give another example. So for instance suppose we know um attendance is a cause of final grade right um so this is the end of the semester right students care about the final grade so suppose attendance is a cause of final grade now suppose for me particularly for me uh my attendance is pretty poor and my final grade is a is very poor you may want to see what would have happened to me if you had forced me to say attend more classes right in this case you can See you have to try to see my hidden information from what you observed. point and you can just try to see that compared to other students actually other students who had similar attendance my performance was relatively poor meaning that maybe I don't pay attention even if I attend classes right and this is the hidden feature and then when you try to see the consequence of intervention say forcing me to go to more classes then you have to incorporate that part right that feature of of this particular person then you can get more informative answer. That would be a very interesting model to have actually. So um and I I can also imagine that this counterfactual reasoning can be used for emergency situations for example. Yes. In order to infer the what if we handle this situation differently and then uh enhance our operational readiness for future uh situations. um and all sorts of uh applications. Um I I want to go back to the explanability part and maybe talk a bit about the reasoning um capabilities of large language models that we have today and I know this is a field where we have different ways of doing the explanability bit and coal kosal AI is is one of the major ones. Um how can we use causal AI in order to understand how large language models generate texts? Um that's a great question. So right now you can see uh a lot of system including uh chbt right make use of auto reggressive generation together with transformer architecture. So uh in some way it's very powerful right because uh you have a lot of data you can really memorize so many pattern dependence the patterns but on the other hand you can see clear problems we collected a lot of examples um so the this kind of platform right can really make some naive mistakes why so in some way you can see uh there's a there could be a bigger gap between our reasoning process and what the language can do, right? Large language model are trained to try to mimic the human uh reasoning process. However, you can see that primarily language is just a way or tool for communication instead of reasoning or thinking. there could be a big gap and as a consequence uh in some situation you can see the result is very good is very convincing but in some other scenarios there could be very naive mistakes I think you are aware of many right such examples so um from my perspective the issue is um there are multiple issues one of them is that large learning models right now may not really have the right types of representations For us when we look at different things we have we encode concepts relation among concepts right we make use of our understanding in terms of concepts and relations and logical uh implication and so on right to answer questions but now basically large model make use of dependence pattern association of information without directly deriving such kind of concepts and relations and so on. As a consequence sometimes some patterns would be completely lost. So we cannot really answer some particular type of questions. So uh for instance we uh we make use of particular architecture or structure in the brain right to in order to really uh encode the different kinds of information uh including causal information logical relationship right and so on. But for large models uh there are no such ability. So as a consequence right now with only text data at the input I believe the power of large language models is kind of limited. So there could be multiple ways to go further right. Essentially I believe what we expect from large lang models is uh something like to have a uh artificial brain right the large lang model can really connect different kinds of input or different modalities right including um video data that we can see pictures we can see audio information uh we have access to and the text and so on and then at the same time you may want to touch things and see the feedback and then you really have the proper way to encode what's going on right in the re in reality and how things are related and then like lang model can can be much more reliable but now I would say uh given the gap between what language contains and what we expect in term of the reason process yes they are the long way to go okay do you think um this will require a fundamental change in the transformer architecture and how we currently build large language models mod and foundational models or it's a tweak to the existing architecture. From my perspective, my own perspective, I believe is um we have to come up with a different way to uh represent the uh the uh the learned information uh inside the large langu. So for instance you may have a uh some kind of a uh transformer architecture but somewhere you have to encode the say temporal information uh concepts relation among concepts causal implications and so on. So find in our brain right we have neuron clusters and we also encode say different concepts by making use of multiple neurons and at the same time we also have the connectivity right specific connectivity um involving specific regions right those things uh should be included in the large language models at some point I I um I want also to spend some time speaking about some of the um practical use cases and you have uh focused on the uh real world applications of causal AI. Um one of the papers that you you you published is about causal reasoning and persuasive dialogue. Um and how we can use the causality in order to understand the causes for a person to react a certain way and in in in that particular paper to augment the donation amounts. um how how can can you explain the use of coal AI in order to influence uh the dialogue and the exchange? Wonderful. So now suppose there are two um different uh salesmen right they try to sell something one of them could be very convincing um the other uh may use the same way to deal with different people right and you can imagine how the outcome is so what's the difference essentially the first one very good one is very good at discovering the hidden information of the particular say a customer, right? By making use of the interactions. For instance, right, very quickly the salesman may discover that um oh, this person basically um say is very considerate and as a consequence this customer may want to get something for uh his friends, right? Or his kind of partner. So, and then such information you can see is hidden, right? It's not directly given. But once you discover such information, you can do something uh accordingly. Right? And for the uh second one, although this person can make use of a lot of information, but the hidden information, the hidden personality, the hidden desire would still be hidden. You cannot really make use of it. That's why if you have a system that can really discover the hidden features like a personality, desire and intention and so on and then clearly can make use of it, right? we can understand each other and we can try to do the right thing to satisfy each other. So every interaction is a basically a source for additional information that your model is able to capture and use uh causal AI in order to find causes of that enhance uh the the response um and to do something pro to do something uh proper for this particular person, right? To get a different way of behavior. uh this is also one of the application of counterfactual reasoning I guess right yeah to the right yes okay so what if I so basically in real time the model would uh consider different versions of how they should address a certain customer and use the one that basically provides the best value for a specific uh feature exactly yeah based on understanding you can predict which one is really the kind of most effective one brilliant yes um one of the other uh domains you I know you're looking uh in is the in education uh can you speak a bit about the work you're doing maybe with ministry of education here in UAE and how can we use coal AI uh in education in general uh wonderful so I personally believe um so in AI there are not many problem we really um we should address right so uh those uh domains include include healthcare education and probably clim I mean the science. So in education uh first of all healthcare is very important because we can really make people uh live kind of more happily right and at the same time we care more about the next generation that's why education might be even more important and you can also see the issues uh with education right now and first of all you can really see a big difference between a talented teacher talented and a devoted teacher and a kind of mediocre teacher right so the first type of teachers would be able to really understand the hidden information of each student and do something accordingly in order to make a desired difference so that each of the students can become different right in a better way right so um but for the second one uh you know it's very hard right you may just want to do the same thing to different students then some student may benefit from that but some student can really get hurt so that's why you can see with AI and the data we can really discover discover this kind of personalized information and make use of it so that all of them collectively will be better and that each of them will be a better version. So that's our purpose and there are different way to achieve that right we just make use of um all kind of data we can observe and try to see the underlying um reasons for different behavior and then given that reason um there will be some uh some suitable recommendation or intervention we can try to apply in order to for this person to become different. Okay. So in an ideal um student experience uh there will be some kind of uh assistant to the teacher uh where it it it will provide them with uh customization for each of the students maybe in their homework maybe in how they interact and uh provide that customized uh learning path depending on different behaviors and different data from each of the students. And even in terms of curriculum design, maybe you want to make it specific or you you want to make it personalized because students learn things in very different ways. We we we can even push this maybe to the students where where are the areas of interest of the students, what kind of career path they can follow. uh but this also brings other questions about should we use this uh for uh for ethical reasons. So there's always that balance between we can provide a better experience for the people we interact with but we can also use this technology uh in ways that the ultimate customer the ultimate user uh does not approve of. How how do you have any specific thinking about this particular problem? Wonderful. Yes, a great question. Uh before that let me add one one sentence to the previous one. So we um we are doing student centric and also teacher centric and school centric and parent centric uh kind of modules. Why? Because you can see clearly students should be uh should be the center right of education. At the same time we should try to find a way so the teachers can be more sustainable right and uh and that the school could be um could allocate the resources in a better way right and so on that's why we have different modules and now come back to uh the question you just asked I believe you're totally right so it's um it's a very hard question and from my perspective um in order to do so we have to make sure the system is first of transparent and the second of all certified. So certification is very important. You should make the you know we use a fridge with the cars right every day. You can see those those things are clearly certified right and you can see uh the system is although the system may not be completely transparent to you you can trust it right because experts already did a lot of work in order to make things reliable in many ways for AI system is the same thing right and first of all we can try to uh as you mentioned right explanability would very important we can make it transparent in terms of explanations and second of all the system should be um the information should be available at least to experts and after that the system should be certified so that we can we we can see why this system is developed and deployed who will benefit from it and why and how and whether some other people basically will have to sacrifice right for this reason and so on and then we can make the final decision or collective decision based on that. Um I I want also to touch on maybe other use cases. We talked about education. Um we talked also about um uh climate analysis uh at some point. Um finance is also a big topic of research of yours and finance uh volatility and how you modelize the market is a big domain of research for everybody since forever. Um how do you think coal AI can uh support the research and uh have practical applications in finance? Wonderful. Oh so you ask about climate science and finance right? Maybe I can just say a few words about um uh both of them. So um so in climate science you can see we aim to understand why say um uh that climate change phenomenon and we aim to see whether we can do something to prevent the very poor phenomena right from happening and we try to improve the situation and so on. That's why we have been trying to recover the underlying driving force of the change and see how the pattern even the uh climate patterns how the patterns actually changed in the past years and how to locate the the uh fundamental causes and how to understand them and how to find the best or possible ways to change them. So that's our final goal in uh climate science. So you can see why causality plays a very important role because you try to really understand the fundamental causes of the change and give interpretation and find a way to make a difference and go into maybe the depth of specific features in the climate uh per region per yes uh specific location maybe and there are there are kind of global factors and also local factors. You're totally right. Okay. So understanding the cause and use that in order to infer what will be the implication in a local specific place maybe uh in a particular feature heat uh humidity whatever uh moving forward yes and then you can based on the understanding of the causes you may want to find a way to do manipulation right for instance uh if you see actually this cause is heavily related to say uh uh greenhouse guess right they can try to yeah you can try to find some some way to reduce the omission of uh the uh so modalize depending on the trajectory and the uh greenhouse gas effects and uh maybe also depending on local adaptation yes uh scenarios where maybe we have mangrove uh being built does it have uh an effect on different uh areas yes exactly and a kind of very tall buildings um as right and so on. So there are many factors and if you can quantify the effect of all those possible factors and try to and finally we can just see uh what price we have to pay right to maintain good life at the same time we can really try to see what we can do right to make life uh not only happy but also sustainable. Absolutely brilliant. Yes. Uh okay let's let's go now now to the finance part. Um so what what are the application that you you are working on on the finance world? Um right now we are really working on kind of a trading system. So I would say um at this moment we haven't explicitly made the use of causal AI. Um but you can already see the power of uh uh I would say even traditional AI if you formulate the problem in the right way. Right. And the the key point is that follow the finance or investment is not about prediction. In traditional kind of uh investment you try to find a kind of fixed model right that can make good prediction all the time. Then you try to make use of this model. However, in finance, we don't really care about the performance of fixed model. At last, you can identify relatively rare rare. Let me say again, as long as you can identify the relatively rare opportunities where you can predict the future and the market change a lot, then you can just make money, right? That's the fundamental basically idea. So uh the issue is how you can just identify the those opportunities. Okay. Right. Because this is because in finance things are heavily non-stationary. A lot of things change over time. Right. And at the same time uh from the causal perspective we are also doing a project trying to really understand the driving force of the change. If you can identify the driving force and also make prediction um uh in this way and since most people don't really realize the existence of a particular hidden thing while you can do it then you can also find out new opportunities right to uh make use of the u the predicted. So given the complexity of the system, it's not about predicting the causes of everything uh because we have a high volatility. Exactly. Just the ones that we have we are able to identify those hidden features that we can predict in short term. Yes. Uh to to to basically get some additional value out of this. Wonderful. So if you compare the very experienced investor and a kind of a new one. Now you can see the very experienced one can detect and make use of many opportunities right but still I don't think this investor will basically uh will um do trading all the time sometimes right the person may not be confident right you don't have to do anything but for the new one probably yes the person cannot really identify uh many opportunities in a reliable way right so basically with the AI we can really try to be a very smart investor who can identify as many opportunities as possible reliably. Perfect. Um there is also and maybe I want also to touch on another field uh very briefly you speak you spoke about the mental disorder and uh the applications in healthcare. Um how we detect ADHD uh how we diagnosis uh different uh effects and causes for specific treatment. How do you see this moving forward uh as uh another way for us to detect, treat and then have a better diagnosis for different illnesses? Wonderful question. So uh you can see when we try to deal with mental conditions or kind of mental disorders, right? First of all, we have to understand the underlying difference across different uh subjects, different people, right? And we have to see the reasons and this is basically a causal problem because we cannot directly identify the causes. we can only design say um questionnaires and make use of rating scores pro provided by the kids or parents right or doctors and that's let's provide a way to automatically um look into the hidden world right you can try to see the reasons and different reasons and then once you go to this step you can try to see what would be the causes and then based on understanding you can try to design very safe uh therapeutics to make a difference for instance In one of the projects we did, we try to identify what type of game digital game would be very effective in order to change a particular dimension of the uh reasons behind uh ADHD. First of all, it's safe and second you can see for kids right playing game for 20 minutes every day would be fine. But if the game is designed in a specific way then the particular reason of HID can be different in a very natural way. So that's uh that's why understanding the underlying reasons and u trying to figure out what kind of safe intervention could be used to change that reason would be so so relevant and important. It's brilliant. So beyond the diagnosis part even the treatment part we can actually experiment without impacting the patients. Exactly. Yeah. And you can do countros right because now we can see once you have the causal model basically the world is transparent. Yeah. Right. Even if those things never happened you don't care you can use the contractual reasoning to see the consequence and to the comparison and finally you can find the optimal action. Brilliant. Um before we end I want maybe to ask you two questions. One for the research world and um your what do you think AI is going to look like in five years from now? Uh and with that vision what would you advise the current researchers working on different AI projects? That's a great question, very broad and um from my own perspective um I care very much about the consequ impact of AI. The impact of AI should be transparent but now I believe uh this kind of research is heavily lagged or even underexplored. So um I believe we should care more about the impact of AI truth um and we should care more about the future of human society in terms of say confidence right human confidence uh intelligence and uh independence right um so in some way right now those factors are ignored in some Okay. So I would say when we do AI um even when we try to formulate AI problems probably in the future we should try to see the consequence of this particular development in the first place. Right? Let's talk about recommended systems. Right? Recommended system is so good because they can make our lives much easier. Right? you can easily see what you can what you want. At the same time, you can see a clear consequence of traditional recommended systems. Essentially, first of all, it could be just a way to repeat history, right? And the second way uh second of all, you can see it might be a way to just get addicted to wh