 I'm going to tell you a little bit, so we have 35 minutes, I believe, about AI and financial services and explain to you a little bit about what we have been doing in this area. And so the first thing I'm going to do is to really tell you a little bit about the actual area of AI. And AI in some sense, I wanted to share is a young science because AI, though it started in the, you know, somehow in the early 30s without untuning or trying to have computers basically reproduce human behavior. It was this conference in 1956 at Dartmouth College that marked the beginning of the field of AI basically with this kind of like proposal to have this study being carried over for two months with 10 people. And the interesting thing is this statement that the proposal mentioned that every aspect of learning or any feature of intelligence could be in principle so precisely described that the machine could be made to simulate it. So the interesting thing here was about this any other feature of intelligence and that's what led our field to be somehow this field of AI, a field of components. So I, we need to understand about these concept of the data that is being processed, the input to the reasoner like natural language, vision, speech, and then the decision making, which is like this search algorithms, the planning algorithms, the optimization, and the multi-agent, the learning, and then eventually the actual bot, the execution part. So I want this, I mean, I just want you to understand that when we have AI projects that we pursue, there is this somehow joint understanding of the field as an integrated kind of like goal, but also these individuals, individuals, individual components that people pursue. So that's why machine learning is one very strong component because somehow, as we know, intelligence is well associated or well merged with the concept of learning. And therefore, it's a very important component. So this is like the beginning and because of these, at Carnegie Mellon, where I've spent, I mean, I don't know what the introduction of me was, but I was actually for many years a professor at Carnegie Mellon University in computer science and then AI and in machine learning in which we try to combine a perception through machine vision, reasoning through computers on a robot and in fact, execution by enabling a robot to be mobile. So this robot here, as you see, is moving by itself. So it's a complete AI system and it's able to actually integrate these multiple components. You can also speak to it, it can also talk with you, but of course it's very much like a navigation robot. These robots and these AI systems, one more thing that I wanted us to understand is that they are not completely capable of everything. As we know, processes in many things we want to address with AI are very complex and AI systems are not fully capable, of course, of doing what humans can do. So there is this concept of inevitable AI limitations, which led us to introduce this concept of symbiotic autonomy in which the AI systems ask for help and learn. So then we built this capability with these cobalt robots that would ask for help pressing elevator buttons because they had these intrinsic limitations that they didn't have arms. And also cognitive limitations, but from which they would ask issues about like where's coffee or how can I, I don't know this thing, explain it to me, it goes to the web, but interestingly all the cognitive limitations, the robots or these AI systems can learn from that interaction. So what I'm saying is like this, I wanted just to get across these two concepts or at least these three concepts, one, the concept of like AI being a science of components, then the fact that putting the components together really will produce or produces interesting complex AI systems, but still with limitations and they have to ask for help. And this is the only thing I'm going to say about my many years of robotics, our robot experience, but I will take questions at the end. And I'm going to just delve into explaining now in the financial world what does this all mean or examples of what we are working on. So one interesting aspect is that J.B. Morgan Chase, which I joined in 2018, is a very large kind of like company. It has many employees, it has many customers, these 60 million households, it has many companies and a lot of digital users and the loss of spread throughout the world in terms of like being very global. So what this means is that this is a fantastic opportunity to understand both data and operations and execution at a very large scale. So I would just want to say that it's very, my first slide is to say that my understanding of systems that ask for help is kind of a cultural change in the way you think about how humans interact with these AI systems because you can imagine that the AI system gets all these data and is able to create a model. And then when you have a model, you are able to make a prediction for humans about what the model predict. And somehow based on the observation of that prediction, humans can give feedback on whether it's actually a good prediction or not. And the model gets updated with that feedback. So the reason why I always pick on this as a cultural change is because somehow you can contrast this type of learning from experience from basically thinking about learning just from data in which you get the data, you feed it into some powerful machine learning tool like a neural net or some other tool. And basically you then classify the data. This is not really just a classification problem or a classification one, this is about being able to learn from experience, be online learning, being always improving. I find this very important because you may want to understand that in some sense we are like literally sometimes thinking that AI systems are so powerful and AI systems are so how do you say so complete that we really view them as more powerful or at least that we can design them more powerful than I think they are. So the important thing here up to now in this talk is just to explain that the AI systems are learning machines over time through feedback from humans. Now I'll just review a few projects that we are working on just to give you a feeling for the impact that it can be to think in the third formative way, to think about the way we in the financial world do things. So here for example is like a trader's floor and I'm sure that all of you have at least seen these trading environments in movies, the same thing with me, I have only seen it in movies, but people are surrounded by lots of screens with visual imagery of time series data, of histograms, of a lot of analysis of what's going on, visualizations, but interestingly what we came up, we understand is that these humans, these traders end up making decisions literally from observing images. So it's not that they are basically computing necessarily processing the time series data or processing other information, maybe they are talking, but the image plays a big role in the decision and they can interpret the images. So interestingly then we knew from our experience in AI how powerful our neural nets have been to really classify from images any class of objects and it had been like objects that were have been associated with these classification power that neural nets provide, but instead what we did was we looked at these images of time series data, we chopped them into little kind of windows which we represent, it's basically a candlestick representation, it doesn't matter, and then we trained the neural net with such kind of little images and the associated by no buy decision that humans have done, and interestingly these model trained on images from the actual time series data performs at very similar at 95% accuracy when we trained on historical S&P 500 data. So this very brief first project I mentioned within these AI for financial systems is about indeed showing that images of in particular of time series data, but it could be other types of images like histograms, pie charts, all sorts of like visual representations of information can be associated those types of images with decisions that humans make. So we were used and we have been mostly used to use these neural nets to classify images in terms of the objects, this is a cat, this is a dog, this is a chair, this is a table, this is a motorcycle, this is like a car, this is a bike, this is an orange and apple, so we have been like so familiar with the object classification problem and instead we are now associating decisions not classes, orange and apples, but decisions by no buy to the actual images, and I think that you can ask me like what about if you use the actual values of the time series, you also do well in fact, but surprisingly you do equally well with the actual image with actually the image the pixels with the black and white marking the actual function and other images like I told you, but this is the first thing I wanted to tell is this Mondrian system that is able to really connect classification with decisions through images. The second thing I'm going to tell you is like even when we do this there is this problem of trying to have even a trader or one day an AI system to explain why the decision was made and to try to understand why the decision was made, again based on the vision on vision we know that humans end up looking at parts of the screen and therefore we made an experimental setup in which we were able to track the eye gaze of people looking at the screen to make decisions and that the gaze patterns in some sense help us track the sequence of attention points which in principle can explain the behavior or at least can be why did you make a decision for buy after you look at these three points, but you didn't look at anything else you will go into all these and you made a decision. So here is an example also from a complicated screen in which here on these orange dots we are counting along your attention was here followed by here followed by here followed by here so this is important to say that this probably you know was more relevant than these and eventually these sequence mattered and other parts were not used to make this decision or people at least did not look at them they could have seen with peripheral view but at least we are capturing what people really target with their view and here is a little kind of like video in which you can see three regions of this screen being tracked and the eyes moving from one to the other and you can actually eventually understand that this blue region look at how always that you look first at the blue then and the yellow then and the red so that is this kind of like dependency here and eventually you could then interpret these as an explanation of how decisions are made so so this is the second thing so I mentioned to you first the actual concept of Mondrian to learn from images and I just mentioned to you the concept of using eye gaze so I'll just sample a few other projects and then I will ask questions I'll answer questions so another project that is very important for us and that you want to understand from an AI point of view again the value of data is tremendous as you know most of AI and machine learning rely on data and JB Morgan has a lot of real data and the historical data is easy eventually to access but sometimes there is a lot of privacy privacy issues and we have been like wondering how to build upon the real data so I see so I'm lacking a slide sorry so there are two ways in which we have looked at so we have created the system an approach to generate synthetic data from these real data and one of the approach is to have somehow a simulation environment for example to generate market data in which we are able to create different types of agents different types of agents in the market and by really changing the values of those agents that play in some sense in the market play the game according to the rules they are parameterized but they enable us to create a lot of synthetic data that is real that is realistic but not real so it's almost as if you have a way of generating like chess player chess games or often or go games or any type of like other game in which you have the rules of the game but then you generate games that never exist in reality so we can create an infinite an infinite number of similar market days we can use by we can generate by using using similar configurations of agents and we can create different regimes and we can actually generate market days that are used as synthetic data so this is very important also and I really invite you to connect with us or go to our website eventually because we are making a very very large effort into making all these same synthetic data available for research and for understanding what is actually the domain all about so the last thing I'm going to tell you about so is about the fact that we create a lot of documents powerpoint documents all sorts of like ways to translate some representation into another representation which in particular is like slides or reports so there is this we looked at these as a translation of representation so what we did is that the we were able to build an approach these what we call AI BPDX in which humans can actually talk with this AI BPDX system asking for specific slides specific reports being generated and in fact this is the same very similar to the COVID examples in which we can ask the robot to bring coffee or to do this so you communicate with the AI system through language and then eventually those commands those languages human language commands are converted into actually executables that generate the output in this case slides so here is an example that says at the slide with the title satisfaction score so you say this in language and certain instagram from the excel file data csv we say this in language use column C of the excel file for this histogram so all of these are language statements and the outcome on the right is automatically generated by AI BPDX look the title is satisfaction scores SS of clients like was required in the title space then a histogram is given as this pie chart or and then eventually and then we use the column C for this so here is an example also of more complicated decks of slides being created instead of a single deck of slides you can have I'm sorry I see that instead of a single slide you can have a deck of slide and here is an example in which the users say create a company briefing that's deck so that's the name of the deck many slides and now the language that the human asks for is mapped into the parameters of that deck and you say okay run the analysis and eventually these particular kind of slides are generated completely automatically by the system based on the the requests made by language and in particular the the and the mapping into the actual template of decks of slides so it executes that specific order automatically and then the interesting thing here is that you can ask for very so changes you know add another company to the deck remove another one change the the the actual how can I say the actual time to just one year and in one second again you say I'm done and then these these changes to the actual executes all that the change is required again when the requirements were given through language and they are converted now into executables so this is very important and we are able to do these also for other reports where here like the left hand side is generated automatically by looking at data and converting it to the actual language and then I just want to say so I I mentioned to you the the Mondrian project and I mentioned the ideas project then I mentioned all these issues about being able to convert these these language to PowerPoint slides and basically trying to tell you a little bit about what we work on in terms of bringing AI to the financial domain and we have many other projects and just to finalize I want you to understand I mean I want to share with you that somehow all of our work is organized along these kind of like foundational topics which are very very related to the to the to the financial domain so how to eradicate financial crime how to deliberate the data safely how to perfect and affect economic systems and then the stakeholders how to perfect client experience how to empower employees and how to automate the policy compliance kind of monitoring and understanding and overall we want to work on establishing ethical and socially good AI so I know I want I on purpose focus the presentation on these a few a few projects in fact I forgot to mention when I summarize the project about like the artificial synthetic data but so these are like building blocks that are examples of the many projects we have and yet these jp morgan.com slash AI you you can access many of these projects and you can definitely send me an email if you are interested and as such I will be available now to answer questions for whatever time is left over