 to Texas Heart Institute educational programs on innovative technologists and techniques. The topic of today's presentation is artificial intelligence, current status and future applications. My name is Juan Recrecia. I'm an international cardiologist at Texas Heart Institute and Baylor St. Luke's Medical Center in Houston, Texas. Join me today for this program is Dr. Ben Azzang He's a Abricovir Professor of Electrical and Computer Engineering and also director at Rice Neuro-Engineering Initiative at Rice University in Houston, Texas. And also Dr. Mehdi Rezavi and he's a director of Electrophysiology Clinical Research and Innovations at Texas Heart Institute. And he's a cardiologist and electrophysiologist at Texas Heart Institute and CHI Health, Baylor St. Luke's Medical Center in Houston, as well as the adjunct professor of bioengineering at Rice University. Welcome gentlemen to this program. Here are our disclosures. So Dr. Azzang, what is artificial intelligence? As you see from the slides, it's a subfield within computer science. And basically the idea is to learn how humans make decisions and how they exercise their intelligence and try to do that with machines, basically. And the kind of tasks out there are visual perception and speech recognition, decision making and many tasks of the sort that we would like to build machines that could carry out these tasks for us. So can you give us a broader perspective on artificial intelligence? Yes, indeed. From my perspective, not being a computer scientist, there is a broader perspective here. And that is try to understand how humans make decisions. And we, as we all know, we make decisions based on learning and we learn from our experiences over the years. Dr. Azzang, please tell us about the history of artificial intelligence. The history of artificial intelligence goes way back in the 40s. And we got to remember that in the 40s and even earlier, scientists were always interested in how humans make decisions. And then computers were coming along and computing devices were coming along. And then the idea of how computers could mimic human decision making was very intriguing. And you, but as you would imagine that that's not easy task because we didn't have that much data and to learn from. And also that as engineers, we always look at the problem and we try to model the problem and they make decisions based on the model. You can imagine that coming up with models of these kind of engineering problems is very, very challenging. And the models are often either intractable or insufficient. That's why you could say that over the last 70 or 80 years it took 70, 80 years to make progress at this level. The reason is that many different fields had to come together and make advances so that we'd be in a position to have a lot of data, to learn from the data and make decisions based on what we learn from the data. So it looks like those early scientists, Warren McCullough and also Walter Pitts, tried to do exactly what you mentioned there to mimic the human brain and design the computer to think using something similar to neural networks. I guess this was the essential step to move forward with this technology. Is that correct? That's correct. But it was basically what I'm trying to say that it was impossible to get it done back then because many different fields had to come together. Many theories and many models have to be built for it to be accomplished. And back in the 40s, it was just a great idea, a guide idea that inspired many, many people to work in this area. And many things had to come together. Many fields had to make advances to be able to accomplish what we have. Basically, as you all know, Turing came up with the idea of computation. Shannon came up with the idea of information theory, how to quantify information. And Wiener came up with many different ideas of this sort. All of that had to happen for this to realize, but obviously it hasn't fully been realized yet, but we've made great progress over the last 10, 20 years. As I mentioned before, many different disciplines had to come together and make advances to bring us to a point that we are today, that we can build systems that can mimic human intelligence. And the steps that we needed to go through were developing integrated circuits, building sensors, building devices, understanding how signals carry information, process the data, and go from data to control. And many, many disciplines had to advance to bring us to a point that we can say today that we have made progress. We are beginning to build artificial intelligence systems. So you already mentioned to us that it took close to 77 years from the original description in the 40s of artificial intelligence until the present times. And you mentioned that all those technologies had to merge and we had to gain advances. And we had to also have a power to record this information. And at that time, the computers were really not designed to that degree to be able to store all this data on a larger scale basis. Not only store, but actually process, right? They had to collect the information, collect the data. They had to store the data. We had to process the data and get key features from the data to be able to mathematically formulate the problem, solve, and then further implement and carry out experiments. All of these fields, all of these disciplines had to come together and make great strides to bring us to a point that we are today. Right, so you did mention that several things had to happen. And of course, one of the things is to figure out the algorithms that are going to work well in computer science for artificial intelligence to work well, but also to implement logic and mimic the neural network function. All of those things are very, very important. And now we have so many algorithms that we're using that did not exist way back when this initiated originally. Absolutely, originally it was a brilliant idea, brilliant concept that was not time to implement. I think we needed to make all of these advances over the last 77 years or so to come to a point that we have now some ideas, we have some ideas that have already been implemented and we can identify a cat from a dog and all of those examples that we talk about. Dr. Azan, can you tell us a little bit more about machine learning and what are the essential components for machine learning to be effective as far as artificial intelligence is concerned? Absolutely, I consider machine learning as a subfield within the broad term of artificial intelligence. That's my personal angle to this. Basically how a computer, a machine can learn from data. Basically, that's basically what it means. How can it effectively learn from data? And the data is now available, the data doesn't come with the model and the model has to come from the data and has to be extracted from the data. Turns out that we have tremendous amount of data today and we can learn from the data through these machines and then make decisions, make predictions, make the classifications and many other tasks of this sort. And they have come a long, long way. So give us some examples how machine learning works. The examples are right on the screen, right? Basically, we've been always very sort of intrigued by images and the question is that if you look at an image, which is basically now there's a digital image and is all data and can you tell whether this is a picture of a cat or a dog? And there is no real simple model of a cat. There is no simple model of a dog. Therefore, the system has to learn features from the data that help it identify that this is a cat and learn features from the data that helps it identify is a dog and then run these algorithms and make a decision whether this is a cat or a dog. And this has been an interesting research problem for 70, 50, 70 years. And now we have made tremendous stride in that area and there are algorithms and there are ways to do this thing today that you could look at many, many pictures of a cat and many, many pictures of a dog and decide what are the important features and then decide whether this was a cat or a dog or a bird as an example. It can be done today. So this is just one of the algorithm. Of course, there are many, many other algorithms that we could use. And here is another example of artificial intelligence. Can you tell us a little bit about the neural network? This is closer to your area of expertise. Basically, a neural network is a circuit, basically, is a network that was inspired, that was designed and built and inspired by human brain. And it was just inspiration. Nobody is suggesting that this is exactly how our brain works. This is actually not how our brain works, but it is our brain inspired us to build deep neural networks of this sort and use data to train the various weights of these circuit and then train the network so that it could identify some features of a cat, features of a dog, features of a bird. And then at the end, as you see, the end of it has only one layer. And at the end, if it is a zero, then it's a cat, if it's a one, it's a dog. And basically, this is an example of a machine learning methodology that can help us learn and differentiate between a dog and a cat. From my perspective, a deep neural network like that is one of the many examples of machine learning methodologies out there. And it's very powerful, it's very effective, and in many, many, many cases, it works beautifully. Just that it needs training, it needs a lot of data, and it can identify features of a cat and features of a dog and make decisions that this was a cat or dog. Basically, if you think of it, AI is the big picture in building artificial intelligence. Machine learning is a subfield of that that basically tries to learn from the data. And neural network is an example of a machine learning technology methodology that actually is a circuit, actually can be built, can be programmed, and in many, many, many cases, it works very, very well. So this is one of the simpler description of a neural network, how they work. Here we have only two hidden layers, layer one and layer two. Actually human brain and neurons are connected close to maybe 15 or 20 hidden layers, so-called hidden layers. If not more, yeah, if not more. If not more, which we didn't get to that point yet with computers, but that is obviously our goal, right? Absolutely. We would like to be able to give enough degrees of freedom and have enough layers and enough nodes in the layers so that the network could focus on features and at the end come up with a decision based on many, many features that it identifies. This was the inspiration. And it works, as I said before, it works quite well in many, many, many cases. So yeah, this is true. The aviation industry has embraced artificial intelligence and simulation earlier than any other field that also includes medicine as well. And obviously this had to be developed as the technologies and computer science evolved, but autopilot uses standard sensors that are present in different locations on the airplane, whether it's the wings or what is the engine, what is the hydraulics and many other parameters. So the pilot in early stages of aviation industry provided the commands that computer would execute and then later the autopilots and computer technology became so sophisticated that it no longer needed pilot's assistance to identify certain things. And actually autopilot and computer technology can detect and monitor whether conditions better than a pilot or track fuel consumption better than a pilot and also address issues and problems with air turbulence, wind share, and many other problems that could occur during the flight. So it is true that until relatively recently, autopilot was able to do the job well, but still humans were trusted more and we thought that humans can do it better, particularly in very complex scenarios and it was up to the pilot to make that final decision what are the best option in certain scenarios. As you see here, examples are Siri, examples are temperature control at home, automatic automated driving cars, there are tons of examples and we are there, we can rely on them, it is because of our experience with them and when you tell Amazon Echo to set the temperature at 78, it does it, it is no doubt about it and we are beginning to trust that and drones, there are tons of examples of practical machine learning, artificial intelligence systems that actually are in our day-to-day life and we don't notice them anymore. Many, so tell us a little bit about driverless cars and here are some examples, we know Tesla and several other ones are at this point. Do you think that this will become a reality in very near future? Yes, I think that it will become a reality, I think that it will actually probably end up being safer than what we have right now and so I think it's going to solve a lot of problems in terms of fatalities and things of that nature but of course there can be some other problems, newer problems associated with it but I think that driverless cars are actually probably one of the shining examples of how things, how quickly the field has evolved. So this is a perfect example where deep learning and machine learning and also the neural networks that we are using in this technology can achieve what we were only dreaming about a decade before or whatever. It's amazing how fast this field is moving. Many, what do you think about, let's say 10 or 15 years from now, do you envision that all of the vehicles on the freeway will be driverless, they will have sensors and will be able to identify the other cars that are near them and avoid problems, collision and so on or do you think this is still a very futuristic thought? I think that it's going to probably happen before that period of time. Right now, for example, I drive a car that is already doing half of these functionalities without it being labeled as using artificial intelligence and so I think it's even gonna be sooner than that and my feeling is that the acceleration it's just going to be logarithmic how quickly things evolve. I think within three to five years, we will have quite a few percentage of these driverless car or self-driving cars. Now, if you think about what are the current challenges with the AI, we can point to accidents that happened when the car was a driverless car but we got to keep in mind that accidents happened with cars with human drivers every day and thousands and thousands of them, right? And then when AI-controlled driverless car has an accident that makes the news but then many, many accidents on a daily basis do not make the news. Could you give us some other examples of AI algorithms going astray? So, and they were more powerful. So the plane ascended and descended faster but also there was a requirement to change the angle of attack in a takeoff and in landing and to make sure that there is no stall occurring at those points, the Boeing engineers thought that artificial intelligence computer can manage the takeoff and landing much better than the pilot would be able in this type of a scenario. So what happens is that if you leave this plane on autopilot from the early beginning and then you try to correct it because you think that something is incorrect, then the autopilot will override you and take control and this is where the disasters and problems happen. So obviously it's not necessarily AI's fault. It's a fault of not educating pilots how to use this particular system and also freedom for the pilot to disengage the autopilot if the pilot thinks that something went wrong. But obviously a human being and artificial intelligence have to communicate and be able to interact with each other rather than just taking over fully control where the computers are running the show and you as a human being have nothing in your control to do it. So Dr. Zung, tell us a little bit about the past, the present and future growth of artificial intelligence. Of course, we know that artificial intelligence is growing by leaps and bounds but let's look at some of the numbers. What do they show us? As you see in these numbers, they show us tremendous potential for growth and then the growth has been ongoing and it will continue to grow as we improve the systems, as we find more applications and more industries and more businesses that artificial intelligence can penetrate and the future for AI machine learning is very, very bright. So we can see here from the projections that there is really exponential growth as far as revenues and investment is concerned in the AI in comparison, for instance, from 2016 to 2025, from investments of five billion in 2016 to it's estimate is gonna be 37 billion in 2025, which is a tremendous, tremendous growth and we look at the number of devices, six billion devices in 2018 were available using AI in applications, which is really a staggering number. Now, what do you think about, I would like to ask both of you and that is obviously of concern to all of us. Well, for that particular reason, AI replace a significant number of jobs in the United States. Is there a concern about it and what can we do about it? Definitely there is concern, but then one way to turn this question around would be to see how AI could generate new businesses, new jobs and how could it expand its application base and how can it create jobs, right? And you mentioned in these slides that AI is moving into the medical field into education into many, many different fields and then when it expands it requires building system, it requires designing system, requires engineers, it requires computer scientists, requires many, many people from many, many disciplines to make that happen. I think as it eliminates jobs, it creates new jobs as well. Maybe you can give us a conclusion of all the important things related to artificial intelligence and machine learning. So what does the future look like? I think the future is brighter than some people may be fearing. I look in my own practice and right now I'm using techniques in the EP lab that honestly are replacing our expertise in the sense of looking at morphologies and shapes of circuit, different arrhythmic circuits. Does that mean that I fear for my job? No, in all honesty it has actually increased the number of patients and cases that we do because with AI as a tool you take on challenges and you can take on things that had a level of difficulty for example that prior we would say, well we can't do this surgery, it's just that the circuitry, things of that nature or the success rates are just too low. But now if you have AI and honestly it does improve, we see this day to day in the EP lab, it does improve and give us a tool to improve our outcomes. Then you take on more complex, more sicker patients and so yes, it's true that in the course of the case it may be replacing some of the functionalities that we would but when you take a look at it at a bigger picture it actually increases the amount of benefit that these procedures could have, the number of patients that qualify for these procedures. And I think at the end of the day, as Dr. Ajang said, the technologies themselves will require support systems and will require still having that human touch if you will. I also think that to say I know exactly how this is going to affect the job market is inaccurate. Anyone who says that in my opinion doesn't really know what they're talking about but I find solace in looking back at history and every time there's been shifts or paradigm shifts that have been of concern in terms of what their long-term projections will be industrial revolution, things of that nature it always ends up being that there are jobs created that you are not expecting. And so I think that we're gonna be in a good place. To conclude, I would say that machines understand verbal commands, they distinguish pictures they can drive cars, airplanes and all they may argue how well do they do it? Increasingly they're doing it as well if not potentially better than humans. They can certainly do complex calculations better than us and play games better than we do. The question that you have right now how much longer will it be before they walk among us? That's a great question. How much longer? Not too long hopefully. But as we all talk about, we need to adjust. I think when we were a farming society for centuries and then when industrial revolution happened we adjusted and we became manufacturers and all that. And now with AI and artificial machine learning and some of these tools we need to make another adjustment in creating jobs, in doing things better and differently. It's just all a matter of adjustments. So in reality the robots, machines are already with us they're here, they're not necessarily walking on the street but for all practical purposes we use them on daily basis. So this is really nothing new. I would like to complete this presentation with this quote from Sondar Pichai who's a CEO of Google. And he gave this very profound statement when he was asked about AI. He said AI is one of the most important things humanity is working on. It is more profound than electricity or fire. That is truly very optimistic and ambitious statement but we are all optimistic and let's hope that we can gain a lot of benefits from AI and not only in industry or any other field but also in cardiovascular medicine because we are at least Dr. Rosave and myself we're physicians practicing on daily basis on very complex scenarios. And we certainly would welcome AI in our fields to offer our patients safer and better outcomes and also to be able to reduce the cost and expenses with all the procedures that we do and hospitalizations as well. Well gentlemen, it has been a tremendous pleasure to participate with you in this program and thank you once again for joining me for this recording of the Texas Heart Institute educational programs on technologies and techniques. And thank you once again and hope we'll be able to record another program with you in the near future. Thank you, thank you very much. Thank you for having us, we really appreciate this.