 Well, geophysical inverse theory has a special place in my heart because, first of all, everything we earth science do is related with geophysics. We deal with something that we cannot see directly. So we have to take measurements outside and figure out what it is. So I knew that geophysical inverse theory was important. When I went to study graduate school at MIT, there was a professor named Ted Madden. He was a physicist by origin, but a famous seismologist, Kei Aki. Newt Madden, Professor Madden, said he did not know about inverse theory because he was a physicist at MIT. But it was Kei Aki, a scientist from Japan, who taught him about the inverse theory. So we are all really excited to learn about this inverse theory from the great man. But Professor Madden is very bad at lecture. So it really, really screwed up our mind, you know? Because thinking about n-dimensional things was difficult itself, but I always felt guilty that I did not understand inverse theory. But it was geophysical oceanographer, Carl Wynch, he's a physical oceanographer. He understood theory, and that's where I learned. Another thing was that in 1956, this my hero, Conelius Lanzos, he's an Irishman, and he wrote a book called Applied Analysis. This book really inspired me and changed. This was before the age of computer, and he described how to calculate matrices, over-determined under-problem system, and what kind of problems are in... So I think you physicists should... This book is cheap, over-addition, can get inspiration from this work by great man. And I think this is really what mathematics is about, you know, being a few steps ahead of your time. Okay? Well, another thing that inspired me was this very complex paper by Albert Tarantola. Albert Tarantola was a geophysicist at IPGP, and this paper was very complicated, but it's basically the essence is, how does one update one's information? There's a figure here which I scanned because the quality is not good, but bear with me, please. So when you have your knowledge, shown here as a big dash line, and then your knowledge has certain uncertainty, everything in the world has uncertainty. There's nothing perfect. And then you have data, and the data has uncertainty. So this dash line is what your view of the world was before any observation, but data comes in. And then how do you adopt your model? How do you update your information? Well, isn't this what you would do? You would find where there is data, you would perturb your knowledge. And wow, it's not just mathematical, it's just common sense. So geophysical invert theory borders on human basic knowledge, our way of understanding the world. So even though I said geophysical invert theory, it's like common sense, all human. If we did not know invert theory, essence of invert theory, we would not exist. Now the final part of my first part, I think I'd say the computer is cost-gift to people with disability. We all use computer, but this really helps people with disability because it can be many things. We normally say when people with disability have three difficulties, we basically classify economic difficulty, social difficulty, and conflict within family. But making them more independent, providing jobs, to get a job, you need education. Then you can improve the lives of the student with disability. I mean, the problem with disability is all disabilities are slightly different. So there is no one-for-all solution, but at least we can take a bite out of a certain part. And this is my philosophy. It has been education, job, and then independence for better living. Now here in Korea, because of government social welfare problem, people with disabilities are employed to do very menial basic job with low ages. Thanks to this AIH, this event detection and object detection, this is very boring. So normal people don't want to do this. So there are now companies in Korea where they employ people with disability to object detection, which goes as a cloud background information for automotive driving and all this AI related for industrial revolution, this big dam of data science. So in order to benefit from artificial intelligence in New York, you have to make the data into digital form so that somebody can copy and download from the computer to run their software. And this business called data labeling is something that they do. People were impressed. I went even further. Now for me, typing on the keyboard is very difficult because I have to press one by one. And as the real programmers don't use graphic user interface or mouse, they do everything with shortcut on keyboard. So what happens is that they end up injuring their hands. And so I have a way of judging whether you are a good computer programmer or not. I said, do you have a hand injury? If they say no, you're not a mature. If you have a serious hand injury, then you are a professional. So because everything is done on keyboard and as you were aware of, there are graphic user interface. And this is fundamental to all high level programming. Now I found out that several real professional programmer could not use their hand because they overused it. The guy here can only type 15 minutes a day. So for their own use, they develop a system called voice coding. This is not to help people with disability, but for their own use. And I show you on the left is a complicated Python code. And on the right is something doing this with voice. No, I'm sorry. Abilities. This is using VEM input as well. This works for all system, Mac, Windows, Linux and everything. Logging, enter, state, import, phrase, OS, enter, args, quote, phrase, void, space, args, phrase, engine, space, star, comma, tip, char, star, comma, const, tip, you and eight, star, comma, tip, size, right, right, quote, our Perrin, enter, looks right, right, delete, state, death, escape, zero, ship cap. Okay, I'll let you follow the remainder yourself. But it takes less than about 10 minutes for this prior to complete all this task. And what's even better is that somebody with having difficulty pronunciation, this voice engine adopts to their language use. And so can even help people that have difficulty pronouncing. State, death, down. Now let's see. Okay, can you see this now? Okay, so what I'm trying to do is to, this is the data of public data collected by a British industry and government to look for oil. And I am now training a team of severely disabled person with the funding from a foundation so that they can make these data all available. I mean, this is an example in UK, but in Korea, all the data that belongs to government available for public so that they can extract new knowledge. So that's the end of my first part. And what do you do with $35 million? Well, I used to take student with disability around the world, around Korea, and every year take them to overseas places. And obviously, computation is very important in aviation. And so I went to one place I visited was Boeing where, you know, you cannot make a plane and say, oh, let's adjust the wing a little bit. You know, you have to get it right at the first time. So that's, this is a picture after we visited Boeing. I moved to the second part of my lecture, which has less slides. No, this was the first part. Oh, I ended up shutting down the second part. So let me just, I'm doing this all myself without the help. So you can see the power of computer. So you would understand why I'm saying computer is God's gift to people with disability. It was meant for people like me. And you are just borrowing from us. So as I mentioned before, my testimony in parliament in 2015, Korea had research vessels, but everything was used by government. And academic community did not have access. We have the ship icebreaker and this thing. But in 2015, I was summoned up to the parliament and asked to testify against the government on this practice because I was very famous. No minister would go against me. That was very, very against, so I had a public thing. But up to this point, I only have people with disability. And I didn't know how much far I can go for our science community. So this was a trial for me. After my testimony, Nature article published an article that says Korea opens up its ocean science and it describes my act. Initially, I became public enemy number one because everybody thought I was breaking the community. But right now, with the ship every year, Korean academic communities are able to not only cover the Pacific, but all the way to Indian Ocean, this is done every year. And so a new law was passed. New law, shared use of the research vessel. And law is quite powerful and they have to do this. And I made a homerun, a history. Let me go, what is the inverse problem? Well, inverse problem, you don't just start with anything, but you have a model that can be your a priori information. And you have data. And your idea is solved basically AX equal B problem. A matrix, B is X vector. So it's just a linear algebra, very simple. But when a professor and when your advisor said, hey, Sangmo, please solve this AX equal B, you're not supposed to solve this. Instead, you are supposed to solve AX equal lambda x eigenvalue problem. The reason is that we don't really care about the solution. We care about the structure of the problem. So I said, when your professor, advisor tells you to solve AX equal B and you solve this, you will never graduate. But when you become smart and then solve this problem, your professor, hey, Sangmo, we'll give you the degree. And this is very important. And I cannot stress the importance of this inverse theory and the way it works as a basic human way of understanding nature and phenomena describing things. But in machine learning, everything is lumped into a small chapter called regression. Of course, fitting a line through the data is inverse theory. But I was appalled by when I talked with engineers and they all say, yeah, well, no, no, it's much more than that. You are missing the fundamental way we conduct science. So when you do global tomography, you need to understand observation and you use inverse theory. One good way is global tomography. Another, the reason I praised Ponelli-Lanzos is that everything, all the matrix problem, you are never given a square matrix to invert and come up with a solution. It's either over-determined or under-determined. And you seldom use this, but he philosophizes a way of understanding all this problem under the name of singular value decomposition. So I tell people, if you understand SVD, then there's nothing else to learn. Just drop linear mathematics class. But then there is everybody that's talking about deep machine learning and learning how great it is. And so, well, I like to be on the band, but I'm skeptical. Well, for one, we don't have enough data. So what people do is that they use some kind of numerical simulation to make artificially more data so that this can feed into machine learning or deep learning. Of course, the data itself can give us ridiculous answer. So in some try to improve this by selecting a sample of data, whether it matches with simulation so that we have a control on the outcome. The reason I say this is that if you just rely on data, you don't know where you will be ending up. You can be really stupid. Because many things in Earth are not yet observed. Basically, in my view, to understand linear inverse theory, I think machine learning is very good. Because machine learning, unlike deep learning, there is a theory behind it. And there is a rich flavor of theory. For instance, if you want to understand a singular support vector machine, there is a beautiful mathematical story and theory behind it. Decision tree, entropy, and all this thing. Basically, machine learning is OK. But I have trouble accepting deep learning. Maybe I'm not that smart, but anyway, it's basically optimization of data. So this is, as you know, familiar program of deep learning neural network. You try out, and because you know examples, what it really works, you back propagate it and try to adjust the coefficient. So it's basically linear algebra in different disguise. But when it gets complicated, it's like a black box. Everybody thinks it's like a black box. And if I gave a positive spin on it, I tell people, when a finite element method was created, engineers knew how to use it. But mathematicians had difficulty understanding. So maybe it's like finite element. Maybe the practicality goes first, and then the theory follows. What they do is that, using all this data, what they do is that here is a basic essence of deep learning and all this thing. Here we have round and blue and green symbols. It's hard to differentiate draw lines to draw the border using straight lines. What they do is that make some transformation and things so that these two sets of data apply the same thing, are discernible in some m-dimensional way, which is difficult for us to visualize that. But we have to believe that. So basically, it's about classification and dividing. So if my data follows here, it'll be dog. If here, it'll be cat or whatever. But as I said, this kind of method works when there is convolutional neural network, reinforced neural network. All works well when there is a lot amount of data and image. So I think the atmospheric science is the greatest beneficiary of artificial intelligence, as we understand at the moment. Now, some issues, we don't need an accurate answer. We have to make decision quickly. Like if there is an earthquake, what do we do about it? These kind of things, when you need action immediately in real time, almost real time, that's when deep learning becomes important. The accuracy is not a matter. We have to make quick decision. But here is some failures. Here, it says it's written in Korean. Well, it says machine learning identifies this as deep learning identifies PIC, but add real little bit of noise. And when you show this, it's an airplane. Here, banana, but they say it's a toaster. It says it's a stop sign, but it says speed sign to go 45 kilometers per second. So it can, with a small data, it can do a miserable job. Also, it cannot be generalized, because here you have the problem. You change brightness, or you change the grid size, then it cannot adjust. It becomes a whole new set of problem. And I think this is why we need so many data scientists, because millions of them will be working on a different size pixel problem, not 18 by 18, but one will work on 20 by 20. So of course. So I think inability to generalize, because we don't understand the working of it, is a bit serious issue. Now, here you have the Google, I mean, the Tesla car bumping into a van, because they did not understand what this was. So now there is a strong movement, because sometimes, like in medical science, the ability to explain is very important. So there are now areas of interpretable machine learning and explainable AI. But I think this is not full force yet. This is just starting, but still, a lot needs to be done. I think this all comes down to how do you add a priori information to make sure that AI does not give crazy solution. I showed you this picture by Albert Tarantola. We have the dash line as a priori information, and we have data, so we do a reasonable job of updating. I think this is the important shortfall of machine learning, deep learning, whatever. We need way to provide guidance, a priori, so-called a priori information. Again, I said, you never have enough data. And so although I teach computational science, I'm not a great fan of it, except for pedagogical and educational purpose. If people ask me what inspires you, I said, not machine learning. Now what inspires me is the mystery of plate tectonics. There is a problem called LAB, Lysosphere Ethnospheric Boundary Issue. We know that crust and mole and mantle, there is a very well-defined seismically well-defined mole. But the real physical action takes place at the boundary of Lysosphere and Ethnosphere, so they call this LAB, Lysosphere Ethnosphere Boundary. But the problem is that this boundary is not sharp and well-defined. It's very hazy. So with the development of measurements and instruments, before this problem was intact, was not treatable, now it's becoming feasible. So I participated in bringing my ship together with the very professional scientists from Japan at University of Tokyo, ERI, our competitor. They don't think that's as competitor. And with Korean ship and Japanese instrument, we went to the oldest piece of the Pacific and for one year made broadband OBS and OVM measurement. Why we chose here is that way, way back ago in Jurassic, when the Pacific was forming, the plate moved in different direction than now. Now it's along the East Pacific rise, but before there was a triple junction and it moved differently. So if we are good enough, we can see two different LAB and if we've discovered that this will be a remarkable achievement. So this project was called Pacific Array and we are working on this. Now this was instrument that was used on Korean ship and after, I mean OBS, ocean bottom size monitor as a long history, but when it comes to deep ocean bottom electromagnetic sound which will provide us information, not on rigidity, but on the conductivity and resistivity, electrical property, the flow property, fluid like property of ocean, not match has been done. So this is a very cutting edge field. I had my student look at this data collected over a year analyzing and he found out, guess what? None of them, his data did not fit anything. So, well, it's like a person touching the elephant. You know, when you have just one observation you can claim but as you have more observation, Earth is definitely more complicated than this two layer, three layer model, it's as convection, melt, hydration cracks and all this thing. So I said, wow, this is really great when everything did not go according to our prediction. That motivates us to do more. So what do we do? We have to develop the next generation of instruments for earth science and this is what real physics is about. Now, I popped this for interesting reason. There was an earthquake and this earthquake was observed at land station and two OBS sites, same earthquake. On the normal ocean bottom seismometer, you have a lot of noise in horizontal component because of movement of water current and wave. It's very noisy. You cannot get land quality. Using this kind of data, your result will be hazy. You will not be able to decide what is the right model. So I have really respect for this Japanese scientist. They come up with a very complex system where they were able to decouple the logger, battery and the sensor in mid-water so that water current will not affect the sensor. This is not yet there. This is the prototype of their model. So this kind of improvement will really help us to understand this. Also, there is OBEM where you measure the electric field across these poles. So there are issues in the Pacific but I have, through my heroic effort facing the firing squad, I can now go to Indian Ocean every year. And in Indian Ocean, there is this problem of how hotspot interacts with mid-ocean bridge. This is a classic bridge-plume interaction. And French scientists and German scientists have done a lot of work in this area near Reunion Island and they could see the hotspot, the plume, magmatic plume from Reunion affecting the mid-ocean bridge. But when we did a survey in the North, we could not see the hotspot. We think we are seeing hotspot coming up from East African region, a far region, much deeper, but we are seeing this. It's shown in passive regional seismic tomography but there is no way to prove this. So the reason I'm really keen on developing instrument is that I will take this instrument and go here and see whether the plume is coming from Reunion hotspot or it's that big, massive plume that is splitting the East African rift, a far plume, all the way down. These kind of things are like questions that it's like looking at another point on another planet. This is at the level of Carl Sagan, in my view. So first, I have to develop instruments because to do this, I need a lot of instruments and deploy them for a year or more. It's like video tape. Nobody will rent your video tape for one year. Only a blockbuster, five, 10 days at maximum. So I need my own instrument and this is the area that I will be looking at. Another area that I'm going to look at is Alaska Illusionar. You know, the United States is very afraid of big tsunami and it turns out that USGS, under the budget cut restrainment in the 80s, they had two research vessels, they got rid of it. So they got good grading for economic, I mean for being a saving budget, but they got rid of the ship. Now they need to ship, they don't have to ship. So they come to me, look at there. Through the effort of Professor Lee, Korean will become available. So I am going to save the property and lives of American people. And this is just my outrageous claim, but because of my effort, Korea has suddenly a lot of ships, scientific ships. And so we can do this and we'll be, I mean, if there is any burden who is in the spirit of inter-age, who wants to come and join our ship, we will try to accommodate. Now the planet A. This is a very good book written by Charlie Langmuir and Wally Broker. It's called How to build a habitable planet. It talks about the geochemistry to reach where the point we are, but we are ravaging our planet, okay? So no wonder. Science has helped to improve lives, but it's going to kill us. Also there is a plan, this is a wonderful book by Paul Collier called The Bottom Billion. Now there are certain number of countries that are living well and there are countries that will soon become better, but there are bottom billion that will never become better. And his advice is that aid is not the problem. Inspiring local champions and giving them the power to change their own society, whether it takes time is the key. So, well, it should be flipped, but here the yellow countries are those countries that are suffering. They're mostly in Sahara and some in Asia. And our ship bows from Pacific all the way here. So this is, I'm telling people now, we are now in the position to give out, we used to receive foreign aid, now we are in the position to give out aid. We need to build capacity of people in this region. This is a good way to help, not direct material aid. Well, material aid is also good, but this is not. So, with 1010 project, I have started so-called data science competition to relate on earth science issues. And recently, several months ago, a big corporation in Korea. This is the third largest corporation in Korea, ranked three at the moment. One of their subsidiary company called SKENS, which stands for energy and solution. They are in charge of powering electric power plants in Korea using gas. And so in this new era, they have to adapt to new change and they approach me. Because I was so famous in Korea. I'm not, I'm a solid earth scientist. I don't have much experience in environment. But you know, with my fame, they put me in charge so that they will invest every year, $15 to $20 million so that I can help young companies to do important research and such like. So, another area that I'm focusing is not only data science, but to actually build hardware and instruments that we can deploy from the ship and make new ground, new understanding in earth science. So this is a good one is by DARPA, U.S. Defense, and they do a lot of things. And it's very strenuous and competitive. And also I have this commitment in disability. So I use young kids with raspberry pi, and at the same time are doing all these embedded solutions in the board computer to inspire them, age 10 and 15 into science so that they will also be an important contributor of our human knowledge and effort. So this is my final slide. As Alec mentioned, it was 2020 postponed two years will be next year. Hopefully we get better and be able to do this offline. But please mark your calendar. And I hope to see you there. Thank you. Thank you very much.