 And we're back live. We are young talents making way only here on Think Tech, Hawaii. I'm Andrea Gabrielli. I'm your host. And every Tuesday, we keep an eye on the future with our most brilliant school students as we talk about their science projects. And joining me today is James Shigemoto, who just graduated from Kalani High School and carried out a science project regarding climatic predictions, and particularly global weather patterns. Welcome to the show, James. Thank you so much for having me. It's so interesting. Very nice to have you here, especially at the beginning of the dawn of this new hurricane season in Hawaii. So terrified. Interestingly enough, I think a while ago, or I'm pretty sure it's still going on, it was actually La Nina season that was forecasted by NOAA. And interestingly enough, my data set, oh, it's kind of cold. This is exactly what we are going to be talking about today. So why did you get interested in predicting weather? So my first interest was computer science. I always wanted to do something that related to computer science. And I love psychology. So I think, what if I could combine them? And when I first found out about machine learning, I was so happy. And then thinking, you know, let's try to predict cancer. And then, as you can see, it didn't really work out well. Was that part of a previous science project? No, this was completely new. I'm actually continuing that, the cancer research right now just as a side project. But I'm also continuing my weather prediction. But the way the weather prediction happened was I was talking to a mentor I had at the time. And she suggested, you know, predicting cancer is really difficult because of the data sets. That's the most important thing for machine learning. And I decided, OK. Well, weather, we're in Hawaii. Sometimes it's hot, sometimes it's cold. Why is that? And then that became El Nino because around that time, one of my family members was in Puerto Rico during one of the hurricanes. The hurricanes. El Nino, I believe, let's have our first slide up so we can understand better what El Nino is. Yeah. So El Nino is a change in both wind direction and speed, as well as sea surface temperature. Right. It really depends on how it works. In certain seasons, it's expected for it to be higher than normal. So I believe during the December, around the December season, well month, it's supposed to be slightly higher than usual. And because of that, you have to change it to, you have to change the way it's measured. So in this cartoon, for example, the one on the left illustrates a normal weather pattern. So we can see we have South America and Southeast Asia. The trade winds basically are responsible for blowing hot, warm ocean water towards Asia. During El Nino years, we sort of see a reverse of the patterns. So how often does an El Nino event occur? And what are the consequences of such inversion, oscillation in the wind and ocean temperatures patterns? So El Nino, it really varies. So it can happen every two to seven years. And in that period, it depends on the region. So in certain places, you'll have flooding. In certain places, you'll have droughts in certain places. Because of the droughts, the trees will become a lot more brittle, which can make fires a lot more likely, and do those fires in the lack of water. You're not going to be able to put them out. So due to that, El Nino has a wide variety of issues, just depending on the region you're in. So for Hawaii, that's going to be more hurricanes, more flooding, more water. And since our hurricane season here in the Central Pacific begins actually June the 1st. So that's only a couple of days from now. I invite you to join our conversation here. You can just call this number here, 808-374-2014. And you can join me and James while we're talking about hurricanes and El Nino. And let's have our next slide up so we can understand better some of the effects of these El Nino years. So here are some figures, some examples. Can you describe what we have here? In only one instance of El Nino, over 24,000 people have died, as well as over $34 billion in damages. And that's just in one instance of El Nino. That's not a combined event. That's 1997, 1998. Yes, that was a strong. That was considered one of the worst in history, if I remember correctly. Right. When was the last one? Well, the last El Nino happened. In 2015, yeah, we had a couple of. It happened a while ago. I know we just had El Nino last year, but El Nino happened like three, like two, three years ago. Three, four years ago. Yeah, it happened a lot. And so we have a variety of, for example, flooding. We saw some pictures before. Charleston, Tennessee. We saw fires in Indonesia. Maybe let's have the slide one more time up. So these are talking about the figure on the right. The fires are large. I mean, we're talking about, this is a picture from space. We can see Borneo and Sumatra. So really large events we're talking about. El Nino, it's insane, the scope of it. And it's insane the damage it causes to people as well as property. And that's why I think it's very important that we're able to identify a possible El Nino event. Because while we are pretty much unable to stop it, we do have an opportunity to try and keep people safer and alert them. So the work you carried out was basically to try and predict the future El Nino events based on data that we have, weather patterns or data that we have. So that's very interesting. How did you do that? So I first started off trying to learn exactly what machine learning was. Turns out it's a very big scope. It really is, yeah. So what I did was I was looking for data that I could trust. Because for me, that's the most important thing. I want data that is included by any other motive. So I chose to stick with NOAA. NOAA's databases. Yeah, because I trust them as a reliable source for all of this. And the problem is they only go back to the 1950s for sea surface temperature, which is genuinely what they believe is a decent or at least a good operator, predict El Nino. If you go on their website, they have a bunch of different methods predicted, but the two they most care about are the sea surface temperature in region known as Nino 3.4, which is around the equilateral of the Pacific, as well as the wind direction and wind speed in that area. So they're able to measure where it's going, how fast it's going. That's generally what they care most about. What are some of the effects of an El Nino year, for example, here in Hawaii? I believe we have a slide about this. So maybe that can help us to end. Oh, OK. Yeah. So an El Nino year in Hawaii can be, it varies greatly. Sometimes it could just be nothing. Sometimes it would just be unnoticeable. But most of the times it would be hurricanes, floods. This is a modest image, so from a NASA satellite. We are looking at, is that Hawaii in the red square? Yeah. Yeah. And this was 2015. So we see three really strong-looking hurricanes. So this is something that can potentially affect Hawaii. But you mentioned also droughts in Indonesia and wildfires are flooding. It's something that really doesn't affect only the Pacific, but it's really something global. Yes. So let's see some of the data that you have used in terms of trying and predict future El Nino events. So these are three figures that you brought us. So what are they? So the first one is known as the anomalies plotted through the years. So this starts at 1950 and it shows all the anomalies plotted. So as you can see, the higher-stiking anomalies are kind of indicative of an El Nino event, while the lowers can also be said to be a La Nino. So what's an anomaly? How is it defined here? An anomaly is defined as, well, El Nino is generally defined as three season period in which, well, five season, sorry, five season period in which the average sea surface temperature is either 0.5 degrees high or 0.5 degrees lower than South America. Closer to South America, as we showed in the cartoon in the first slide. Yeah, OK. And the second one is wind speed and sea surface temperature graph. I was really upset that we only were measuring the wind speed in around the 1980s, 1990s. It's fairly new. So the blue one is basically the wind speed. And what's the orange one? The orange one is the sea surface temperature. That's why there is a big collusion, because I had it on the same graph, so it would be easier to illustrate. And the orange one is just the overall sea surface temperature. OK, so the scale on the left is only related to the temperature. Yes. What's the average wind speed that you observed? The average wind speed I observed. Oh, but it's more or less the direction it's going. Oh, the direction. Yeah. But the average wind speed was, I think, or not, or so. I'm not 100% sure, and that fact to be clear. No, OK, but we're talking about the direction. Yeah. Yeah, OK. And so in La Niña years, basically we have strong trade winds in the southern hemisphere, which blows from South America to Asia. And then in El Niño years, do you observe a reversal or lowering the intensity? You observe the reverse of El Niño. So El Niño and La Niña are both extreme effects, but they're the opposite side of the coin. So while El Niño is hotter, La Niña is colder. While El Niño goes one way, La Niña goes the other way. So they're both extremes just in the separate direction. And the third figure that you brought us? That one's my favorite one. All right, this looks complicated. That took me an hour just to graph because I wanted to make it look nicer. So that's actually a three-dimensional graph of what it looks like. So the x dimension is each observation. So you can see it progress. Each observation of a sea surface temperature and wind direction. Yes, OK. While y is its actual sea surface temperature. So you can observe each plot over there. And then the z represents its wind speed. So you can observe it. Unfortunately, I couldn't give you the gift of it. But I was able to rotate it around to show another perspective. So the top down looks a lot more interesting as well as from a 40-degree angle. What's telling us this figure? It's telling us that there is actually a huge correlation between sea surface temperature, wind speed, and an El Niño event. I wanted to create this just because I wanted to be sure for myself. I did trust Noah, but I wanted to see it for myself that there is a correlation. But that's what science is about. You're verifying and trying to see, OK. Very good, OK. This is a very good project. And so based on this Noah date of sea surface temperature as well as wind direction, wind speed, yeah. Oh, I believe we have a phone call here. So let's try to answer a call. So I press this, yeah. OK, let's see. I'm still new how it works, yeah. OK. OK, but in the meantime, as we were talking about this, so these El Niño events really, it's something that we can experience effects in the global Pacific as well. Yeah. OK, so maybe this is this time. Let's try. Answer. OK. Hello, hello. Hello, hello. Hi. Hello. Hi, hi there. How are you? What's your name and where you're calling from? Hi, this is Linda. I'm calling from Manoa. And I have a question for James. It's such an interesting conversation you're having. Awesome. Thank you, Linda. Thank you. I really appreciate it. So what's your question? You have decades of data. And I was thinking, oh, a lot of data. How did you find computers fast enough to process all of your data? That's an amazing question. So there are a lot of open source resources. So you can actually compile your data on your own computer. But what I found to be the best was what you do is you actually have it run on Google's computers or Amazon's computers. So they have a service where you're able to run all of your data and have them compile everything for you. And you just get the results. So. Oh, wow, OK. Yeah, it's a very powerful method. It's Google Cloud services or platforms. I'm a little confused. If you want that, but it's such a wonderful resource to have. When you first sign up, you get $300 in free credits for a year. And it's so powerful. A breakdown of it was I used GPUs. When you compile data like this, you generally want to use GPUs because they work faster. So I used four GPUs, ran it for 24-7 for a month, only brought my bill up to $100. So it's a very powerful, powerful thing to have. And I highly recommend that for anyone who's working with large amounts of data for machine learning. Thank you. And we also thank Linda for asking this question. Thank you, Linda, from Manoa. Thank you for watching us. Thank you. So this is how long did it take for you actually to compile using this technology we just mentioned to finish from the beginning of your science project all the way to the results? Honestly, it's hard to say because it's an ongoing work. Because I just left it on. I'm like, you know what? Maybe I'll get better results this time. Right, right. I just left it on. And then maybe about a week from my science fair, I grab it and I'm like, OK, this is my best results. So I start circling it. OK, time to take a break now. But we'll be back soon. Thank you. Aloha. I'm Kili Ikeena. And I'm here every other week on Mondays at 2 o'clock PM on Think Tech Hawaii's Hawaii Together. In Hawaii Together, we talk with some of the most fascinating people in the islands about working together, working together for a better economy, government, and society. So I invite you into our conversation every other Monday at 2 PM on Think Tech Hawaii Broadcast Network. Join us for Hawaii Together. I'm Kili Ikeena. Aloha. Do you want to be cool? If so, watch my show on Tuesdays at 1 called Out of the Comfort Zone. I sang this song to you because I think you either are cool or have the potential to be seriously cool. And I want you to come watch my show where I bring in experts who talk all about easy strategies to be healthier, happier, build better relationships, and make your life a success. So come sit with the cool kids at Out of the Comfort Zone on Tuesdays at 1. See you there. And we're back live. We are young talents making way, talking about weather, and with James Shigemoto here on Think Tech Hawaii. Hi, thank you. So you've got these huge data sets to try and predict future weather, El Nino. You process them. But how do you actually train a computer to predict future El Nino based on these extrapolations based on past data sets? So there are a lot of different methods you could have used to do it. And at the time, I was not near main expert on computers, well, not computers on machine learning. So I decided that the best course of action for me was running an optimization. I used something known as genetic algorithms to optimize it to make it work the best with the least amount of time on my part. So I could focus that on reformatting my data. The science, yeah, the science behind it, yeah. So let's have our next slide up. I believe it's slide number five, the one you brought us. You used genetics algorithm for this particular. What is a genetic algorithm? A genetic algorithm. It basically tries to mimic Charles Darwin's theory of natural selection. And we are looking at him on the right hand side. Is that your slide? He's a very good looking man. We're basically trying to mimic his approach of natural selection into a computer. So what you do is you have to, I don't want to say random because that's kind of undermining what it does, but I don't know a better word. So basically two random possible solutions. And what you do is you run them multiple times. Well, technically you run all of them multiple times and you find the highest one. So you basically list out its solution. So I guess it's not random. Run the two best fit ones and then you give it a random mutation, like a random change. And you combine them to create this new one. And you continue running that until you get the best fit score or its highest accuracy in my case. So instead of me doing calculus and trying to calculate the derivatives for it, I actually just use a genetic algorithm to do it for me. You mentioned this is similar to machine learning, but it's really deep learning. Machine learning and deep learning are very similar. Deep learning tries to mimic more of the human brain. So they're basically neurons. So it's more like neural networks. Yeah, neural networks is way closer to that. Machine learning is just a broader source. So what I generally use for machine learning was a support vector machine. While for deep learning, I don't know what the correct term was, I created something using TensorFlow and Keras. So I use TensorFlow as a back end and Keras as my front development. And I believe we have a slide which really can emphasize more of these concepts. You're talking about the slide six, I believe it is. Yeah, maybe if we can have. Yeah, this is exactly what you were mentioning. So machine learning, it's very similar to artificial intelligence. It's actually underneath a subsection of it, but not exactly it. It's more of a subsection. It uses a lot of data. And that's the most important thing for it. You don't have a lot of data. Well, that's also another thing. If you have too much data, it's not good. So there's an issue known as overfit. And also, you don't want to start fitting the noise which is present in there. So you want to train this computer, this brain, this brain of the computer, just right, just enough so that we can give us meaningful scientific results. You don't want to over-train it. You don't want to train it too less as well. So you're using this for, and we're going to have a look at your results in a minute. What are other applications for machine learning? I'm curious. So machine learning, it's such a wide thing. Everything, well, all of our smartphones use machine learning. So with a nice example, text prediction uses machine learning. It looks at the most likely outcomes you're like, hey, I'm going to. You're going to probably say home or the mall instead of an outer space. So they look at, for example, in this case, data, most common messages, I'm going home. I'm going to the hospital. I'm going to school. On the way more technical side, what they're trying to do right now, or I know one researcher is, they're trying to remove humans from human testing. What they're doing is they're putting all of their data, all of their genetics into a computer. And then they're giving the other data for the medicine. So for example, a certain molecule that has a certain effect, like an inhibitor for example. And what they're doing is they're testing to see how that works on the computer, instead of doing it on an actual human. Because on a computer, you could kill the human, but it's just a bunch of data. But men, people are always there to control. People are always there to control that. Yeah, yeah, make sure of that. Not like crazy movies with the stuff that happens. I don't know. Right. So let's talk more about your results. What are the outcomes of this modeling? And I believe we have our last slide for today. Yeah, OK. So what are your results? So my results for using a support vector machine, I use a method known as support vector regression. So normally, a support vector machine is just good for classifying. But in my case, I used it to calculate the probability. And that's difficult because that's a zero, well, a negative infinity to an infinity possible result. So it finds the most likely outcome. And what I did was, when I finished doing all the work, I got a 53% accuracy rate consistently. Well, is that considerably? I believe it's considerably good getting the exact temperature to the 10th decimal for the 3% amount of the time. And I looked through that past data, and most of the time, it would get about 90% of the time when it's not looking at its 10th decimal. Right. Do you use your model to try and predict future events at this time, at this stage? Yes. Once you have your data fitted, you have to write about five to six lines of code to have it predict. So it's looking at its next possible outcome or the outcome of the next possible outcome. And part of the code you used, is that the box on the right we're looking at in this image here? Yes, the box on the right shows the prediction value. I kind of cleaned it up because I didn't really like people seeing all my ugly code. My code is very ugly. Code is beautiful. No, I mean, code is beautiful if you make it. It's powerful. You can do a variety of things. My code was kind of jumbled. I mean, it works. But if a programmer sees it, they'll probably not understand it, because it's really ugly. But what are the outcomes of this code? So can you tell us when the next El Nino is going to be? Or are you still working on it? I'm still working on that. It stays right now. I can predict the sea surface temperature, which is important. What I'm working on right now is predicting its wind speed temperature and having that combined together with my sea surface temperature to create nice correlational data. And you've been presenting this science project of yours at the most recent Hawaii state science and engineering field. How was this experience sharing your results, your code, your work with the judges and your peers, your fellow students? It was both really stressful and really humbling. I got to see a lot of great minds talk about not just professors, but some of my other competition. They had very, very interesting and intricate projects. And it was very stressful, because all I'm thinking was, don't screw up, don't screw up, don't screw up. Just stop staring, James. So very, yeah. And now you mentioned you graduated from Colony High School. So congratulations on your graduation. What's next in the future? You're going to continue doing, you know, you mentioned your interest in psychologists as related to what we're doing here, what your project was about. Are you going to continue doing this as part of your college experience? Yes. Machine learning, once I got into it, I'm kind of hooked. I'm going to continue this in my college years. So I'm going to be majoring in statistics at CopyLine and Community College. Oh, OK. And then going to go over to UH to see what takes me after that. Right, but your interest is basically statistics, psychology, machine learning. But what would you see yourself doing in the future, for example? As a scientist or more like psychologist, because we have these two branches. I see myself being a data analyst. I really enjoyed and hated working with the data. I enjoyed it because when you've got. It's a love and hate relationship. You know, when you get it right, you just feel so happy. When you get it wrong, you just feel so sad. No, no, absolutely. But I did enjoy it in the long run. And I did learn a lot about how to clean up data, a lot about how to test data to see, oh, is this data helpful or is this data just noise? And I became decent at that. And I was able to become better at that. And so I feel like when I get older, that's why I'm taking statistics. I want to be a data scientist or data analyst. And we have a variety of data which is now being acquired by sensors or space. So big data analysis is really something that everybody is focusing on, including the economy. We have about one minute left for our conversation now. So how would you like to summarize your work that you've been doing as part of your science fair experience here in Honolulu? The best way I can summarize it was a lot of research, a lot of going on YouTube and trying to figure out, OK, so this correlates to this. Time to write this code, a lot of. And even more research coming, I guess, yeah? OK, thank you very much, James. Thank you, James, for being with us today. And so you've been watching Young Talent's Making Way on here on Frank Tech, Hawaii. And I'm Andrea Gabriele. I'm your hostess. And next Tuesday, we're going to be back for more. Stay tuned.