¿Qué demonios hago en Corea del Sur? - Deep Learning Camp Jeju 2018

Key Takeaways

I'm going to tell you a story, okay? Well, there's one hour left before my flight to Manchester, and pay attention to the route I have to take to get to South Korea. As I said, I'm in Helsinki, and to collect the supplement, I have to fly to the UK, to Manchester, and on a flight that's 4 hours long, the clock goes back two hours so I can go to the gym, and I'll arrive at 9:00 AM. From the airport, I have to wait until 11:40, that is, a three-hour layover, to get from Manchester to Hong Kong on a flight that's going to be 12 hours long, and on top of that, the clock goes forward 7 hours, so I'm going to arrive a day later at 7:00 AM. At that point, my plan is pretty messed up, but after Hong Kong, I have to wait at the airport from 7:00 AM until 3:00 PM for a 7-hour layover, and then I'll fly from Hong Kong to the mix of people on a 4-hour flight. [Music] And I've left the airport, and it's started to rain. It's pouring rain because apparently a typhoon is expected tomorrow. Anyway, this is the first day, I 'm already at the hotel, everything is going smoothly as expected, and I'm going to rest because I don't even know how many hours of sleep I've accumulated or subtracted. So, tomorrow we can't... tomorrow everything starts, which would be... bed, sleeping crosses, taking advantage of these precious slow-motion shots that I'm going to use to answer a very important question: what the heck am I doing in South Korea? The answer is that two months ago I applied to participate in a summer camp where 24 machine learning researchers from all over the world would be selected to live together for a month on the beautiful island of Jeju, South Korea, to work individually on a deep learning project. The main sponsor of this event is Google, and therefore we have expenses paid, a salary for this month, and a thousand euros of cloud resources to spend on our project. In other words, a unique opportunity to travel to a paradise island, live alongside other professionals who love the field of machine learning, and work individually on an amazing project. Sounds good, right? This is from... Berlin, capital and it 2018 [Music] Hi guys, I barely have time because we're doing a lot of things, but anyway, now we're going to have a session on your tips, Tensor Professor, you need CPUs. These are the processors that are specialized for tensor processing, and here we're going to be using them with the Google Cloud platform, which is very interesting, and they're going to give us a course now on how to use it. In short, everything's great. The food is spicy and strange for breakfast; they have rice and soup, a bit weird. And the typhoon is coming, but it's not arriving yet. Let's wait and see what happens. [Music] And we're continuing with a bit of introduction. We all had a 5-minute talk this morning explaining how to play Mine, said [Music] Carrizosa. Without explaining it, we already ate at noon, and now to maintain a game again, in terms of investing money in different projects to switch to waiting. I don't think I've even explained what project I'm here to do; I'll tell you so you understand the idea. In the project I'm going to work on, we first have to travel back in time, specifically to one of the news stories included in the April news video. Let's recap it. I want to thank Carlos Jose Diaz for sending it to me via Twitter, and what a crazy article it is! It's about an image synthesis system, and in this case, we have images where a subject appears in a specific pose. We can define a new target pose, and the system will learn to create a new image where the subject appears in that pose. Yes, you heard right. You take a reference image, for example, this one here, and then specify which pose you want it to imitate, and that's it. The system is able to generate a new, quite realistic image where the person appears in this new pose, even maintaining consistency in the rest of the image elements. To achieve this, unlike other solutions, the presented system decides to divide the problem into different tasks. First, the system learns to separate the image background and the person into two layers. Then, the person's figure is divided into different parts corresponding to arms, legs, and other body parts so that each element can be repositioned. The desired pose is then synthesized, filling in the gaps in the background layer that have been revealed after moving the body elements, and finally combining all the parts to obtain the final image. It's that simple. This entire system we've described will generate this architecture and will include the evaluation of a discriminator network, thus forming a generative adversarial network structure. This ensures that the results obtained have more realistic details and are consistent with the original pose. A very interesting project that could be applied to graphic design tools for image and video manipulation, and as Carlos Jose points out in his tweet, this, combined with other face and voice synthesis projects, brings us closer to a future where we won't be able to tell if a video of a particular person is real or not. We will maintain a positive view of this project, as there is a small possibility that we will hear more about it on the channel in the near future. Stay tuned, because we will definitely hear about this project. Seeing this system capable of synthesizing images from a given pose made me wonder what would happen if we combined this with another system that, starting from a If the image were able to predict the next sequence of poses, my proposal is to use the image synthesizer to generate an image that can be given as input to the pose predictor so that it uses this information and predicts what the subject's next pose will be. Then we can use the predicted pose to synthesize a new image, and so on for several generations, creating a loop that, if it works, will allow us to generate a video predicting the subject's future movement on screen. Similar works exist that have used a person's pose information to generate videos with very promising results, as you can see in these examples. My work this week will be to implement this video generation system and compare it with the results of some of these projects. As you can see, it's a super interesting project that I'll tell you about in more detail next week when I start working on it more thoroughly. This week we've had our share of work [Music], but to be honest, the goal of this week was for everyone in the group to get to know each other, and our weeks ended up looking a little more like this [Music] [Music]. I think it takes it YouTubers here sometimes [Music] and [Music] yes [Music]