Welcome to the specialization on Generative Adversarial Networks or GANs for short. GANs are an emergent class of deep learning algorithms that generate incredibly realistic images. This specialization, will teach you how to build and apply cutting edge GANs. For example, you learn how to build a GAN to generate pictures of people that had never existed, or make someone maybe yourself, look younger or older. You can also use a GAN to take a low res video and turn it into a great-looking high res video. Or on the supervised learning side, if you are tying to classify scratch the objects, or classify medical X-ray images, but you don't have enough pictures, don't have enough examples. You can also use a GAN, to synthesize more data to feed to your learning algorithm. I think GANs are on the path to transform image editing and more broadly, media and entertainment. I'm here today with Sharon Zhou, who will be your instructor for this series of courses. Welcome Sharon. Thanks Andrew, pleasure to be here. One of the things about GANs is that, they seem a bit complex or mysterious. But I don't think they necessarily have to be. It doesn't take a lot to actually implement a GAN, and to the learners watching this, you actually get to implement a GAN in your very first week. Sharon, to demystify GANs, I've heard you when you're teaching, sometimes you use the art forger and the art inspector analogy to explain GANs. Can you say more about that? Yes, I love this analogy, because it strips away the neural network components and the details, the math, everything. It really leaves us with two different networks that are, one is the art forger who's trying to mimic pieces of art or realistic artworks. May be the Mona Lisa, maybe Starry Night. Then there's an art inspector who's looking at a pile of real famous art and also this fake art that's forged by this art forger, and trying to figure out which ones are real and which ones are fake and giving that feedback back to the art forger to improve over time. That's how you get a really great GAN that generates really realistic images. As a GAN Implementer, what one is doing is really implement two neural networks. The art forger, that's trying to paint these amazing looking pictures, and the art inspector at neural network, that's trying to give feedback to the art forger, so that the art forger neural network can become better and better at generating these images. Exactly. One of the really fun things about working with GANs is, I've seen a lot of GAN researchers take a lot of pride in their creation. Because as someone working on GANs, it's like you created your artwork. I'm actually excited about this specialization. Hoping a lot of learners create their own art work, can you feel that pride of creation as well? I would definitely agree with that. I actually think of this as the IKEA effect, because with IKEA furniture, you get to build it yourself. People actually have seen that you like your furniture more if you build it yourself. Just like your artwork, if you build your GAN yourself, you'll like that output, that artwork, those masterpieces you create. But they're again, much more. To learn there, I think one of the things that might be really fun would be, if in this specialization, you implement your own GAN and generate some really cool looking pictures, that you feel proud of and want to share with your friends and family. Or with us, get creative. Finally, GANs are relatively advanced, deep learning topic. Let's also go over the background that this specialization assumes you have. For the background of this specialization, for you to get the most out of this specialization. You should know what a neural network is, including a convolutional neural network and be able to cut out models in Python and also know how to use a deep-learning frameworks such as TensorFlow, Keras or PyTorch. If you've completed deeplearning.ais and deep learning specialization, you'll be all set. It's okay if you're a little rusty on these topics, you'll get to review them in the first couple of weeks. To those of you that had done the deep learning specialization, I hope that this specialization on GANs will be a good next step to learn about a very advanced sets of deep learning algorithms. I'm excited about the specialization, hoping many learners, hoping many of you get started in this cutting edge field. With that, let's get started with course one, where you get to build a basic GAN in your very first week, which is awesome. Second, you get to build upon that GAN, and some convolutional neural network aspects to it, to build a more powerful GAN. You keep improving on it such that your GAN can train stably in your third week. Finally, in your fourth week, you get to actually tell your GAN what to generate. If your GAN is generating types of dogs, you can tell your GAN to generate a golden retriever for example. You'll also get to control the outputs of your GAN a little bit more. So if your GAN is generating faces, you can make this face look younger or older. Sounds like it will be an exciting few weeks. Let's jump into it and lets get started. Let's get started. I'll see you in the classroom soon.
