Currently, generation tasks that create new data such as images and sounds are attracting attention as an application field of AI. Among them, generative adversarial network (GAN) is the technology that is attracting the most attention. Since it was proposed by Goodfellow et al.
On the other hand, with the exception of some companies, there is a problem that application to business has not progressed. Therefore, Skill Up AI will hold a course where you can systematically learn about various GAN derivatives .
In this course, you can learn about GAN from the latest technology trends to how to apply it to business through hands-on, and it is possible to solve the problem of “I do not know which GAN to use!”
- GANs Evolving at Extraordinary Speed
- GAN active in many fields
- Use case of GAN
- Understanding 100 GANs around StyleGAN
GANs are currently one of the most researched models in the AI community. Since it was proposed by Goodfellow et al. in 2014, many derivative models have been proposed. The number of papers on GANs is increasing year by year, and even now, about 70 papers on GANs are submitted every month. In addition, GANs are being applied not only to image generation, but also to various fields such as video generation and audio generation.
In this way, GANs are evolving at a tremendous speed, but because too many GANs have appeared, people say, “I want to solve business problems with GANs, but I don’t know which GANs to use.” The problem arises, “I want to do something with the data I have, but I don’t know what I can do.”
GANs, which are rapidly evolving, are being applied to various fields. Below are some of the tasks in which GANs have been applied.
|image task||image generation||Learn the image data you have and generate a new image|
|style conversion||Converting the style of one image to that of another image
(coloring line drawings, converting photos to anime style)
|High resolution||Increase the resolution of low resolution images|
|image generation from text||Generate images from text|
|time series task||voice generation||Learn the voice data you have and generate new voice|
|voice conversion||Convert one person’s voice to another person’s voice|
|natural language tasks||text generation||Generate meaningful sentences
(generate sentences from words, summarize sentences, etc.)
In addition to the above tasks, applications are progressing in various fields.
Yann Lucan, director of Facebook’s AI lab, refers to GANs as ‘the most interesting idea in machine learning.’ In this way, it can be expected that the applied research of GAN will increase more and more in the future.
Here, let’s take a look at an example where GAN is actually used in business. ” Kiwami Prediction AI Human
” provided by CyberAgent, Inc. generates a person model suitable for the target of each company or brand by GAN, and can reduce the time and effort required for conventional advertisement shooting. In this ” Polar Prediction AI Human “, GAN plays the role of generating a large number of person models. By generating a large number of fictional characters, it has become possible to select models that are more suitable for each company or brand.
In this course, we will first learn about StyleGAN proposed by NVIDIA researchers . StyleGAN contains many of the important elements of previous GANs, and is the foundation for the GANs proposed later. By learning about StyleGAN, you will be able to easily understand the main points of other GANs.
In addition, we will explain the implementation code of StyleGAN and not only learn more about the role and function of each element, but also perform experiments to actually generate images with StyleGAN and experience its performance.
The figure below is an example of an experiment on the function “Style mixing” of StyleGAN. StyleGAN can combine the features of two images by switching latent variables in the middle, which also has the effect of suppressing over-learning by doing it during learning.
Through experiments, we can understand the changes in the image due to Style mixing of StyleGAN. In this course, we will also conduct various image generation experiments related to StyleGAN.
Next, we will extract 100 particularly important variants of GANs, and explain their main points. At that time, we will introduce application examples in the business scene and think about how to actually use GAN.
What do you think? This is a must-see course for those who want to learn more about the possibilities and cutting-edge knowledge of GANs, and for those who want to use GANs better than anyone else in the field. It is an interactive lecture where you can not only watch videos online, but actually move your hands and experience active learning. There is also a free trial that allows you to watch part of the course, so if you are interested, please consider taking it.