Channel: NeoScribe

Category: Science & Technology

Tags: generative adversarial networkdeep learningai newsganaiai created paintingai paintingai imaginationtechnologymachine learningartificial intelligencenvidia

Description: Hello, welcome to NeoScribe. Using our imagination is easy. We can all close our eyes, and think of ice cream, or cake, or even better, cake and ice cream. But teaching AI to imagine things has been very difficult up until a few years ago, with the advent of Dueling Neural Networks, also known as Generative Adversarial Networks or GANS. GANs is the next frontier in Machine Learning and it involves using two Neural Networks with opposing objectives to train one another, resulting in mind-blowing results. The first network is the generator which is programmed to generate images from random noise, with the goal to fool the other network called the discriminator, which is programmed to detect whether images are real or fake. First, the discriminator network is trained on a dataset, commonly referred to a training set. An example would be showing the network many images of faces until they are able to recognize images of faces from other images. Once the discriminator network is trained, its pitted against the generator like a 2-player game. The generator produces an image from random noise, trying to fool the discriminator as mentioned before. Then the image, along with a stream of real images is sent to the discriminator which predicts the probabilities of the images being real or fake. Whenever the generator creates an image that the discriminator identifies as fake, it learns from the failure and generates more realistic images. Likewise, when the discriminator incorrectly identifies a fake image as real, it learns from it and becomes more accurate at authenticating. This process creates a rapid feedback loop until the network reaches a point of equilibrium, successfully completing the training phase. So, after the training phase is complete, you have a generator network capable of creating realistic images based on what it learned. Here are some images generated from a GAN project done by Nvidia where they trained the worked with a huge dataset of 30,000 images of real celebrity faces. Researchers taught the discriminator network how to identify real faces of celebrities, while the generator was programmed to create images of faces with the goal that they will be authenticated. Then the networks dueled each other off of very simple 4x4 resolution images of a few skin-colored pixels. Then the researchers gradually trained the networks, layer by layer, expanding to 8x8 resolution, then 16x16 all the way up to the 9th of 1024x1024 resolution. This process took over 30 hours involving the authentication of over 10 MILLION images. And by the time the training was complete, the generator network was able to produce these incredible images of fake people. Look at that, these aren’t real people, they’re essentially imagined by artificial intelligence! Now I want to turn our attention to a group of artist and researchers out of Paris called Obvious. Obvious applied GANs algorithms to teach generator networks how to produce images of fine art portraits. The group fed the system with a dataset of 15,000 portraits painted between the 14th and 20th century. After the training process, Obvious had the generator network produce an 11-portrait series designed to resemble a fictional family called the Belamys. Look at these creepy images that came out of the network’s model. I like this one the best, the network really captured the sorrow of living in the 1600’s without electricity, Wi-Fi and smart phones. And someone must have really, really liked this one called The Portrait of Edmond de Belamy because it was recently sold at Auction for $432,000! While these applications of GANs are amazing, it appears that researchers are just scratching the surface. Researchers are starting to use GANs to teach AI how to form images from text. They are exploring the possibilities of using GANs to generate drugs for previously incurable diseases. Or using them to generate new anti-cancer molecules. GANs have only been around for about 5 years when they were introduced by Ian Goodfellow in 2014. Think about 10 years from NOW, imagine what networks can IMAGINE then? Alright, that’s all I have, for now, I hope you enjoyed your journey, if you did, please leave a like and subscribe. I am NeoScribe and I’ll see you on the next journey. Sources: (Not a complete list because description character limit) AI PORTRAIT: theverge.com/2018/10/25/18023266/ai-art-portrait-christies-obvious-sold EXPLAINER: towardsdatascience.com/generative-adversarial-networks-explained-34472718707a NVIDIA RESEARCH: research.nvidia.com/sites/default/files/pubs/2017-10_Progressive-Growing-of/karras2018iclr-paper.pdf FUTURE APPLICATIONS: blog.statsbot.co/generative-adversarial-networks-gans-engine-and-applications-f96291965b47 Tero Karras: youtube.com/watch?v=G06dEcZ-QTg&feature=youtu.be