Generative Adversarial Networks (GAN) and GAN-inspired Innovations in Computer Vision
"The most interesting idea in the last 10 years in ML.” - Yann LeCun, Facebook AI research director.
In this talk, we will focus on Generative Adversarial Networks, one of the most interesting concepts in deep learning. A GAN is a generative model, which captures the patterns in the data so that it can generate new data points from the estimated data distribution. In the recent years, there has been tremendous research in the field of GANs, some of which include text-to-image synthesis, photo realistic image generation from doodles and a lot more.
We will cover the working of GANs with implementation and some of these interesting applications in this talk.
Keywords : StackGAN , DCGAN, Generators, Autoencoders, VAE
Outline/Structure of the Talk
- Generative vs Discriminative Models
- Why GANs?
- Introduction to GAN
- How do GANs work?
- Generators and Discriminators
- Cost function and optimization
- GANs vs Autoencoders and VAE
- Recent applications/case studies of GANs
- Text-to-image synthesis
- Pose Guided Person Image Generation
- Nvidia’s GauGAN (Doodles into photo realistic images)
- Deep Fakes
We will also cover an implementation of DCGAN using Jupyter notebook and keras for better understanding of the implementation and the concept.
By the end of this session, the audience will have a clear idea about what Generative Adversarial Networks are, how they work and the latest innovations that have taken place using GANs.
Anyone who is interested in latest developments in deep learning and wants to understand GANs and their applications.
Prerequisites for Attendees
Basic understanding of deep learning and how neural networks are trained. Beginner level knowledge about Python and Keras will be helpful in understanding the concepts more efficiently.