Generative Adversarial Networks, An Introduction

While deep learning has made historic improvements in speech recognition and object recognition in recent years, almost all of these gains have been in supervised learning of now fairly well understood discriminative models. In the larger context of machine learning, less is understood about both unsupervised and generative models, but Generative Adversarial Networks have emerged as a promising approach to making progress in that direction.

We are going to introduce Generative Adversarial Networks (GAN), a deep learning generative model. GANs have primarily been used to generate samples of realistic images but other recent uses have included generating song lyrics, images from captions and video. We will begin with a gentle background of the theory of generative models and GANs in particular and show how GANs are being used today. We will then step through the code for training a basic GAN and we will show how to use a pre-trained GAN to generate images. The objective of this talk is to provide a basic introduction to generative models and Generative Adversarial Networks such that you can walk away from this talk with enough understanding to train and test your own GAN.

This talk is sponsored by the Data Science DC meetup, which meets monthly to discuss diverse topics in predictive analytics, applied machine learning, statistical modeling, open data, and data visualization. Its members are professionals, students, and others with a deep interest in these fields and related technologies. Meeting topics are varied and range from tutorials on basic concepts and their applications, to success stories from local practitioners, to discussions of tools, new technologies, and best practices.

For more information and to register, see the Data Science DC Web site.

The presentation and Jupyter notebook for the talk are available here.