IC Colloquium: Progressive growing of GANs
Cancelled
Event details
Date | 17.12.2018 |
Hour | 16:15 › 17:30 |
Location | |
Category | Conferences - Seminars |
By: Jaakko Lehtinen - Aalto University
Abstract:
Generative adversarial networks (GAN) are a powerful and exciting family of generative models that learn purely from observing samples. The quality of samples has, however, remained less than ideal. We describe a new GAN training methodology that yields samples of unprecedented quality. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. A form of curriculum learning, this both speeds the training up and greatly stabilizes it. We also propose a simple way to increase the variation in generated images. This talk will also feature still higher-quality results than presented in the ICLR article.
Bio:
Jaakko Lehtinen is an associate professor at Aalto University, Finland, and a principal research scientist with NVIDIA Research. Prior to that, he spent a few years as a postdoc with Frédo Durand at MIT. Jaakko works in the intersection of computer graphics and computer vision, including imaging and generative modelling. Of the latter line of research, Progressively-grown GANs* have recently gathered significant attention.
More information
Abstract:
Generative adversarial networks (GAN) are a powerful and exciting family of generative models that learn purely from observing samples. The quality of samples has, however, remained less than ideal. We describe a new GAN training methodology that yields samples of unprecedented quality. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. A form of curriculum learning, this both speeds the training up and greatly stabilizes it. We also propose a simple way to increase the variation in generated images. This talk will also feature still higher-quality results than presented in the ICLR article.
Bio:
Jaakko Lehtinen is an associate professor at Aalto University, Finland, and a principal research scientist with NVIDIA Research. Prior to that, he spent a few years as a postdoc with Frédo Durand at MIT. Jaakko works in the intersection of computer graphics and computer vision, including imaging and generative modelling. Of the latter line of research, Progressively-grown GANs* have recently gathered significant attention.
More information
Practical information
- General public
- Free
- This event is internal
Contact
- Host: Wenzel Jakob