Sticky Reasoning within Learning Representations
Abstract
We first proposed an architecture and learning procedure that automatically learns representations of images in which the content is separated from the pose. The method is completely unsupervised, not requiring that the pose or the identity of the object in the image be provided [Mathieu NIPS 2016].While this work used generative adversarial networks (GAN), we found the concept wanting and proposed a new formulation of GAN scalled Energy-Based GAN. Our method was the first to allow the generation of realistic images at 128x128 resolution. Generative models, such as GANs are a key component of unsupervised procedures that can generate complex patterns from latent variables drawn from a simple distribution [Zhao ICLR 2017] (over 550 citations to date).
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 18, 2019
- Accession Number
- AD1096472
Entities
People
- Yann Le Cun
Organizations
- New York University