event
Thursday 19 Mar 2020: [Cancelled] Image Attribute Editing Using Cyclic Reverse Generator
Dr. Hacer YALIM KELES - Ankara University, Turkey
Harrison 170 14:30-15:30
Generative Adversarial Networks (GANs) recently provide unreasonably realistic images via two competing networks: generator and discriminator. Generator usually maps a lower dimensional latent space from a known distribution, such as Normal or Uniform distribution, to a higher-dimensional data space that matches the real data distribution. Hence, when a generator is provided with a point from the latent space, it produces an image from the learnt distribution. Although this is impressive, it has very limited practical uses if we don’t know how to generate and/or modify a particular image.
What if we want to generate images in a more controlled fashion and edit some of its attributes as we like? Although there are some supervised solutions to this problem, which mainly impose some conditions in Generator training, the type of editing is limited by the predefined conditions, i.e. attribute labels. In this talk, I will present our unsupervised approach where explicit definitions of conditions are unnecessary during training. For the widely studied face image attribute editing problem, I will present our Cyclic Reverse Generator (CRG) architecture that helps embedding a given image in the latent space with high accuracy, and explain how we determine the attribute editing direction in the latent space.
Short Bio.
Dr. Hacer YALIM KELES received her B.S., M.S. and Ph.D. degrees in Computer Engineering from Middle East Technical University, Turkey, in 2002, 2005 and 2010, respectively. Her Ph.D. Thesis received the Thesis of the Year award by Middle East Technical University Prof. Dr. Mustafa Parlar Education and Research Foundation. From 2000 to 2007 she worked as a researcher and senior researcher at The Scientific and Technological Research Council of Turkey (TUBITAK). During the years at TUBITAK, she primarily worked on different pattern recognition problems using multimedia data including audio and video. In 2010, she received an R&D grant from Ministry of Industry and Trade of Turkey and established her R&D company. Her follow-up project SOYA is funded by TUBITAK in 2011 and later awarded by TUBITAK as one of the ten best venture projects in the country. She has been working as an Assistant Professor in the Department of Computer Engineering at Ankara University since 2013. Her research interests lie predominantly in the areas of computer vision and machine learning, particularly in deep learning. She is also interested in optimization of computational problems using GPUs. She is the Principle Investigator of a Nationally Funded Research Project on Turkish Sign Language Recognition Problem and a University funded project on Image Attribute Editing Using Generative Adversarial Networks. She is currently working as a visiting researcher in the Dept. of Computer Science at the University of Exeter.
