In my presentation, I am going to show a wide spectrum of deep learning models that given a dataset are capable of generating artificial examples almost indistinguishable by humans from the ones coming from the original datasets. I'll show how useful the models are in many machine learning tasks and what we can expect in the future in this area of research.
Deep learning is particularly successful in computer vision field. Deep models have been achieving state-of-the-art results for a few years in many popular CV tasks like classification, object detection, and segmentation on many benchmark datasets. Besides that, deep learning models are capable of generating artificial samples of data indistinguishable by humans from real ones.
I am going to start my presentation by presenting a topic of image generation in its historical context. I will pay attention to the renaissance of deep learning in the 2000s. Later, I will introduce two most popular classes of models used nowadays - Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). I will discuss how they work and their process of learning. Various extensions and modifications will be also presented. Further, both models will be examined for their applications to semi-supervised learning. At the end, I would like to show some visualizations of data generated by previously discussed models and tell how they can be implemented using popular python libraries like Theano and Lasagne.