: 08h00, ngày 24/12/2020 (Thứ Năm)
: P104 D3, ĐH Bách Khoa Hà Nội
: Machine Learning và Data Mining
: Nguyễn Hữu Minh
: Viện Toán ứng dụng và Tin học, ĐH Bách Khoa Hà Nội
Tóm tắt báo cáo
GANs is a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that tries to capture the data distribution and a discriminative model D that estimates the probability that a sample came from the training data rather than G. Although the GANs frameworks have witnessed rapid advances, generating realistic images in high resolution is still difficult. Some obstacles are the memory constraint or the gradient problem. In recent years, some proposals have achieved incredible results and the most remarkable models are Progressive Growing GANs, StyleGAN and StyleGAN2. Some key ideas are to grow both the generator and discriminator progressively, to use style transfer literature and to regularize the generator. Moreover, in these works, the authors have shown some insight of the relation between the latent space and the image space.