The Institute of Information Technology of ANAS held an online scientific seminar on "Analysis of learning algorithms in generative rival networks" of the 2nd department.
Presenting the report, Firangiz Sadiyeva, an employee of the institute, spoke about the Generative Adversial Network (GAN) and the problems encountered in their teaching. She noted that generative rival networks have three main problems: disruption of the regime, unsustainable learning and lack of quality metrics.
Ms. Sadiyeva spoke about a number of variants of generative rival networks (VanillaGAN, WassersteinGAN, DCGAN, InfoGAN, conventional GAN, f-GAN and GAN Laplace pyramid) and the purpose functions used in them.
She noted that generative rival networks have shown themselves better in the generation of different images. She said that in mode colllapse, the network begins to generate a small number of images, not all of the intended object images. The researcher also presented the main characteristics of the proposed AdaGAN, VEEGAN and Wasserstein GAN architectures to solve this problem.
The speaker then highlighted the applications of generative rival networks, solutions to improve learning, data sets used in GAN training and testing, and GAN quality assessment metrics. Although about 30 metrics are suggested in the literature, Google Inseption noted that only two metrics based on the deep neural network - Inseption Score and Freche Inseption - gave acceptable results in image generation. She noted that these metrics also place serious restrictions on the format and size of images.
The head of the department, PhD in Technical Sciences, Associate Professor Yadigar Imamverdiyev recommended to deepen research on the application of generative rival networks in the field of information security, to explore new directions on the subject, and made a number of proposals.
Discussions were held around the report, questions were answered.