The training efficiency of the StyleGAN2 neural network for rendering has been improved fourfold. This was reported by the press service of the National Research University Higher School of Economics.
Modern neural networks can produce fake images that are almost indistinguishable from real ones. It is possible to create faces, especially of people who have never lived. One of the most successful types of neural networks for these tasks are generative adversarial networks, in which one algorithm builds an image and the other tries to separate it from the real one. As a result, with a gradual enumeration, the picture takes such a form that the differences are minimal. At the same time, a significant difficulty in operating such a scheme is the need to collect a large number of high-quality images for training. For example, correct generation of random faces requires a database of at least 100,000 real photos. However, there are ways to partially circumvent this limitation: for example, if there are very few required images (photos of a certain style or people with a certain face), then you can train the network on ordinary images and then “train” it by changing tens of millions of parameters.
Experts from the HSE Center for Deep Learning and Bayesian Methods have described a new approach to retrain the StyleGAN2 generative model. This is a generative neural network that transforms random noise into a realistic picture. By training an additional domain vector, the researchers were able to optimize its training by reducing the number of parameters (weights) trained by four orders of magnitude.
The StyleGAN2 network architecture means that the input random vector can be changed by gender, age, etc. It has special transformations (modulations) where it controls the semantic properties of the output image. The scientists proposed to grow an additional vector that describes the output image field through similar modulations.
“In addition, if we only train such a domain vector, the field of the generated images changes, just as if we were retraining all the parameters of the neural network. This greatly reduces the number of optimized parameters, since the size of such a domain vector is only 6000,” said Aibek Alanov, one of the creators of the new algorithm. ;
As a result, the authors hope that their invention will significantly speed up the training of generative neural networks and simplify their operation.
Ancient Russian scientists created Coating that protects against the destruction of magnesium-based implants.