A method for improving a generative adversarial network by using adaptive control learning

Abstract

The invention belongs to the technical field of an adversarial network, and discloses a method for improving a GAN (Generative Adversarial Network) by using adaptive control learning, which is suitable for an adaptive hyper-parameter learning process of the GAN to improve the training stability of different data sets so as to ensure the quality of generated data (such as images and characters). The method is realized by guiding a training process of a data set with relatively complex categories and modes through a learning curve of training elements obtained under the data set with relativelysimple categories and modes; The present invention also analyzes an adaptive generative adversarial network (Ak-GAN) model having a multi-layer perceptron (MLP) and a deep convolutional neural network(DCGAN) architecture. According to the method, the stability of general GAN training can be improved indeed, and the method can be well popularized to various improved models and data sets; For future work, plan analysis better generates a sample metric, which may encourage convergence of the GAN; The effect of the proposed adaptive control mechanism in GAN multi-modal learning is hoped to be analyzed.

Description

technical field [0001] The invention belongs to the technical field of confrontation networks, and in particular relates to a method for improving generation confrontation networks with adaptive control learning. Background technique [0002] Generative Adversarial Networks (GANs) can efficiently synthesize samples for various applications, such as image generation, industrial design, speech synthesis, and natural language processing. The goal of GAN is to alternately train the generator model G and the discriminator model D. G and D of GAN usually choose to use multi-layer perceptron (MLP) or convolutional neural network (CNN). The generator accepts a set of noise priors to simulate the real data distribution, and the simulated distribution is called the generated data distribution, while the discriminator is trained to extract the discriminative features of the real data. More specifically, the discriminator judges the probability that a piece of data is from the real da...

AI Technical Summary

Problems solved by technology

Two major issues complicate the problem are: 1) the imbalance in the capabilities of the generator and discriminator models during training, which prevents the generator from effectively learning the real data distribution; 2) the lack of computable, interpretable convergence criteria

[0009] In the prior art, two main issues complicating the problem are: the uneven growth of the model capabilities of the generator and the discriminator during training, which hinders the effective learning of the generator; the lack of computable, interpretable convergence criteria

Images