Paper link: https://arxiv.org/abs/1406.2661v1
Preface
Generative Adversarial Nets (GAN) was proposed by Ian Goodfellow, Yoshua Bengio and others in 2014. Until now, it is still the most popular generation task algorithm. After GAN was proposed, a large number of excellent works based on GAN continued.
GANs can learn generative tasks in semi-supervised or unsupervised application scenarios. At present, GANs have achieved amazing results in the fields of computer vision and natural language processing. They are widely used in image generation, video generation, text generation and other fields. They can generate samples that are fake and real, and are highly creative. and potential commercial value. GANs are one of the most promising methods for unsupervised learning on complex data distributions in recent years.
However, the training process of GANs is relatively complex and is very sensitive to hyperparameters and network design, requiring careful adjustment and optimization to obtain good generation results. At the same time, the generated samples may have some uncontrollable biases and uncertainties. Therefore, when using GANs for generation tasks, you also need to carefully consider the characteristics of the application scenario and model.
Generative and discriminative methods:
Generative and discriminative methods are two different modeling methods in machine learning. The difference between the two lies in their learning goals and modeling methods:
1) Observation goals: Generative methods try to understand the real distribution behind the sample data and learn how to generate similar data samples; discriminative methods focus on input data and output Conditional probability distributions between labels, i.e., learning decision boundaries, focus on how to map input data to corresponding output labels.
2) How to use: Generative methods can be used to generate new data samples, fill in missing values, perform data enhancement and other tasks. Common generative methods include probabilistic graphical models, generative adversarial networks (GAN), and variational autoencoders. (VAE), etc.; discriminant methods are mainly used for tasks such as classification, regression and prediction.
3) Data requirements: Generative methods have higher requirements on the distribution of sample data and require more training data to capture the statistical characteristics of the data; discriminative methods pay more attention to the relationship between input data and output labels. For training Data distribution requirements are relatively low.