|
|
The discriminator of feature learning N can learn the deep features of the data. These features can be used for other machine learning tasks such as classification, clustering, etc. Disadvantages of N The training process is complex The training process of N is more complicated and requires more hyperparameters to be adjusted. For example, the learning rate of the generator and discriminator, the choice of optimizer, the choice of noise distribution, etc. will all affect the training effect of N. Mode Collapse Problem N may suffer from mode collapse e lle problem where the generator always generates the same image.
This is because during the training process, the generator may find a "shortcut" that can fool the discriminator. It only generates a certain type of images and ignores other images. This makes the generated images lack diversity. Training process of training stability problem N The Rich People Phone Number List abilities of the generator and the discriminator need to be synchronized as much as possible. If the ability of the discriminator is too strong, the generator may not be able to find a suitable direction for optimization. On the contrary, if the ability of the generator is too strong, the discriminator may be deceived and cannot correctly guide the training
of the generator. . This instability makes the training process of N require very careful selection and adjustment of hyperparameters. Long training time Since N contains two neural networks and needs to be trained alternately, the training time of N is usually long. It is difficult to quantitatively evaluate the results generated by N Data quality is difficult to quantitatively assess. Although it can be assessed manually, this method is highly subjective and inefficient. Although there are some quantitative assessment methods such as Inein re, FI, etc., these methods have their own limitations. Generation
|
|
發表於 16:17:35