Hands-On Generative Adversarial Networks with Keras

Adding noise

It is reported that adding noise to the training procedure can improve training because it makes the support of the real and fake distributions less separable, thus potentially eliminating vanishing gradients and training instabilities. Take a look at the following list of some of these strategies:

Source: Amortised MAP inference for Image Super-Resolution (https://arxiv.org/abs/1610.04490)

• Flip labels occasionally: By flipping labels occasionally, real becomes fake and fake becomes real. The preceding diagram shows the effect on a multivariate-Gaussian distribution with two modes. Note that flipping the labels does not create ...