The loss error battle between the discriminator and the generator when generating Type IIIs. This illustrates the GAN’s learning pattern. Notice how no instance of training is the same. During training, the generator and discriminator losses should converge, which is an indication that the network is producing high-quality simulated images. This can be observed in plots (a) and (c). One key feature when training GANs is convergence failure seen in plots (b) and (d) (when generator loss spikes for a period of epochs; Goodfellow et al. 2014). This occurs when there is an inability to find the equilibrium between generator loss and discriminator loss. Images generated during this period are very poor and noisy.