A new framework could solve mode collapse in generative adversarial networks

Generative adversarial networks (GANs) are widely used to synthesize complex and realistic data by learning the distribution of genuine real-world samples. However, a significant challenge faced by GANs is mode collapse, where the diversity of the generated samples is significantly lower than the diversity of the actual samples. The complexity of GANs and their training process makes it difficult to uncover the underlying mechanisms of mode collapse.

A research team led by Professor Yang Zhouwang of the University of Science and Technology of China (USTC) of the Chinese Academy of Sciences (CAS) thoroughly investigated the root causes of mode collapse and proposed a new framework, Dynamic GAN (DynGAN), Quantitatively detect and resolve mode collapse in GANs. Their works are IEEE Transactions on Pattern Analysis and Machine Intelligence.

Through theoretical analysis, the team found that when multiple modes exist in the real data, the generator's loss function is non-convex with respect to its parameters. Specifically, the parameters that result in a generated distribution that covers only a portion of the modes of the real distribution are the local minima of the generator's loss function.

To tackle the mode collapse problem, the team proposed a unified framework, DynGAN. This framework can set thresholds on the output of observable discriminators to detect samples that the generator could not produce (called decay samples). The training set is partitioned based on the collapsed samples, and a dynamic conditional generative model is trained on the partitions.

The theoretical results ensure that the progressive mode of DynGAN is covered. Experimental results on both synthetic and real-world datasets show that DynGAN outperforms existing GANs and their variants in resolving address mode collapse.

This work not only advanced the theoretical understanding of GANs but also provided important implementation strategies to improve the mode coverage of generative models.