Ensembles of Generative Adversarial Networks for Disconnected Data
This addresses a fundamental limitation in generative modeling for computer vision datasets with disconnected sets, offering a practical solution for researchers and practitioners.
The paper tackles the problem of representing disconnected data distributions, such as images from different classes, with generative networks, proving that continuous networks cannot do so without error and showing that ensembles of GANs outperform single continuous or conditional GANs with fewer parameters.
Most current computer vision datasets are composed of disconnected sets, such as images from different classes. We prove that distributions of this type of data cannot be represented with a continuous generative network without error. They can be represented in two ways: With an ensemble of networks or with a single network with truncated latent space. We show that ensembles are more desirable than truncated distributions in practice. We construct a regularized optimization problem that establishes the relationship between a single continuous GAN, an ensemble of GANs, conditional GANs, and Gaussian Mixture GANs. This regularization can be computed efficiently, and we show empirically that our framework has a performance sweet spot which can be found with hyperparameter tuning. This ensemble framework allows better performance than a single continuous GAN or cGAN while maintaining fewer total parameters.