Generative Auto-Encoder: Non-adversarial Controllable Synthesis with Disentangled Exploration

Autoencoders perform a powerful information compression framework with are construction loss and can be a regularization module in different tasks, which has no generative ability itself. We wondering if an autoencoder gains generative ability without using GAN and VAE based modification, which are two mature methods. Here we propose a new method: Disentanglement and ExplorationAutoencoder (DEAE), DEAE using a disentangle and exploration positive iteration achieve semantic controllable synthesis especially controllable mining the novel semantic value. For instance, given only red, green, and blue object color while mining new object color expressions in the image domain. The encoder of DEAE first turn the input sample into a disentangled latent code, then explore the latent codespace by attribute oriented interpolation. To encourage interpolated latent code successfully output a semantically meaningful sample by the decoder, we propose a regularization procedure by ’reuse’ encoder and constrain the output latent value which implicitly improves the quality of the interpolated sample. DEAE can become a generative model and synthesis semantic controllable samples by interpolating latent code, which can even synthesis novel attribute value never is shown in the original dataset. Experiments demonstrate how disentanglement and exploration can boost each other which empowers autoencoder generative ability. We also demonstrate that DEAE can improve the performance of downstream tasks compared with GANand VAE based generative model, especially in controllable data augmentation, dataset bias elimination (Fairness)