PCA-AE: Principal Component Analysis Autoencoder for Organising the Latent Space of Generative Networks

Autoencoders and generative models produce some of the most spectacular deep learning results to date. However, understanding controlling latent space these presents a considerable challenge. Drawing inspiration from principal component analysis autoencoders, we propose autoencoder (PCA-AE). This is novel whose verifies two properties. Firstly, dimensions are organised in decreasing importance with respect data at hand. Secondly, components statistically independent. We achieve this by progressively increasing during training, covariance loss applied codes. The resulting produces which separates intrinsic attributes into different space, completely unsupervised manner. also describe an extension our approach case powerful, pre-trained GANs. show on both synthetic examples shapes state-of-the-art GAN. For example, able separate colour shade scale hair, pose faces gender, without accessing any labels. compare PCA-AE other approaches, particular ability disentangle space. hope that will contribute better spaces powerful models.