Explicit Information Placement on Latent Variables using Auxiliary Generative Modelling Task

Deep latent variable models, such as variational autoencoders, have been successfully used to disentangle factors of variation in image datasets. The structure of the representations learned by such models is usually observed after training and iteratively refined by tuning the network architecture and loss function. Here we propose a method that can explicitly place information into a specific subset of the latent variables. We demonstrate the use of the method in a task of disentangling global structure from local features in images. One subset of the latent variables is encouraged to represent local features through an auxiliary modelling task. In this auxiliary task, the global structure of an image is destroyed by dividing it into pixel patches which are then randomly shuffled. The full set of latent variables is trained to model the original data, obliging the remainder of the latent representation to model the global structure. We demonstrate that this approach successfully disentangles the latent variables for global structure from local structure by observing the generative samples of SVHN and CIFAR10. We also clustering the disentangled global structure of SVHN and found that the emerging clusters represent meaningful groups of global structures – including digit identities and the number of digits presence. Finally, we discuss the problem of evaluating the clustering accuracy when ground truth categories are not expressive enough.