Feb 15, 2018 (modified: Mar 27, 2018)ICLR 2018 Conference Blind Submissionreaders: everyoneShow Bibtex
Abstract:What would be learned by variational autoencoder(VAE) and what influence the disentanglement of VAE? This paper tries to preliminarily address VAE's intrinsic dimension, real factor, disentanglement and indicator issues theoretically in the idealistic situation and implementation issue practically through noise modeling perspective in the realistic case. On intrinsic dimension issue, due to information conservation, the idealistic VAE learns and only learns intrinsic factor dimension. Besides, suggested by mutual information separation property, the constraint induced by Gaussian prior to the VAE objective encourages the information sparsity in dimension. On disentanglement issue, subsequently, inspired by information conservation theorem the clarification on disentanglement in this paper is made. On real factor issue, due to factor equivalence, the idealistic VAE possibly learns any factor set in the equivalence class. On indicator issue, the behavior of current disentanglement metric is discussed, and several performance indicators regarding the disentanglement and generating influence are subsequently raised to evaluate the performance of VAE model and to supervise the used factors. On implementation issue, the experiments under noise modeling and constraints empirically testify the theoretical analysis and also show their own characteristic in pursuing disentanglement.
TL;DR:This paper tries to preliminarily address the disentanglement theoretically in the idealistic situation and practically through noise modelling perspective in the realistic case.