A framework for the quantitative evaluation of disentangled representations


Nov 03, 2017 (modified: Nov 03, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Recent AI research has emphasised the importance of learning disentangled representations of the explanatory factors behind data. Despite the recent focus on models which can learn such representations, visual inspection remains the primary method for evaluating the degree of disentanglement achieved. While various desiderata have been implied in recent definitions, it is currently unclear what exactly makes one disentangled representation better than another. In this work we propose a framework for quantitatively evaluating the quality of disentangled representations learned by different models. Three criteria are explicitly defined and quantified to elucidate the quality of learnt representations and compare models on an equal basis. Experiments with the recent InfoGAN model for learning disentangled representations illustrate the appropriateness of the framework and provide a baseline for future work.