Correcting Flaws in Common Disentanglement Metrics
This work addresses the problem of accurately evaluating disentangled representations for researchers in representation learning, though it is incremental as it builds on prior metrics.
The paper identifies flaws in existing disentanglement metrics that can incorrectly assign high scores to entangled models, and proposes two new metrics to correct these issues. It also introduces a classification-based approach for compositional generalization, showing that performance is poor, correlates with most metrics, and is most strongly correlated with the new metrics.
Recent years have seen growing interest in learning disentangled representations, in which distinct features, such as size or shape, are represented by distinct neurons. Quantifying the extent to which a given representation is disentangled is not straightforward; multiple metrics have been proposed. In this paper, we identify two failings of existing metrics, which mean they can assign a high score to a model which is still entangled, and we propose two new metrics, which redress these problems. We then consider the task of compositional generalization. Unlike prior works, we treat this as a classification problem, which allows us to use it to measure the disentanglement ability of the encoder, without depending on the decoder. We show that performance on this task is (a) generally quite poor, (b) correlated with most disentanglement metrics, and (c) most strongly correlated with our newly proposed metrics.