Revisiting precision recall definition for generative modeling
[edit]
Proceedings of the 36th International Conference on Machine Learning, PMLR 97:57995808, 2019.
Abstract
In this article we revisit the definition of PrecisionRecall (PR) curves for generative models proposed by (Sajjadi et al., 2018). Rather than providing a scalar for generative quality, PR curves distinguish modecollapse (poor recall) and bad quality (poor precision). We first generalize their formulation to arbitrary measures hence removing any restriction to finite support. We also expose a bridge between PR curves and type I and type II error (a.k.a. false detection and rejection) rates of likelihood ratio classifiers on the task of discriminating between samples of the two distributions. Building upon this new perspective, we propose a novel algorithm to approximate precisionrecall curves, that shares some interesting methodological properties with the hypothesis testing technique from (LopezPaz & Oquab, 2017). We demonstrate the interest of the proposed formulation over the original approach on controlled multimodal datasets.
Related Material


