English

Separation Results between Fixed-Kernel and Feature-Learning Probability Metrics

Machine Learning 2021-11-02 v4 Machine Learning Probability Statistics Theory Statistics Theory

Abstract

Several works in implicit and explicit generative modeling empirically observed that feature-learning discriminators outperform fixed-kernel discriminators in terms of the sample quality of the models. We provide separation results between probability metrics with fixed-kernel and feature-learning discriminators using the function classes F2\mathcal{F}_2 and F1\mathcal{F}_1 respectively, which were developed to study overparametrized two-layer neural networks. In particular, we construct pairs of distributions over hyper-spheres that can not be discriminated by fixed kernel (F2)(\mathcal{F}_2) integral probability metric (IPM) and Stein discrepancy (SD) in high dimensions, but that can be discriminated by their feature learning (F1\mathcal{F}_1) counterparts. To further study the separation we provide links between the F1\mathcal{F}_1 and F2\mathcal{F}_2 IPMs with sliced Wasserstein distances. Our work suggests that fixed-kernel discriminators perform worse than their feature learning counterparts because their corresponding metrics are weaker.

Keywords

Cite

@article{arxiv.2106.05739,
  title  = {Separation Results between Fixed-Kernel and Feature-Learning Probability Metrics},
  author = {Carles Domingo-Enrich and Youssef Mroueh},
  journal= {arXiv preprint arXiv:2106.05739},
  year   = {2021}
}
R2 v1 2026-06-24T03:03:28.313Z