Testing and Learning on Distributions with Symmetric Noise Invariance
This work addresses the issue of noise interference in distribution analysis for researchers in machine learning and statistics, offering a targeted improvement.
The paper tackles the problem of irrelevant variability in distribution comparisons by proposing distances and features invariant to additive symmetric noise, enabling robust two-sample testing and learning algorithms.
Kernel embeddings of distributions and the Maximum Mean Discrepancy (MMD), the resulting distance between distributions, are useful tools for fully nonparametric two-sample testing and learning on distributions. However, it is rarely that all possible differences between samples are of interest -- discovered differences can be due to different types of measurement noise, data collection artefacts or other irrelevant sources of variability. We propose distances between distributions which encode invariance to additive symmetric noise, aimed at testing whether the assumed true underlying processes differ. Moreover, we construct invariant features of distributions, leading to learning algorithms robust to the impairment of the input distributions with symmetric additive noise.