Wasserstein Distance
A distance metric from probability theory and optimal transport theory used to measure the discrepancy between two probability distributions. It is sensitive to the geometric structure of distributions and is widely used in the theoretical analysis of generative models.
Timeline (last 90 days)
In risk-averse machine learning, besides CVaR, there also exist methods such as variance penalties, Wasserstein distance constraints and distributionally robust optimization (DRO).
Drift detection algorithms generally monitor changes through statistical tests (such as ADWIN and the Page-Hinkley test) or distribution distance metrics (such as KL divergence and Wasserstein distance).
Wasserstein distance bypasses the label permutation problem by representing clustering partitions as probability measures and solving an optimal transport problem, measuring the minimum total cost required to transport one partition into another.
The study introduces a Wasserstein-type posterior summarization method that compresses a large collection of posterior samples into a small number of weighted representative clusterings using a newly designed noise-aware loss function.
The Wasserstein distance measures the minimum cost required to transport one distribution into another, where the magnitude of the cost is determined by the underlying cost function.
The 2-Wasserstein distance is defined as W₂(μ,ν)² being the infimum of the optimal transport cost over all couplings with marginals μ and ν
The error bound in this paper holds for all Wasserstein-p distances with p ≥ 1, rather than being limited to a specific metric.
The paper proves that the learned generative distribution satisfies a finite-sample error bound in Wasserstein-p distance: W_p ≲ n^(-1/d) + n^(-1/(2p))(log(1/ξ))^(1/(2p)).
Data drift detection originated as a concept in the machine learning field; detection methods include statistical tests (such as the Kolmogorov-Smirnov test and chi-squared test) and distribution distance measures (such as KL divergence and Wasserstein distance).
The Wasserstein distance can provide meaningful gradient signals even when two distributions have no overlap at all, alleviating the mode collapse and vanishing gradient problems in original GAN training.
4 more timeline events
All Facts (14)
In risk-averse machine learning, besides CVaR, there also exist methods such as variance penalties, Wasserstein distance constraints and distributionally robust optimization (DRO).
50%UnverifiedDrift detection algorithms generally monitor changes through statistical tests (such as ADWIN and the Page-Hinkley test) or distribution distance metrics (such as KL divergence and Wasserstein distance).
50%UnverifiedWasserstein distance bypasses the label permutation problem by representing clustering partitions as probability measures and solving an optimal transport problem, measuring the minimum total cost required to transport one partition into another.
50%UnverifiedThe study introduces a Wasserstein-type posterior summarization method that compresses a large collection of posterior samples into a small number of weighted representative clusterings using a newly designed noise-aware loss function.
50%UnverifiedThe Wasserstein distance measures the minimum cost required to transport one distribution into another, where the magnitude of the cost is determined by the underlying cost function.
50%UnverifiedThe 2-Wasserstein distance is defined as W₂(μ,ν)² being the infimum of the optimal transport cost over all couplings with marginals μ and ν
50%UnverifiedThe error bound in this paper holds for all Wasserstein-p distances with p ≥ 1, rather than being limited to a specific metric.
50%UnverifiedThe paper proves that the learned generative distribution satisfies a finite-sample error bound in Wasserstein-p distance: W_p ≲ n^(-1/d) + n^(-1/(2p))(log(1/ξ))^(1/(2p)).
50%UnverifiedData drift detection originated as a concept in the machine learning field; detection methods include statistical tests (such as the Kolmogorov-Smirnov test and chi-squared test) and distribution distance measures (such as KL divergence and Wasserstein distance).
50%UnverifiedThe Wasserstein distance can provide meaningful gradient signals even when two distributions have no overlap at all, alleviating the mode collapse and vanishing gradient problems in original GAN training.
50%Unverified一维Wasserstein-1距离等于两个累积分布函数之间面积的积分 W₁ = ∫|F_A(x) - F_B(x)|dx
50%Unverified切片Wasserstein距离（SWD）将高维分布随机投影到多个一维方向上，在每个方向上计算Wasserstein距离后取平均
50%UnverifiedWasserstein距离是一个真正的度量，满足对称性和三角不等式，而KL散度不对称且在分布支撑不重叠时为无穷大
50%Unverified一维情况下Wasserstein距离有闭式解，只需排序即可，复杂度为O(n log n)
50%