MMD
A measure of the discrepancy between two probability distributions that maps distributions into a reproducing kernel Hilbert space (RKHS) via a kernel function and quantifies the difference using the distance between their mean embeddings. It is widely used in GAN training, two-sample hypothesis testing, and domain adaptation.
Timeline (last 90 days)
Traditional offline RL methods mitigate distribution shift by imposing divergence constraints on the action distribution (such as KL divergence, Wasserstein distance, MMD, etc.).
Exactly achieving population independence under finite samples is almost impossible; in practice, batch-level statistics (such as HSIC, MMD, or correlation penalties) are usually used for approximate constraints, and the paper's theoretical guarantees are also discussed within this approximate framework.
Under weight-free gradient flow, any stationary point whose MMD value is far from zero is unstable, and gradient flow will not stably remain at these bad stationary points.
MMD (maximum mean discrepancy) is a classic measure in machine learning for measuring the difference between two distributions, widely used in generative models, domain adaptation, and kernel two-sample tests.
If and only if the kernel is a characteristic kernel (such as a Gaussian kernel), MMD(P,Q)=0 is equivalent to P=Q.
The contractivity under kernel-induced geometry enters the convergence bound only through a linear correction term, rather than the dominant term.
When the kernel is a characteristic kernel (e.g., a Gaussian kernel), MMD equals zero if and only if the two distributions are identical.
MMD is widely applied in tasks such as the training objective of Generative Adversarial Networks (GANs), two-sample hypothesis testing, and domain adaptation.
The paper's experiments cover smooth soft-particle energy, maximum mean discrepancy (MMD) minimization, component ablation experiments, and scaling studies in both the number of particles and dimension directions.
Common UDA technical approaches include adversarial training (e.g., the DANN method), Maximum Mean Discrepancy (MMD), and self-training/pseudo-labeling
1 more timeline events
All Facts (11)
Traditional offline RL methods mitigate distribution shift by imposing divergence constraints on the action distribution (such as KL divergence, Wasserstein distance, MMD, etc.).
50%UnverifiedExactly achieving population independence under finite samples is almost impossible; in practice, batch-level statistics (such as HSIC, MMD, or correlation penalties) are usually used for approximate constraints, and the paper's theoretical guarantees are also discussed within this approximate framework.
50%UnverifiedUnder weight-free gradient flow, any stationary point whose MMD value is far from zero is unstable, and gradient flow will not stably remain at these bad stationary points.
50%UnverifiedMMD (maximum mean discrepancy) is a classic measure in machine learning for measuring the difference between two distributions, widely used in generative models, domain adaptation, and kernel two-sample tests.
50%UnverifiedIf and only if the kernel is a characteristic kernel (such as a Gaussian kernel), MMD(P,Q)=0 is equivalent to P=Q.
50%UnverifiedThe contractivity under kernel-induced geometry enters the convergence bound only through a linear correction term, rather than the dominant term.
50%UnverifiedWhen the kernel is a characteristic kernel (e.g., a Gaussian kernel), MMD equals zero if and only if the two distributions are identical.
50%UnverifiedMMD is widely applied in tasks such as the training objective of Generative Adversarial Networks (GANs), two-sample hypothesis testing, and domain adaptation.
50%UnverifiedThe paper's experiments cover smooth soft-particle energy, maximum mean discrepancy (MMD) minimization, component ablation experiments, and scaling studies in both the number of particles and dimension directions.
50%UnverifiedCommon UDA technical approaches include adversarial training (e.g., the DANN method), Maximum Mean Discrepancy (MMD), and self-training/pseudo-labeling
50%UnverifiedDomain adaptation assumes access to unlabeled data from the target domain and aligns the feature distributions of the source and target domains through adversarial training (such as the DANN gradient reversal layer), MMD minimization, or optimal transport methods.
50%