Maximum Mean Discrepancy (MMD) is a non-parametric statistical test and distance metric that measures the difference between two probability distributions by comparing their mean embeddings in a reproducing kernel Hilbert space (RKHS). In domain adaptation, MMD serves as a differentiable loss function that quantifies how different the source and target feature distributions are, enabling direct minimization of domain discrepancy without adversarial training.

Why MMD Matters in AI/ML: MMD provides a statistically principled, non-adversarial measure of distribution distance that is differentiable, easy to compute, has well-understood theoretical properties, and directly plugs into neural network training as a regularization loss—making it the most mathematically grounded approach to domain alignment.

• RKHS embedding — Each distribution P is represented by its mean embedding μ_P = E_{x~P}[φ(x)] in a RKHS defined by kernel k; MMD²(P,Q) = ||μ_P - μ_Q||²_H = E[k(x,x')] - 2E[k(x,y)] + E[k(y,y')], where x,x' ~ P and y,y' ~ Q • Kernel choice — The Gaussian RBF kernel k(x,y) = exp(-||x-y||²/2σ²) is most common; multi-kernel MMD uses a mixture of Gaussians with different bandwidths for robustness; the kernel must be characteristic (Gaussian, Laplacian) to guarantee that MMD=0 iff P=Q • Unbiased estimator — Given source samples {x_i}ᵢ₌₁ᴺ and target samples {y_j}ⱼ₌₁ᴹ, the unbiased empirical MMD² = 1/(N(N-1))Σᵢ≠ⱼk(xᵢ,xⱼ) - 2/(NM)ΣᵢΣⱼk(xᵢ,yⱼ) + 1/(M(M-1))Σᵢ≠ⱼk(yᵢ,yⱼ) is computed from mini-batches during training • Multi-layer MMD (DAN) — Deep Adaptation Network (DAN) minimizes MMD across multiple hidden layers simultaneously: L = L_task + λΣₗ MMD²(S_l, T_l), aligning representations at multiple abstraction levels for more robust adaptation • Conditional MMD — Class-conditional MMD aligns source and target distributions per class: Σ_k MMD²(P_S(f|y=k), P_T(f|y=k)), preventing class confusion that can occur with marginal MMD alignment alone

Maximum Mean Discrepancy is the mathematically rigorous foundation for non-adversarial domain adaptation, providing a differentiable distribution distance in kernel space that enables direct minimization of domain discrepancy, with well-understood statistical properties, unbiased estimation from finite samples, and seamless integration as a regularization loss in deep neural network training.