Domain adaptation enables models to generalize across different data distributions. This paper proposes that "erratic discrimination of the target domain mainly stems from its much smaller feature norms" compared to source domain features. The authors introduce Adaptive Feature Norm (AFN), a parameter-free approach that progressively adapts feature norms across domains toward larger values. The method unifies standard and partial domain adaptation while addressing negative transfer risks. Results show substantial improvements: "11.5% on Office-Home and 17.1% on VisDA2017" over prior work.

The paper identifies a key observation through visualization: target samples "collide in the small-norm (i.e., low-radius) regions which are vulnerable to slight angular variations of decision boundaries." Two hypotheses are proposed: the Misaligned-Feature-Norm Hypothesis, suggesting domain shift depends on misaligned mean norms, and the Smaller-Feature-Norm Hypothesis, proposing that domain shift relies on less-informative smaller-norm features. The authors argue that adapting targets away from small-norm regions enables safe transfer.

Existing domain adaptation methods minimize statistical discrepancies like Maximum Mean Discrepancy (MMD) and correlation distance. Adversarial domain adaptation dominates recent research, with methods including RevGrad, ADDA, and MCD. Partial domain adaptation addresses scenarios where "source label space subsumes the target one," with PADA attempting to mitigate negative transfer through sample reweighting. These approaches all focus on distribution alignment but overlook the role of feature magnitude.

The framework includes a backbone network G and classifier F with l fully-connected layers. The authors first introduce L2-preserved Dropout, a modification that scales outputs by 1/sqrt(1-p) instead of 1/(1-p), ensuring the expected squared values remain invariant across training and evaluation phases. This addresses the systematic feature norm discrepancy introduced by standard dropout during domain adaptation.

Hard AFN (HAFN) constrains both domains' mean norms toward a shared equilibrium R. The optimization objective combines source classification loss L_y for task-discriminative features and feature-norm penalty L_d for domain-transferable features. However, HAFN has limitations: setting extremely large R causes gradient explosion, preventing exploration of whether "rigorous alignment between feature-norm expectations is necessary."

SAFN encourages progressive feature-norm enlargement via individual sample penalties, using a residual scalar delta_r to control step size. Rather than forcing all features toward a single target R, SAFN incrementally increases each sample's feature norm by delta_r. This approach proves more stable than HAFN and permits larger effective feature norms without gradient issues. The key insight: "once we properly adapt the target samples towards the large-norm regions, the rigorous alignment becomes superfluous."

Experiments span VisDA2017 (synthetic-to-real, 280K+ images, 12 categories), Office-Home (4 domains, 65 categories), Office-31 (3 domains, 31 categories), and ImageCLEF-DA (3 datasets, 12 categories). In the vanilla setting, SAFN achieves 68.5% average accuracy on Office-Home versus prior SOTA 63.8%, and 76.1% per-class accuracy on VisDA2017. In the partial setting, results show 11.5% improvement on Office-Home and 17.1% on VisDA2017 over the PADA baseline. t-SNE visualization confirms that target samples separate and align with source clusters post-adaptation. Ablation studies verify that entropy minimization provides complementary gains of 0.8-1.4%.

The paper reveals that "model degradation on the target domain mainly stems from its much smaller feature norms" and demonstrates that progressive norm adaptation toward larger values yields significant transfer gains. The approach is parameter-free, easy to implement, and robustly handles both standard and partial domain adaptation scenarios while mitigating negative transfer effects. The simplicity of the method suggests that feature magnitude is an underexplored yet fundamental factor in domain adaptation.