On the Transferability and Discriminability of Repersentation Learning in Unsupervised Domain Adaptation

TL;DR

The paper tackles the limitation of standard adversarial-based unsupervised domain adaptation that relies solely on distribution alignment and source-domain risk. It introduces an information-theoretic perspective, defining good representation learning as requiring both transferability and discriminability, and shows that targeting target-domain discriminability is necessary. Building on this, the authors propose RLGLC, an adversarial UDA framework that combines a novel AR-WWD-based global alignment with a Local Consistency Module (CNCE-based) to preserve target discriminability, supported by Bayes error-rate bounds linking information measures to generalization. Empirically, RLGLC achieves state-of-the-art performance across multiple benchmarks (Office-31, Office-Home, VisDa-2017, DomainNet, Digits) and tasks, and its ablations confirm the contribution of both global and local consistency components. The work thus bridges theory and practice in UDA, showing that enforcing both transferability and discriminability yields robust, transferable representations.

Abstract

In this paper, we addressed the limitation of relying solely on distribution alignment and source-domain empirical risk minimization in Unsupervised Domain Adaptation (UDA). Our information-theoretic analysis showed that this standard adversarial-based framework neglects the discriminability of target-domain features, leading to suboptimal performance. To bridge this theoretical-practical gap, we defined "good representation learning" as guaranteeing both transferability and discriminability, and proved that an additional loss term targeting target-domain discriminability is necessary. Building on these insights, we proposed a novel adversarial-based UDA framework that explicitly integrates a domain alignment objective with a discriminability-enhancing constraint. Instantiated as Domain-Invariant Representation Learning with Global and Local Consistency (RLGLC), our method leverages Asymmetrically-Relaxed Wasserstein of Wasserstein Distance (AR-WWD) to address class imbalance and semantic dimension weighting, and employs a local consistency mechanism to preserve fine-grained target-domain discriminative information. Extensive experiments across multiple benchmark datasets demonstrate that RLGLC consistently surpasses state-of-the-art methods, confirming the value of our theoretical perspective and underscoring the necessity of enforcing both transferability and discriminability in adversarial-based UDA.

Figures (6)

• Figure 1: The outline of the proposed RLGLC. RLGLC is an adversarial-based representation learning method for UDA. First, we feed the source-domain data $X_s$ and target-domain data $X_t$ from the original input space into the feature extractor $\varphi$, resulting in feature representations $Z_s$ and $Z_t$. Next, we pass $Z_s$ into the classifier $\psi$ to compute the cross-entropy loss on the source-domain samples. We then input both $Z_s$ and $Z_t$ into the network $f$ to calculate the proposed AR-WWD metric, ${W_{1,{W_{2,{d_\Omega }}}}}(P_s^\varphi, P_t^\varphi)$. Finally, we feed $Z_s$ and $Z_t$ into the network $\phi$ to compute the proposed conditional mutual information $I_{\text{CNCE}}(X_s; X_t \mid Z_t, \phi, K)$. For optimization, we begin by fixing $\varphi$ and $\psi$. We maximize ${W_{1,{W_{2,{d_\Omega }}}}}(P_s^\varphi, P_t^\varphi)$ and $I_{\text{CNCE}}(X_s; X_t \mid Z_t, \phi, K)$ to update $f$ and $\phi$, respectively. Next, we fix $f$ and $\phi$ and minimize Equation (\ref{['Eq:bbb']}) to update $\varphi$ and $\psi$. This two-step procedure ensures proper adversarial interplay between the feature extractor, the classifier, and the networks $f$ and $\phi$.
• Figure 2: The red solid circles represent the source domain samples, the blue solid circles represent the target domain samples, the red dashed area represents the source domain data distribution, and the blue dashed area represents the target domain distribution area. The left subfigure shows that the two domain distributions are strictly aligned, and the right subfigure shows that the target domain distribution is included in the source domain distribution.
• Figure 3: Distances on ${\rm{P}} \to C$ task of Office-home dataset.
• Figure 4: The influence of $\alpha$
• Figure 5: The influence of $\lambda$

Theorems & Definitions (21)

• Definition 4.1
• Definition 4.2
• Theorem 4.1
• Theorem 4.2
• Proposition 4.1
• Theorem 5.1
• Theorem 5.2
• Lemma 8.1
• proof
• Proposition 8.1