Implicit Class-Conditioned Domain Alignment for Unsupervised Domain Adaptation

ICML 2020  ยท  Xiang Jiang, Qicheng Lao, Stan Matwin, Mohammad Havaei ยท

We present an approach for unsupervised domain adaptation---with a strong focus on practical considerations of within-domain class imbalance and between-domain class distribution shift---from a class-conditioned domain alignment perspective. Current methods for class-conditioned domain alignment aim to explicitly minimize a loss function based on pseudo-label estimations of the target domain. However, these methods suffer from pseudo-label bias in the form of error accumulation. We propose a method that removes the need for explicit optimization of model parameters from pseudo-labels directly. Instead, we present a sampling-based implicit alignment approach, where the sample selection procedure is implicitly guided by the pseudo-labels. Theoretical analysis reveals the existence of a domain-discriminator shortcut in misaligned classes, which is addressed by the proposed implicit alignment approach to facilitate domain-adversarial learning. Empirical results and ablation studies confirm the effectiveness of the proposed approach, especially in the presence of within-domain class imbalance and between-domain class distribution shift.

PDF Abstract ICML 2020 PDF

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Unsupervised Domain Adaptation Office-31 Implicit Alignment (with MDD) Avg accuracy 88.8 # 1
Unsupervised Domain Adaptation Office-Home Implicit Alignment (with MDD) Avg accuracy 69.5 # 1
Unsupervised Domain Adaptation Office-Home (RS-UT imbalance) COAL Average Per-Class Accuracy 58.4 # 2
Unsupervised Domain Adaptation Office-Home (RS-UT imbalance) MDD Average Per-Class Accuracy 55.44 # 4
Unsupervised Domain Adaptation Office-Home (RS-UT imbalance) Source Only Average Per-Class Accuracy 52.81 # 5
Unsupervised Domain Adaptation Office-Home (RS-UT imbalance) DANN Average Per-Class Accuracy 56.91 # 3
Unsupervised Domain Adaptation Office-Home (RS-UT imbalance) Implicit Alignment (with MDD) Average Per-Class Accuracy 61.67 # 1
Unsupervised Domain Adaptation VisDA2017 Implicit Alignment (with MDD) Accuracy 75.8 # 3

Methods


No methods listed for this paper. Add relevant methods here