Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Class Overwhelms: Mutual Conditional Blended-Target Domain Adaptation

Pengcheng Xu, Boyu Wang, Charles Ling

TL;DR

This work demonstrates that domain labels are not directly necessary for BTDA if categorical distributions of various domains are sufficiently aligned even facing the imbalance of domains and the label distribution shift of classes, and observes that the cluster assumption in BTDA does not comprehensively hold.

Abstract

Current methods of blended targets domain adaptation (BTDA) usually infer or consider domain label information but underemphasize hybrid categorical feature structures of targets, which yields limited performance, especially under the label distribution shift. We demonstrate that domain labels are not directly necessary for BTDA if categorical distributions of various domains are sufficiently aligned even facing the imbalance of domains and the label distribution shift of classes. However, we observe that the cluster assumption in BTDA does not comprehensively hold. The hybrid categorical feature space hinders the modeling of categorical distributions and the generation of reliable pseudo labels for categorical alignment. To address these, we propose a categorical domain discriminator guided by uncertainty to explicitly model and directly align categorical distributions $P(Z|Y)$. Simultaneously, we utilize the low-level features to augment the single source features with diverse target styles to rectify the biased classifier $P(Y|Z)$ among diverse targets. Such a mutual conditional alignment of $P(Z|Y)$ and $P(Y|Z)$ forms a mutual reinforced mechanism. Our approach outperforms the state-of-the-art in BTDA even compared with methods utilizing domain labels, especially under the label distribution shift, and in single target DA on DomainNet. Source codes are available at https://github.com/Pengchengpcx/Class-overwhelms-Mutual-Conditional-Blended-Target-Domain-Adaptation.

Class Overwhelms: Mutual Conditional Blended-Target Domain Adaptation

TL;DR

This work demonstrates that domain labels are not directly necessary for BTDA if categorical distributions of various domains are sufficiently aligned even facing the imbalance of domains and the label distribution shift of classes, and observes that the cluster assumption in BTDA does not comprehensively hold.

Abstract

Current methods of blended targets domain adaptation (BTDA) usually infer or consider domain label information but underemphasize hybrid categorical feature structures of targets, which yields limited performance, especially under the label distribution shift. We demonstrate that domain labels are not directly necessary for BTDA if categorical distributions of various domains are sufficiently aligned even facing the imbalance of domains and the label distribution shift of classes. However, we observe that the cluster assumption in BTDA does not comprehensively hold. The hybrid categorical feature space hinders the modeling of categorical distributions and the generation of reliable pseudo labels for categorical alignment. To address these, we propose a categorical domain discriminator guided by uncertainty to explicitly model and directly align categorical distributions . Simultaneously, we utilize the low-level features to augment the single source features with diverse target styles to rectify the biased classifier among diverse targets. Such a mutual conditional alignment of and forms a mutual reinforced mechanism. Our approach outperforms the state-of-the-art in BTDA even compared with methods utilizing domain labels, especially under the label distribution shift, and in single target DA on DomainNet. Source codes are available at https://github.com/Pengchengpcx/Class-overwhelms-Mutual-Conditional-Blended-Target-Domain-Adaptation.
Paper Structure (11 sections, 1 theorem, 11 equations, 6 figures, 4 tables)

This paper contains 11 sections, 1 theorem, 11 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Explicit categorical adversarial alignment
  5. Uncertainty guided discriminative adversarial training
  6. Source-only balanced adversarial training
  7. Low level feature for classifier correction
  8. Experiments
  9. Ablation and analysis
  10. Conclusion
  11. Acknowledgement

Key Result

Theorem 1

For any classifier $\hat{Y}=(h\circ g)(X)$, the blended target error rate is

Figures (6)

  • Figure 1: Left: t-SNE for hybrid categorical feature space of BTDA where features of various classes are pervasive and unstructured. The color indicates the domain and the digit indicates the class. Middle: the sample rate of the same class for each class center's K nearest neighbors. All data are collected from Office-Home (ResNet-50). Right: BTDA distribution shift where features are unstructured and the classifier is biased.
  • Figure 2: The framework of MCDA. The source data utilizes balanced sampling for training the categorical discriminator and is augmented with blended target styles to train the classifier. The target data is randomly sampled, and the predicted pseudo labels with low uncertainty are converted to one-hot labels to train the categorical domain discriminator.
  • Figure 3: Label distribution shift of Office-Home-LMT.
  • Figure 4: Samples below uncertainty threshold and pseudo label accuracy during training process on DomainNet.
  • Figure 5: t-SNE feature visualization on Office-Home-LMT with Clipart as the source. 15 classes are sampled for conciseness. Color represents the domain and the digit represents the class. left: sourceonly, right: MCDA.
  • ...and 1 more figures

Theorems & Definitions (1)

  • Theorem 1