[Paper Review] Domain Confusion with Self Ensembling for Unsupervised Adaptation

Jiawei Wang, Zhaoshui He+5 more|arXiv (Cornell University)|Oct 10, 2018

Domain Adaptation and Few-Shot Learning18 references4 citations

TL;DR

This paper proposes Domain Confusion with Self Ensembling (DCSE), a novel unsupervised domain adaptation method that jointly optimizes feature and class invariance through domain confusion and self-ensembling to address training instability and mistaken confusion in adversarial adaptation. The approach achieves state-of-the-art performance across multiple benchmarks by leveraging both domain-level and sample-level consistency.

ABSTRACT

Data collection and annotation are time-consuming in machine learning, expecially for large scale problem. A common approach for this problem is to transfer knowledge from a related labeled domain to a target one. There are two popular ways to achieve this goal: adversarial learning and self training. In this article, we first analyze the training unstablity problem and the mistaken confusion issue in adversarial learning process. Then, inspired by domain confusion and self-ensembling methods, we propose a combined model to learn feature and class jointly invariant representation, namely Domain Confusion with Self Ensembling (DCSE). The experiments verified that our proposed approach can offer better performance than empirical art in a variety of unsupervised domain adaptation benchmarks.

Motivation & Objective

• To address training instability and mistaken confusion in adversarial unsupervised domain adaptation.
• To improve generalization by jointly learning domain-invariant and class-invariant representations.
• To combine the strengths of domain confusion and self-ensembling for more robust feature learning.
• To achieve state-of-the-art performance on standard unsupervised domain adaptation benchmarks.

Proposed method

• The method introduces a domain confusion module that aligns features across source and target domains using adversarial training.
• It incorporates a self-ensembling component that enforces consistency of predictions across multiple forward passes with data augmentations.
• The model jointly optimizes domain confusion and self-ensembling objectives to learn representations invariant to both domain shift and class distribution.
• Feature and classifier parameters are updated end-to-end using a combined loss function that balances domain confusion and prediction consistency.
• The approach uses data augmentation to generate multiple views of the same input, enhancing the reliability of self-ensembling.

Experimental results

Research questions

• RQ1 How can training instability in adversarial domain adaptation be mitigated?
• RQ2 To what extent can joint optimization of domain and class invariance improve adaptation performance?
• RQ3 Can self-ensembling enhance robustness in unsupervised domain adaptation when combined with domain confusion?
• RQ4 Does the proposed method outperform existing state-of-the-art approaches on standard benchmarks?

Key findings

• The proposed DCSE method achieves superior performance compared to existing state-of-the-art methods on multiple unsupervised domain adaptation benchmarks.
• The combination of domain confusion and self-ensembling effectively reduces training instability observed in pure adversarial approaches.
• The method mitigates the mistaken confusion issue by enforcing consistent predictions across augmented views of the same sample.
• Joint optimization of domain and class invariance leads to more robust and generalizable feature representations.