[Paper Review] Learning to Transfer Examples for Partial Domain Adaptation

Zhangjie Cao, Kaichao You+3 more|arXiv (Cornell University)|Mar 28, 2019

Domain Adaptation and Few-Shot Learning43 references36 citations

TL;DR

ETN introduces a transferability weighting mechanism and an auxiliary discriminative module to down-weight outlier source classes and promote transfer via a learnable, end-to-end partial domain adaptation framework. It achieves state-of-the-art results on several PDA benchmarks.

ABSTRACT

Domain adaptation is critical for learning in new and unseen environments. With domain adversarial training, deep networks can learn disentangled and transferable features that effectively diminish the dataset shift between the source and target domains for knowledge transfer. In the era of Big Data, the ready availability of large-scale labeled datasets has stimulated wide interest in partial domain adaptation (PDA), which transfers a recognizer from a labeled large domain to an unlabeled small domain. It extends standard domain adaptation to the scenario where target labels are only a subset of source labels. Under the condition that target labels are unknown, the key challenge of PDA is how to transfer relevant examples in the shared classes to promote positive transfer, and ignore irrelevant ones in the specific classes to mitigate negative transfer. In this work, we propose a unified approach to PDA, Example Transfer Network (ETN), which jointly learns domain-invariant representations across the source and target domains, and a progressive weighting scheme that quantifies the transferability of source examples while controlling their importance to the learning task in the target domain. A thorough evaluation on several benchmark datasets shows that our approach achieves state-of-the-art results for partial domain adaptation tasks.

Motivation & Objective

• Motivate partial domain adaptation (PDA) where source label space subsumes target label space and target labels are unknown during training.
• Propose ETN to jointly learn domain-invariant features and a progressive per-example transferability weight.
• Reduce negative transfer by down-weighting irrelevant source examples while enhancing positive transfer from shared classes.
• Leverage an auxiliary discriminative pathway to better quantify example transferability and improve PDA performance.

Proposed method

• Introduce a transferability weight w(xs) for each source example that scales its contribution to the source classifier and the domain discriminator.
• Use an auxiliary domain discriminator ˜Gd fed by an auxiliary label predictor ˜Gy with leaky-softmax to quantify transferability and distinguish shared vs outlier classes.
• Compute w(xs) = 1 - ˜Gd(Gf(xs)) with weights normalized within each mini-batch.
• Incorporate entropy minimization on target predictions to exploit unlabeled target data.
• Formulate a minimax objective that jointly optimizes feature extractor, source classifier, and domain discriminators, including auxiliary components.
• Train auxiliary components ˜Gy and ˜Gd to incorporate discriminative and domain information for better weighting.

Experimental results

Research questions

• RQ1 How to effectively transfer knowledge from a large source label space to a smaller target label space in PDA?
• RQ2 Can a learned per-example transferability weight reduce negative transfer from outlier source classes while preserving positive transfer from shared classes?
• RQ3 Does incorporating discriminative information into auxiliary discriminators improve transferability quantification and PDA performance?
• RQ4 What is the impact of entropy minimization on leveraging unlabeled target data in PDA?

Key findings

• ETN achieves state-of-the-art results on Office-31, Office-Home, ImageNet-Caltech, and Caltech-256 PDA benchmarks.
• On Office-Home with ResNet-50, ETN averages 70.45% accuracy, outperforming prior methods like SAN (65.30) and PADA (62.06).
• On Office-31 and ImageNet-Caltech with ResNet-50, ETN attains averages of 79.08% and 79.08% respectively (per Table 2 and 3 context).
• ETN with ablations shows the full model outperforms variants lacking the classifier or auxiliary components, confirming the benefit of the transferability weighting and auxiliary discriminators.
• ETN consistently improves across multiple PDA task combinations, indicating robust transferability quantification and negative transfer mitigation.