Learning to Transfer Examples for Partial Domain Adaptation (1903.12230v2)

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.

• The paper presents the Example Transfer Network (ETN) that selectively quantifies the transferability of source examples to address mismatched label spaces.
• It employs adversarial training to learn domain-invariant features and progressively weight only the beneficial source examples for the target task.
• ETN demonstrates superior performance on benchmark datasets like Office-31 and Office-Home, reducing negative transfer compared to conventional domain adaptation methods.

An Analytical Summary of "Learning to Transfer Examples for Partial Domain Adaptation"

The paper "Learning to Transfer Examples for Partial Domain Adaptation" tackles the challenge of adapting machine learning models to new environments where only part of the label space is shared between the source and target domains. This is a significant issue in domain adaptation (DA), where the objective is to enable models trained on one dataset to perform adequately on another, typically unlabeled, dataset with different characteristics.

Core Contributions

The primary contribution of this paper is the Example Transfer Network (ETN), a framework specifically designed for Partial Domain Adaptation (PDA). Unlike standard DA approaches that assume the same label space shared across domains, PDA deals with cases where the source label space contains additional labels not present in the target space. This mismatch often results in negative transfer, where the irrelevant source information degrades the model's performance on the target task.

ETN employs a mechanism that simultaneously learns domain-invariant features and assigns a progressive weight to each example from the source domain to ascertain its relevance for the target task. The novelty lies in how ETN quantifies the "transferability" of source examples, helping the model to focus on data that is likely to benefit the target domain task.

Methodology in Detail

The authors incorporate a domain adversarial training paradigm while introducing a scheme to quantitatively measure each source example's transferability. The framework incorporates:

• Domain-Invariant Representation Learning: Utilizing domain adversarial networks to minimize the discrepancies between source and target domains.
• Transferability Quantification: A progressive weighting scheme that evaluates the potential benefit of each source example for transfer learning. This is achieved through a transferability quantifier, which is effectively a refined domain discriminator augmented with label prediction tasks.

The computational approach is rooted in adversarial training, where the model alternates between optimizing for classification accuracy of transferable examples and misleading a domain discriminator that tries to distinguish between source and target domains. Importantly, this discrimination accounts for both domain differences and class label distribution, offering a more fine-grained resistance to any detrimental effects arising from outlier classes in the source domain.

Empirical Evaluation

ETN demonstrates consistent performance improvements on multiple benchmark datasets, including Office-31 and Office-Home, compared to baseline and contemporary DA methodologies like DANN, SAN, and IWAN. A standout aspect of ETN is its resilience against the misalignment that typically occurs when negative transfer is prevalent, showcasing its effectiveness even in complex scenarios where the sizes of the source and target label spaces significantly differ.

Implications and Future Directions

The implications of this work are primarily seen in scenarios where large, labeled datasets are leveraged for enhancing model performance in domains with sparse or differently distributed labels. The capability to filter out irrelevant examples efficiently is beneficial, making this approach particularly useful in industrial applications where labeled data can be exorbitant to acquire systematically.

Theoretically, the approach presents a paradigm where domain adaptation is approached as a selective transfer problem, hinting at future research directions such as exploring more dynamic or self-evolving weighting mechanisms that potentially adapt over time or through online learning frameworks.

Future developments could also include integrating ETN with other unsupervised or semi-supervised learning techniques to enhance performance where labeled target data is entirely unavailable or prohibitively expensive to acquire. Moreover, further research can focus on improving the computational efficiency of ETN and its scalability in significantly larger domains or datasets.

In conclusion, the ETN framework proposed in this work advances the field of partial domain adaptation by addressing the practical challenges posed by mismatched label spaces, setting a precedent for future exploration and application in various AI and machine learning fields.