Deep Cocktail Network: Multi-source Unsupervised Domain Adaptation with Category Shift (1803.00830v1)

Abstract: Unsupervised domain adaptation (UDA) conventionally assumes labeled source samples coming from a single underlying source distribution. Whereas in practical scenario, labeled data are typically collected from diverse sources. The multiple sources are different not only from the target but also from each other, thus, domain adaptater should not be modeled in the same way. Moreover, those sources may not completely share their categories, which further brings a new transfer challenge called category shift. In this paper, we propose a deep cocktail network (DCTN) to battle the domain and category shifts among multiple sources. Motivated by the theoretical results in \cite{mansour2009domain}, the target distribution can be represented as the weighted combination of source distributions, and, the multi-source unsupervised domain adaptation via DCTN is then performed as two alternating steps: i) It deploys multi-way adversarial learning to minimize the discrepancy between the target and each of the multiple source domains, which also obtains the source-specific perplexity scores to denote the possibilities that a target sample belongs to different source domains. ii) The multi-source category classifiers are integrated with the perplexity scores to classify target sample, and the pseudo-labeled target samples together with source samples are utilized to update the multi-source category classifier and the feature extractor. We evaluate DCTN in three domain adaptation benchmarks, which clearly demonstrate the superiority of our framework.

Deep Cocktail Network: Multi-source Unsupervised Domain Adaptation with Category Shift

This paper introduces the Deep Cocktail Network (DCTN), an approach designed to address the challenges in unsupervised domain adaptation (UDA) when dealing with multiple source domains. Traditional UDA frameworks often operate under the assumption of a single source domain. However, in real-world scenarios, labeled data are frequently collected from diverse sources which do not share a common underlying distribution. Furthermore, these sources may not cover the same category sets, leading to what is termed as "category shift." This paper explores this practical setup extensively and proposes a solution through the DCTN framework.

Key Contributions and Methodology

1. Multi-source Domain Adaptation Framework: The framework proposes that the target distribution can be expressed as a weighted combination of multiple source distributions. This mulit-source consideration is critical for applications where data collection involves various platforms or environments with shared and unique categories among the sources.
2. Deep Cocktail Network Architecture:

DCTN integrates three main components: - A feature extractor for mapping data from various domains into a common feature space. - A multi-source domain discriminator to manage adversarial learning and estimate the perplexity scores, which reflect the likelihood of target samples belonging to each source domain. - A multi-source category classifier which, together with the aforementioned perplexity scores, determines the class membership of target samples.

1. Training Approach:

The network training consists of two alternating adaptation processes: - Multi-way Adversarial Learning: This step aims to reduce discrepancies between the target and each source domain, ensuring that features are extracted uniformly across different domains. - Target Discriminative Adaptation: This involves pseudo-labeling of target samples, which are subsequently used alongside source samples to update the model components, thus enhancing classification performance on the target domain with limited labeled data.

1. Handling Category Shift: Unlike typical MDA settings, DCTN addresses scenarios where there is a categorical disalignment among the sources. This capability allows the model to maintain high performance and reduce negative transfer, even when certain categories are present in only some source domains.

Experimental Validation

The paper presents comprehensive evaluations using benchmarks such as Office-31, ImageCLEF-DA, and Digits-five. These experiments demonstrate that DCTN significantly outperforms both traditional single-source UDA methods and alternative multi-source unsupervised approaches in both vanilla and category shift settings.

• Office-31 and ImageCLEF-DA: DCTN achieved consistently high accuracy, indicating its robustness to domain shifts across various domains.
• Digits-five: On this complex multi-source dataset, the model showcases its strengths in scenarios with significant inter-domain variability.

Implications and Future Directions

The DCTN framework provides a viable solution to multi-source domain adaptation challenges, suggesting theoretical implications for how domain similarity can be computed and represented in learning algorithms. Practically, it opens avenues for training models on aggregated data sources without homogenizing domains, which is common in practice due to accessibility of diverse yet incomplete datasets.

Future work might focus on enhancing robustness to further shifts in both domain and category, extending the framework to online learning contexts where the domain landscape can change incrementally. Additionally, investigating computational efficiency and scalability of such models when dealing with very large datasets could also be a vital step toward more broadly applicable multi-source domain adaptation frameworks.