Partial Transfer Learning with Selective Adversarial Networks (1707.07901v1)

Abstract: Adversarial learning has been successfully embedded into deep networks to learn transferable features, which reduce distribution discrepancy between the source and target domains. Existing domain adversarial networks assume fully shared label space across domains. In the presence of big data, there is strong motivation of transferring both classification and representation models from existing big domains to unknown small domains. This paper introduces partial transfer learning, which relaxes the shared label space assumption to that the target label space is only a subspace of the source label space. Previous methods typically match the whole source domain to the target domain, which are prone to negative transfer for the partial transfer problem. We present Selective Adversarial Network (SAN), which simultaneously circumvents negative transfer by selecting out the outlier source classes and promotes positive transfer by maximally matching the data distributions in the shared label space. Experiments demonstrate that our models exceed state-of-the-art results for partial transfer learning tasks on several benchmark datasets.

• The paper presents a novel selective adversarial mechanism that aligns shared classes, reducing negative transfer in partial domain adaptation.
• It integrates entropy minimization to promote clear feature separation in the target domain, enhancing classification accuracy.
• Experimental results on benchmarks like Office-31 demonstrate significant improvements compared to existing transfer learning methods.

Overview of "Partial Transfer Learning with Selective Adversarial Networks"

The paper "Partial Transfer Learning with Selective Adversarial Networks," authored by Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I. Jordan, presents an innovative approach to address challenges in partial transfer learning. The motivation stems from the need to transfer models trained on comprehensive large-scale source domains to smaller target domains with potentially unshared label spaces. Traditional transfer learning models assume a fully shared label space across domains, which is often unrealistic in practical scenarios, leading to negative transfers when target and source domains differ in label distribution.

Key Concepts and Proposed Method

The authors introduce a concept of partial transfer learning where the target label space is a subset of the source label space. This adjustment in assumption is pivotal as it acknowledges that negative transfer can occur when source and target classifications are mismatched. To address this, the paper presents Selective Adversarial Networks (SAN), a methodological extension of existing domain adversarial networks which selectively aligns shared classes between domains while isolating outlier classes that could lead to performance degradation.

The novelty of SAN lies in its dual capability:

• Selective Adversarial Mechanism: The SAN model utilizes multiple domain discriminators corresponding to each class in the source domain. It leverages the label prediction probability to adjust class-wise domain adaptation processes. This innovation is targeted at preventing negative transfer by minimizing unmatched class influences.
• Entropy Minimization: The paper enhances this mechanism with entropy minimization, which encourages low-density separation in the target dataset, promoting the learning of more distinct and relevant features.

These elements are incorporated into a deep learning framework enabling an end-to-end training procedure that adjusts both transfer features and classifier parameters.

Experimental Outcomes

The experimental evaluation is comprehensive, involving benchmark datasets such as Office-31, Caltech-Office, and ImageNet-Caltech. The SAN model consistently outperforms other state-of-the-art methods, including variants like Deep Adaptation Network (DAN), Reverse Gradient (RevGrad), and Residual Transfer Networks (RTN). Notably, the SAN methodology achieves average classification accuracy surges, particularly in scenarios with large source domains and fewer target classes.

Noteworthy numerical results from the experiments underscore SAN’s robustness. For example, on the task Office A 31 → W 10, SAN registers an accuracy of 80.02%, surpassing the previous best by a notable margin. Such outcomes highlight the efficacy of selective adversarial learning in reducing negative transfers and adapting to shared label spaces only.

Implications and Future Directions

The implications of SAN for future AI and machine learning research are substantial. By introducing a granularity in adversarial adaptation aligned with realistic data distribution assumptions, the approach bridges the applicability of transfer learning in real-world applications where overlapping class scenarios across domains are inevitable.

Theoretically, SAN contributes a structured approach to handling domain discrepancies at a class level, providing a more nuanced understanding of feature alignment in neural networks. Practically, this method can be pivotal in enhancing tasks like cross-domain image classification, sentiment analysis across different corpora, and more.

Future research could extend this work by exploring alternative domain discriminator architectures, applying the model to multi-modality tasks, and examining the effect of variations in network architecture on domain adaptation quality. Additionally, scalability and computational efficiency of SAN in larger-scale applications remain critical areas for further exploration.

In summary, "Partial Transfer Learning with Selective Adversarial Networks" offers a robust solution to the partial label space discrepancy issue in transfer learning, marking a significant progression in domain adaptation strategies. This work serves as a foundation for more adaptive and selective learning models tailored to practical domain-specific challenges.