Open Set Domain Adaptation by Backpropagation (1804.10427v2)

Abstract: Numerous algorithms have been proposed for transferring knowledge from a label-rich domain (source) to a label-scarce domain (target). Almost all of them are proposed for a closed-set scenario, where the source and the target domain completely share the class of their samples. We call the shared class the \doublequote{known class.} However, in practice, when samples in target domain are not labeled, we cannot know whether the domains share the class. A target domain can contain samples of classes that are not shared by the source domain. We call such classes the \doublequote{unknown class} and algorithms that work well in the open set situation are very practical. However, most existing distribution matching methods for domain adaptation do not work well in this setting because unknown target samples should not be aligned with the source. In this paper, we propose a method for an open set domain adaptation scenario which utilizes adversarial training. A classifier is trained to make a boundary between the source and the target samples whereas a generator is trained to make target samples far from the boundary. Thus, we assign two options to the feature generator: aligning them with source known samples or rejecting them as unknown target samples. This approach allows extracting features that separate unknown target samples from known target samples. Our method was extensively evaluated in domain adaptation setting and outperformed other methods with a large margin in most settings.

• The paper introduces an adversarial learning framework that effectively distinguishes known target samples from unknown classes without relying on unknown source data.
• The method leverages backpropagation to generate a boundary between source and target feature distributions, enhancing class segregation.
• Results on Office, VisDA, and digit datasets show superior performance in detecting unknown classes compared to traditional closed-set approaches.

Open Set Domain Adaptation by Backpropagation

This paper addresses the challenge of open set domain adaptation (OSDA), where the target domain includes samples of classes not present in the source domain. Traditional domain adaptation approaches are designed for the closed-set scenario, assuming complete class overlap between source and target domains. However, this assumption is impractical in real-world applications, where unknown classes can exist in the target domain, necessitating methods that handle open set scenarios effectively.

Methodology

The authors propose an approach utilizing adversarial training to tackle OSDA, exploiting the capabilities of deep neural networks. The approach involves:

1. Adversarial Learning Framework: The framework consists of a classifier and a feature generator. The classifier is trained to differentiate between known and unknown classes by outputting a probability for the unknown class for samples from the target domain.
2. Boundary Formation: The classifier forms a boundary between source and target samples. The generator perturbs target features to either align with known source samples or divert them as unknown targets.
3. No Dependency on Unknown Source Samples: Unlike previous methods that utilize unknown samples in the source domain to detect unknowns in the target, this approach does not require unknown source data, increasing its practicality.

Results

The proposed method is evaluated across several datasets, including Office, VisDA, and digit datasets (MNIST, USPS, SVHN). The results demonstrate:

• Office Dataset: The method achieves superior performance in detecting unknown classes compared to baselines such as OSVM (Open Set SVM), MMD (Maximum Mean Discrepancy), and BP (Backpropagation-based domain classifier). Specifically, it outperforms in scenarios where the number of known classes is increased.
• VisDA Dataset: When applied to adaptation from synthetic to real images, the approach showed robust performance in discriminating known from unknown classes, further demonstrating its effectiveness in diverse domains.
• Digit Datasets: In experiments involving MNIST, USPS, and SVHN, the method consistently surpassed existing techniques, highlighting its efficacy in managing domain shifts between visually distinct datasets.

Theoretical and Practical Implications

The proposed method effectively separates known target samples from unknowns without relying on explicit unknown samples in the source domain, making it applicable to a wider range of real-world scenarios. This capability addresses a significant limitation of existing domain adaptation techniques, promising improved utility in applications such as autonomous systems and image recognition tasks where encountering unknowns is inevitable.

Future Directions

Future research could explore the integration of more sophisticated adversarial strategies or the incorporation of self-supervised or semi-supervised learning paradigms to enhance feature extraction. Moreover, exploring its application to other domains like natural language processing or multi-modal datasets may yield further insights and advancements.

This paper contributes a substantial advancement in open set domain adaptation by providing a framework that handles class discrepancies between domains efficiently, extending the applicability of domain adaptation techniques beyond the constraints of closed-set assumptions.