Zero-shot Domain Adaptation without Source Irrelevant Task-of-interest

Abstract

본 논문에서는 도메인 적응을 위한 데이터 수집 비용을 획기적으로 줄일 수 있는 새로운 source IrT-free zero-shot 도메인 적응 방식을 제안한다. 기존의 zero-shot 도메인 적응 방식들과는 달리, source IrT-free zero-shot 도메인 적응 방식을 사용할 경우 적응시키고자 하는 target 도메인에서의 데이터수집은 우리가 분류하고자 하는 클래스가 아닌, 더 수집하기 쉽고 비용이 적은 다른 클래스의 데이터 샘플만을 수집하여도 도메인 적응이 가능하다. 이를 위해, 본 논문에서는 siamese generator 뉴럴 네트워크 아키텍처와 자가 의미 매핑 제약을 제안한다. siamese generator를 사용하는 siamese GcGAN은 도메인/태스크-불변 제약과 교차-불변 제약을 통해 학습된다. 자가 의미 매핑 제약은 siamese GcGAN이 이미지-이미지 번역시 원래 가지는 의미를 잃지 않도록 방지한다. 각 컴포넌트의 분해 분석 실험을 통해, 각 컴포넌트가 성공적인 적응에 필수적임을 증명했다. 실험은 기존의 합성된 도메인들을 가지는 MNIST 벤치마크와 직접 수집한 WiFi 채널 상태 정보를 이용한 행동인식 데이터셋에서 수행되었다. 정량적 평가에서 제안한 모델이 성공적으로 도메인 적응했음을 보였고, 그 차이는 특히 도메인 적응 없이 수행한 모델의 성능이 낮은 경우에서 두드러졌다. 정성적 평가에서는 제안하는 모델의 한계점이 드러났으며, 이는 도메인/태스크-불변 특성을 잘 조합할 수 있는 decoder의 개발이 연구되어야한다는 추후 연구의 방향성을 제시하였다.

Deep neural networks outperform the traditional state-of-the-arts models in many tasks with a large margin. However, its mesmerizing performance is often degraded when we deploy the model on real world data. One of the typical reasons is that input data collected during the training phase and deployment phase have different domains. Domain adaptation has studied to mitigate the performance degradation of neural network as well as other machine learning models. Especially, unsupervised domain adaptation is attracted in recent years by computer vision area, since it enables the adaptation by utilizing a large amounts of images on target domain without labeling efforts. The approaches, however, are tailored under the assumption that we already have enough amount of unlabeled target domain data. For many real world applications, there should be significant efforts for data collection in new environment for successful adaptation, rather then labeling efforts. Zero-shot domain adaptation (ZSDA) enables domain adaptation without further task-of-interest (ToI) data collection. In this paper, we argue that conventional ZSDA schemes is not enough to guarantee low efforts in two aspects i) it assumes there already exists irrelevant task-of-interest (IrT) data which requires additional data acquisition process, and ii) it requires a paired condition or labels in the irrelevant task-of-interest datasets. Therefore, we propose a new zero-shot domain adaptation task without irrelevant task-of-interest (IrT) samples of source domain, i.e. Source IrT-free zero-shot domain adaptation. We tackle the task using geometry-consistent generative adversarial network (GcGAN). Specifically, our new pixel-level adaptation approach is working without paired condition, unlike conventional ZSDA methods. The method is evaluated on famous zero-shot domain adaptation benchmark datasets and also, a manually collected dataset for human activity recognition (HAR) task using WIFI channel state information (CSI). WIFI CSI-based HAR task is a good test-bed for applying new low-effort domain adaptation schemes for two reasons: 1) it uses multi-path information of radio frequency signal from WIFI access point, which is highly dependent on the nearby environment. Thus, it suffers a significant performance degradation on new environments 2) its data collection process is labor-intensive, and adaptation cost is inevitably expensive for deploying the models on new environments. With above benchmarks, we emprically prove that our model successfully adapts to novel domain in source IrT-free ZSDA setting.