Unsupervised Domain Adaptation by Statistics Alignment for Deep Sleep Staging Networks

Deep sleep staging networks have reached top performance on large-scale datasets. However, these models perform poorer when training and testing on small sleep cohorts due to data inefficiency. Transferring well-trained models from large-scale datasets (source domain) to small sleep cohorts (target domain) is a promising solution but still remains challenging due to the domain-shift issue. In this work, an unsupervised domain adaptation approach, domain statistics alignment (DSA), is developed to bridge the gap between the data distribution of source and target domains. DSA adapts the source models on the target domain by modulating the domain-specific statistics of deep features stored in the Batch Normalization (BN) layers. Furthermore, we have extended DSA by introducing cross-domain statistics in each BN layer to perform DSA adaptively (AdaDSA). The proposed methods merely need the well-trained source model without access to the source data, which may be proprietary and inaccessible. DSA and AdaDSA are universally applicable to various deep sleep staging networks that have BN layers. We have validated the proposed methods by extensive experiments on two state-of-the-art deep sleep staging networks, DeepSleepNet+ and U-time. The performance was evaluated by conducting various transfer tasks on six sleep databases, including two large-scale databases, MASS and SHHS, as the source domain, four small sleep databases as the target domain. Thereinto, clinical sleep records acquired in Huashan Hospital, Shanghai, were used. The results show that both DSA and AdaDSA could significantly improve the performance of source models on target domains, providing novel insights into the domain generalization problem in sleep staging tasks.<br>

SA-GAN: Stain Acclimation Generative Adversarial Network for Histopathology Image Analysis

Histopathological image analysis is an examination of tissue under a light microscope for cancerous disease diagnosis. Computer-assisted diagnosis (CAD) systems work well by diagnosing cancer from histopathology images. However, stain variability in histopathology images is inevitable due to the use of different staining processes, operator ability, and scanner specifications. These stain variations present in histopathology images affect the accuracy of the CAD systems. Various stain normalization techniques have been developed to cope with inter-variability issues, allowing standardizing the appearance of images. However, in stain normalization, these methods rely on the single reference image rather than incorporate color distributions of the entire dataset. In this paper, we design a novel machine learning-based model that takes advantage of whole dataset distributions as well as color statistics of a single target image instead of relying only on a single target image. The proposed deep model, called stain acclimation generative adversarial network (SA-GAN), consists of one generator and two discriminators. The generator maps the input images from the source domain to the target domain. Among discriminators, the first discriminator forces the generated images to maintain the color patterns as of target domain. While second discriminator forces the generated images to preserve the structure contents as of source domain. The proposed model is trained using a color attribute metric, extracted from a selected template image. Therefore, the designed model not only learns dataset-specific staining properties but also image-specific textural contents. Evaluated results on four different histopathology datasets show the efficacy of SA-GAN to acclimate stain contents and enhance the quality of normalization by obtaining the highest values of performance metrics. Additionally, the proposed method is also evaluated for multiclass cancer type classification task, showing a 6.9% improvement in accuracy on ICIAR 2018 hidden test data.

Unsupervised Image-Generation Enhanced Adaptation for Object Detection in Thermal Images

Object detection in thermal images is an important computer vision task and has many applications such as unmanned vehicles, robotics, surveillance, and night vision. Deep learning-based detectors have achieved major progress, which usually need large amount of labelled training data. However, labelled data for object detection in thermal images is scarce and expensive to collect. How to take advantage of the large number labelled visible images and adapt them into thermal image domain is expected to solve. This paper proposes an unsupervised image-generation enhanced adaptation method for object detection in thermal images. To reduce the gap between visible domain and thermal domain, the proposed method manages to generate simulated fake thermal images that are similar to the target images and preserves the annotation information of the visible source domain. The image generation includes a CycleGAN-based image-to-image translation and an intensity inversion transformation. Generated fake thermal images are used as renewed source domain, and then the off-the-shelf domain adaptive faster RCNN is utilized to reduce the gap between the generated intermediate domain and the thermal target domain. Experiments demonstrate the effectiveness and superiority of the proposed method.

DAFI

WiFi-based Device-free Passive (DfP) indoor localization systems liberate their users from carrying dedicated sensors or smartphones, and thus provide a non-intrusive and pleasant experience. Although existing fingerprint-based systems achieve sub-meter-level localization accuracy by training location classifiers/regressors on WiFi signal fingerprints, they are usually vulnerable to small variations in an environment. A daily change, e.g., displacement of a chair, may cause a big inconsistency between the recorded fingerprints and the real-time signals, leading to significant localization errors. In this paper, we introduce a Domain Adaptation WiFi (DAFI) localization approach to address the problem. DAFI formulates this fingerprint inconsistency issue as a domain adaptation problem, where the original environment is the source domain and the changed environment is the target domain. Directly applying existing domain adaptation methods to our specific problem is challenging, since it is generally hard to distinguish the variations in the different WiFi domains (i.e., signal changes caused by different environmental variations). DAFI embraces the following techniques to tackle this challenge. 1) DAFI aligns both marginal and conditional distributions of features in different domains. 2) Inside the target domain, DAFI squeezes the marginal distribution of every class to be more concentrated at its center. 3) Between two domains, DAFI conducts fine-grained alignment by forcing every target-domain class to better align with its source-domain counterpart. By doing these, DAFI outperforms the state of the art by up to 14.2% in real-world experiments.

Profiling of the Ethical Concepts Good / Evil and Justice from the Etymological Perspective

The paper gives a comprehensive insight into the peculiarities of concept profiling through defining the related etymological domains. The aim of the paper is to reveal the peculiarities of the profiling of the concepts GOOD / EVIL and JUSTICE through elucidating their source domains from an etymological perspective. The choice of the analysed ethical concepts is stipulated by their higher contextual actualization frequency, as compared with the other ethical concepts, according to the data obtained from British National Corpus. The research method of trajector / landmark alignment used in this work is based on R. Langacker’s views on the profiling of concepts in language utterances and on the tenets of Conceptual Metaphor Theory. The novelty of this approach consists in the fact that it was elaborated and tailored for the analysis and explanation of the onomasiological principles of ‘wrapping’ abstract ethical concepts into the language form. The underlying idea is that abstract categories were conceptualized on the basis of background central life experience and knowledge of concrete things. It is argued that such things were the source domains in ontological cross-domain mappings for the target ethical concepts. The current research into the source domain etymons of the ethical categories made it possible to determine the underlying images, which are the core for drawing metaphorical correspondences between source and target concepts. The etymological layer of source domain lexicalizers revealed the intrinsic psychological mechanisms of human cognition and perception of the world, which consist in the inherent proclivity of the human mind to make metaphorical parallels in the direction from daily, central experience to complex abstract ideas and notions. The results made it possible to develop the matrix model of the analysed concepts, which was further developed into a ‘conceptual edifice’ multi-layer model, which reveals the conceptualization paths along which the human mind classifies and categorizes abstract ethical ideas.

Comparative Study of Intertexts in Newspaper Headlines (Based on the Material of the Russian and British Discourses of Sanctions Policy Towards Russia)

The article presents a comparative analysis of intertexts in the headlines of British and Russian articles, the subject of which is the sanctions imposed by Western countries on Russia. The source of the material was the authoritative publications of Russia and Great Britain (Kommersant and The Economist). The chronological scope of the study lies within 2014 – current days. As a result of the analysis of the headlines (60 in Russian, 60 in English) containing intertexts, the author identified 5 source domains, common for British and Russian discourses, and one source domain, specific for one discourse (Russian or British): “Set expressions”, “Proverbs and sayings”, “Mass culture”, “Literature”, “Proper names”, “Situations” (for British articles), “Names of organizations / projects” (for Russian articles). The study showed that the source domain “Set expressions” turned out to be the most widely represented, and the share of intertexts from the source domain in the Russian-language headlines is more than twice bigger than their share in the British headlines. The group of intertexts from the source domain “Proverbs and sayings” turned out to be the second most represented. The most common way to include an intertextual element in a headline is lexical transformation, when a word or several words are replaced with others corresponding to the content of the article. As for the effect of including an intertextual element in the headline it can be different: comic, if the author chooses a lexical transformation that contradicts the meaning of the original text; creating a negative / positive (usually negative) image of a person or situation; forming the reader’s opinion, similar to the opinion of the author / edition.

Language of Persuasion: Analysis of Conceptual Metaphors in Political Discourse

The aim of this article is to study the scope of conceptual metaphors as a persuasive tool inherent to political discourse in English. In particular, it dwells upon the use of four conceptual metaphors such as NATION IS A FAMILY, STATE IS A BODY, POLITICS IS A WAR, and POLITICS IS A GAME. For this purpose, the transcripts of twenty-eight public speeches delivered by David Cameron, Hillary Clinton, Theresa May, and Donald Trump were analysed. The results revealed numerous functions of these metaphors in the process of persuasion. Apart from that, the analysis showed that the majority of the analysed politicians resort to the source domain of WAR to conceptualise their political activities, while the source domain of GAME is the least frequently used.