Hosei University Repository
Not a member yet
22008 research outputs found
Sort by
the Establishment and Development Process of Social Allowances in the Social Security System : From the Perspective of Central-Local Relations
Get PDF博士(公共政策学)法政大学 (Hosei University
Estimation of reservoir sedimentation considering land failures triggered by precipitation as well as sediment transport in river channels
Get PDF博士(工学)法政大学 (Hosei University
Improving Prediction Performance and Model Interpretability through Attention Mechanisms from Basic and Applied Research Perspectives
Get PDFWith the dramatic advances in deep learning technology, machine learning research is focusing on improving the interpretability of model predictions as well as prediction performance in both basic and applied research. While deep learning models have much higher prediction performance than conventional machine learning models, the specific prediction process is still difficult to interpret and/or explain. This is known as the black-boxing of machine learning models and is recognized as a particularly important problem in a wide range of research fields, including manufacturing, commerce, robotics, and other industries where the use of such technology has become commonplace, as well as the medical field, where mistakes are not tolerated.Focusing on natural language processing tasks, we consider interpretability as the presentation of the contribution of a prediction to an input word in a recurrent neural network. In interpreting predictions from deep learning models, much work has been done mainly on visualization of importance mainly based on attention weights and gradients for the inference results. However, it has become clear in recent years that there are not negligible problems with these mechanisms of attention mechanisms and gradients-based techniques. The first is that the attention weight learns which parts to focus on, but depending on the task or problem setting, the relationship with the importance of the gradient may be strong or weak, and these may not always be strongly related. Furthermore, it is often unclear how to integrate both interpretations. From another perspective, there are several unclear aspects regarding the appropriate application of the effects of attention mechanisms to real-world problems with large datasets, as well as the properties and characteristics of the applied effects. This dissertation discusses both basic and applied research on how attention mechanisms improve the performance and interpretability of machine learning models.From the basic research perspective, we proposed a new learning method that focuses on the vulnerability of the attention mechanism to perturbations, which contributes significantly to prediction performance and interpretability. Deep learning models are known to respond to small perturbations that humans cannot perceive and may exhibit unintended behaviors and predictions. Attention mechanisms used to interpret predictions are no exception. This is a very serious problem because current deep learning models rely heavily on this mechanism. We focused on training techniques using adversarial perturbations, i.e., perturbations that dares to deceive the attention mechanism. We demonstrated that such an adversarial training technique makes the perturbation-sensitive attention mechanism robust and enables the presentation of highly interpretable predictive evidence. By further extending the proposed technique to semi-supervised learning, a general-purpose learning model with a more robust and interpretable attention mechanism was achieved.From the applied research perspective, we investigated the effectiveness of the deep learning models with attention mechanisms validated in the basic research, are in real-world applications. Since deep learning models with attention mechanisms have mainly been evaluated using basic tasks in natural language processing and computer vision, their performance when used as core components of applications and services has often been unclear. We confirm the effectiveness of the proposed framework with an attention mechanism by focusing on the real world of applications, particularly in the field of computational advertising, where the amount of data is large, and the interpretation of predictions is necessary. The proposed frameworks are new attempts to support operations by predicting the nature of digital advertisements with high serving effectiveness, and their effectiveness has been confirmed using large-scale ad-serving data.In light of the above, the research summarized in this dissertation focuses on the attention mechanism, which has been the focus of much attention in recent years, and discusses its potential for both basic research in terms of improving prediction performance and interpretability, and applied research in terms of evaluating it for real-world applications using large data sets beyond the laboratory environment. The dissertation also concludes with a summary of the implications of these findings for subsequent research and future prospects in the field.博士（工学）法政大学 (Hosei University
The spirit of Japanese martial arts seen in NAKAGAWA Kazumasa’s painting : Inheriting the essence of MIYAMOTO Musashi’s Dokkōdō (‘The path of going alone’) (Awarded an honorable mention at the AKEGARASU Haya Award sponsored by Hakusan City, Ishikawa Prefecture)
Get PDFA study about the feeling of “real self” seen by those who have undergone cosmetic surgery
Get PDFThe medical practice of cosmetic surgery, which has become popular among young people in recent years, is seen positively by some, and negatively by others. In particular, those who perceive it negatively tend to perceive the body before cosmetic surgery as an absolute original, and seem to be concerned about the danger of irreversible changes to such a genuine body. So what do people who undergo cosmetic surgery think of their real bodies? Also, don’t you feel the risk of giving your body an irreversible change? Based on a questionnaire survey and an interview survey, this study focuses on such bodily sensations and aims to clarify how people who have undergone cosmetic surgery perceive their “true selves.” In the questionnaire survey, it was found that experience and interest in cosmetic surgery are related to negative and positive feelings about it. In particular, the irreversibility of cosmetic surgery has caused a sense of incongruity in the sameness before and after cosmetic surgery, and has been a factor in negative perceptions. On the other hand, from the interview survey, it is clear that those who actually experienced cosmetic surgery felt that irreversibility was a positive aspect of it, and that they felt that they were their “true selves” after cosmetic surgery.From the above survey, the following three points can be pointed out. First, resistance feeling to cosmetic surgery is easily influenced by the values of those around them. Second, the sense of “true self” is also created by the values of those around us. Thirdly, this belief of those who affirm cosmetic surgery and think of themselves as their “real selves” after cosmetic surgery is also prone to crumbling
Toward Transnational Japanese Studies
Get PDFThis paper discusses the transnational approach to Japanese studies. As a result of the expansion of practices cutting across nation-states since the end of the 20th century, different approaches that explore our world from a transnational lens have significantly enhanced. Transnational studies that began in sociology, historical studies, and cultural studies have recently applied to area studies, including Japanese studies. The phrase “transnational Japanese studies,” however, sounds contradictory as it combines two words with contrary meanings – transnational and national (Japan). What do we mean by exploring national phenomena from a transnational perspective? Would this approach be a denial of Japanese studies or rather a rebirth of it? I argue that the transnational approach would relativize the nation-state framework while bringing us an alternative view of Japan. This paper will first discuss the concept of transnational by comparing it with similar concepts such as global, international, and boundary-crossing (ekkyō). Then, it will explore how global and transnational turn has evolved in different fields – sociology, historical studies, and area studies – to overcome the so-called methodological nationalism that was deeply embedded in modern social science and humanities. The final section will discuss the main part of this paper - transnational Japanese studies. While to overcome methodological nationalism in area studies, area specialists have moved away from nation-states and focused on the larger region – such as “Asia.” Given the remaining presence of national states in our society, I suggest that instead of evading the nation-state from its analysis, “bringing the nation-state back in” to the transnational analysis will generate new insights into our understanding of Japan
石橋湛山はなぜ政界への進出を志したか : 戦前の言論人としての活動を手掛かりに
Get PDF1910年代から言論活動に従事した石橋湛山（1884-1973）は、戦前は東洋経済新報社の発行する『東洋時論』を経て経済専門誌『東洋経済新報』を中心に経済、政治、外交、文化など多岐にわたる分野に関する評論を行った。また、戦後は1946年に大日本帝国憲法下での最後の総選挙となった第22回衆議院議員総選挙に出馬して国政に関わるようになった。その結果は落選であったものの、言論人が政治家を目指したことは、活動の場を言論から実践に移す試みの一つとして見逃せない。従来の石橋湛山に関する研究では戦前の言論活動と戦後の政治活動との間の関連について十分な検討がなされてこなかった。そこで、本論では、石橋湛山の言論活動について、その代表的な所論である「小日本主義」や金解禁論争に注目し、議論の内容と実際の政策への影響を検討した。その結果、石橋が政策の提案や当局の批判によっては実際の政治を改めることが出来ず、自分自身が政治に関与しない限り所説を実現できないことを自覚したことが、戦後の政界への進出に繋がったことを明らかにした
