| 浅井 俊宏 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 岡田 実 | 諏訪 博彦 | 松井 智一 |
|
title: Application of active learning to unlabeled data using a non-contact daily life sensing system. abstract: Various home action recognition methods have been proposed for application to high-quality lifestyle services such as in-home action recommendation. In particular, there is a need for low-cost, low-effort sensing while protecting residents' privacy. As a system to solve this problem, SALON, a maintenance-free home sensing system that operates wirelessly and without power supplying while protecting privacy, has been proposed. In this study, we propose a method to reduce annotation cost by applying active learning to datasets collected by SALON from home. The proposed method is expected to reduce annotation costs and missing data values and to enable more efficient data collection. 発表題目: 非接触の日常生活センシングシステムを使用した無ラベルデータに対する能動学習の適用 発表概要: 宅内での行動推薦など高品質な生活サービスに応用するため，様々な宅内行動認識手法が提案されている．特に，居住者のプライバシーに配慮しつつ，低コスト・低労力でセンシングすることが求められている．それを実現するシステムとして，プライバシーに配慮しつつ無線・無給電で動作し，メンテナンスフリーの宅内センシングシステムであるSALONが提案されている．本研究ではSALONで収集される一般家庭から得られた無ラベルデータに対して能動学習を適用し，アノテーションコストを低減する手法を提案する．提案手法では，一般家庭における問題点であったアノテーションコストを削減し，より効率的なデータ収集が可能になる事が期待される． language of the presentation: Japanese | ||||||
| 池永 拓海 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 岡田 実 | 諏訪 博彦 | 松田 裕貴 |
|
title : Proposal and Evaluation of a BLE-Based Bus Ridership and Subjective Crowdedness Estimation Method
abstract: Accurate estimation of information on buses and other modes of transportation is essential for improving the efficiency of transportation systems and enhancing passenger comfort. Existing methods have been proposed to estimate the number of passengers on a bus using BLE (Bluetooth Low Energy) signal information transmitted by devices carried by passengers. However, these approaches primarily focus on estimating the number of people on the bus, neglecting important information such as the number of passengers boarding and alighting at each bus stop, and the level of congestion experienced by passengers. This information is crucial for optimizing the transportation system. In this study, we propose a method that uses BLE signal information to estimate not only the number of passengers on the bus, but also the number of passengers and the level of congestion at each bus stop. Experimental results show that with the proposed system, the accuracy based on the percentage of correct answers is 0.614 (for the number of people on the bus) and 0.538 (for the perceived level of congestion), while the Mean Absolute Error (MAE) is 1.203 (for the number of boarding passengers) and 1.357 (for the number of alighting passengers). language of the presentation: Japanese | ||||||
| HARLAND FITRIADI AMIN | M, 2回目発表 | サイバーレジリエンス構成学 | 門林 雄基 | 岡田 実 | 妙中 雄三 | |
|
title: Increasing LoRa-like Physical Layer Capacity by
Introducing Orthogonal Code
abstract: Currently, LoRa suffers from scalability issues, making it still far from a feasible option for LPWAN in large deployment scenarios. During Signal Transmission, no other signal employing the same Spreading Factor should be transmitted simultaneously, as this condition would render both signals unrecoverable. We propose introducing orthogonal codes to the LoRa Physical layer. We demonstrate that the introduction of orthogonal codes significantly enhances the signal recovery rate, achieving a Detection Error Rate (DER) of 99.5% for singledevice scenarios and maintaining robust performance with an 85.33% DER for two devices per SF. language of the presentation: English | ||||||
| 酒井 裕基 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 門林 雄基 | 諏訪 博彦 | |
|
title: Proposal for Utilizing Digital Signage in Distributed Federated Learning
abstract: With the increasing adoption of AI-based services, there is a growing demand for deep learning methods that protect the privacy of participants while enabling low-cost, high-accuracy model construction driven primarily by mobile devices. In this study, we propose an improved distributed federated learning method for building object recognition models in tourist areas. This method incorporates communication not only between mobile devices but also with digital signage that has the capability to build and store high-precision model parameters. Using a model based on MobileNetV2, one of the representative neural networks for mobile devices, the CIFAR-10 image dataset, and mobile device trajectory data from a specific tourist area in Nara City, we conducted a one-day model update simulation. The results showed that the proposed method improved the average accuracy of user models built on mobile devices by approximately 5.3% compared to traditional methods that only use device-to-device communication. language of the presentation: Japanese | ||||||
| 中岡 明義 | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | 渡辺 太郎 | 若宮 翔子 | 矢田 竣太郎 |
|
title: Dataset construction and application for behavior classification
abstract: Understanding the major behaviors in people's daily lives is an important issue in many fields such as sociology and economics. In this study, we defined behaviors in daily life based on 20 types of behaviors used in the Basic Survey of Social Life conducted by the Statistics Bureau of the Ministry of Internal Affairs and Communications. These behaviors were annotated to the “LIFE STORY” dataset collected by crowdsourcing, and a publicly available corpus was constructed. We then built a behavior classification model using the corpus, and evaluated the corpus based on the model's classification performance. In addition, to improve the quality of the corpus, data expansion using ChatGPT-4o was performed. Experimental results showed that the models trained on the expanded corpus performed better than the models trained before the expansion. Finally, we conducted a case study on the behavioral changes before and after Covid-19 using the constructed behavioral classification model. language of the presentation: Japanese | ||||||
| 西岡 竜生 | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | 渡辺 太郎 | 若宮 翔子 | 矢田 竣太郎 |
|
title: The Impact of Interpersonal Skills on Human-AI Collaboration: A Study with ChatGPT
abstract: Collaboration between humans and artificial intelligence (AI) has attracted attention for its potential to outperform AI alone. Recently, the advent of accessible large language models has accelerated research into effective collaboration between humans with applications in fields such as education. These research particularly focus on improving the performance of AI systems involved in such collaborations. However, evaluating LLMs that collaborate with humans requires not only performance when working with humans but also human capabilities. For instance, some users are prone to believing erroneous outputs from LLMs, while others can generate unique solutions through active discussions with LLMs. In the approach to learning in which people work together, called cooperative or collaborative learning, interpersonal skills are considered one of the factors that make learning be more productive than competitive or individualistic methods. Can the same be said of human-AI collaboration? This study analyzes whether interpersonal skills and skills to interact with AI are correlated. Specifically, we designed five tasks to work on using ChatGPT and analyzed the relation between users' communication skills and their task performance using ChatGPT. The results showed that in some cases, people with high interpersonal skills were better at using AI, while in other cases, people with low interpersonal skills were better at using AI. language of the presentation: Japanese | ||||||
| 林 純子 | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | 渡辺 太郎 | 若宮 翔子 | 矢田 竣太郎 |
|
title: Annotating autonomy to episodes abstract: It is known that autonomy is related to depression and well-being, and it has been suggested that measuring autonomy from free writing episodes such as diaries and social media posts can be useful for detecting depressive tendencies and understanding well-being. In the past, autonomy has been simply measured based on the number of times the passive voice is used, but there is a problem that it is impossible to evaluate the degree to which the actions expressed in the sentences are performed with autonomy. In this study, we define autonomy as “controlling one's own actions with one's own will,” and annotate the level of autonomy of the owner of the actions to LIFE STORY, which is a collection of memorable episodes recalled from emotions. If we can construct a resource for estimating differences such as “ate hospital food” (low proactivity) and “made and ate a hamburger steak” (high proactivity), qualitative analysis of features that contribute to proactivity will become possible, and this will lead to a deeper understanding of well-being. language of the presentation: Japanese | ||||||
| HANNACHI SAMAR | D, 中間発表 | ソーシャル・コンピューティング | 荒牧 英治 | 渡辺 太郎 | 若宮 翔子 | 矢田 竣太郎 |
|
title: Exploring Trust in AI in Healthcare
abstract: The integration of Artificial Intelligence (AI) into the healthcare system has the ability to revolutionize patient care and reduce healthcare providers’ workload. This would not be possible without having the component of trust in such technologies from both the patients and the healthcare providers. In this research we try to investigate the factors that influence trust in AI within healthcare environments. Collecting high-quality data was made possible through the launching of a survey into a crowdsourcing platform. The analysis of the data allows the answering of major questions related to the evolution of trust in AI and the importance of cultural and societal factors into such a concept. language of the presentation: English | ||||||
| 北岡 哲哉 | D, 中間発表 | ソフトウェア工学 | 松本 健一 | 中島 康彦 | 石尾 隆 | Raula Gaikovina Kula | 嶋利 一真 |
|
title: Evaluation of the Reliability of Code Obfuscation Methods When Combined With Code Optimization
abstract: Code obfuscation is a software protection technique against malicious analysis and tampering attacks. Since code obfuscation modifies the representation and structure of code in complex ways, many obfuscating transformations cause performance overhead, such as a decrease in execution speed. To improve execution speed while obfuscating code, code optimization is sometimes combined with code obfuscation. However, it is reported that applying optimization to obfuscated code may break the original program behavior. In this research, the behavioral changes factor of obfuscated programs by code optimization is investigated by analyzing the logic and coverage information before and after optimization. The current results show that most of the obfuscated programs preserved the same behavior before and after optimization, while some obfuscated programs were broken by optimization. Also, a comparison of coverage information before and after optimization revealed that optimization removes some of redundant code fragments introduced by obfuscation. In this presentation, I report the results of this current research and discuss how to investigate the factors that change program behavior in more detail. language of the presentation: Japanese 発表題目: 最適化を併用した場合におけるコード難読化手法の信頼性の評価 発表概要: コード難読化は，コードの表現と構造を複雑な方法で変更することで，悪意のある解析や改ざん攻撃の成功を遅延させるソフトウェア保護技術である．コード難読化は実行速度の低下などのパフォーマンスのオーバヘッドを引き起こすことから，実行速度を向上させるために難読化されたコードに最適化が用いられることがある．ただし，難読化されたコードに最適化を適用すると，元のプログラムの動作が壊れ，「信頼できる」プログラムではなくなる可能性の存在が報告されている．本研究では，コード難読化と最適化を併用したときの信頼性を評価することを目的として，最適化前後でのロジックやカバレッジ情報を解析することで，プログラムの挙動が変化する要因を調査する．現時点の調査では，ほとんどの難読化プログラムは最適化前後で同じ振る舞いを維持する一方で，最適化によって壊れる難読化プログラムもあることがわかった．また，最適化前後のカバレッジ情報を比較した結果，最適化により，難読化によって導入された冗長なコード断片の一部が削除されることが明らかになった．本発表では，この調査結果を報告し，より詳細にプログラムの挙動が変化する要因を調査する方法について議論を行う． | |||||||