コロキアムB発表

日時: 07月24日 (Thu) 2限目(11:00-12:30)


会場: L1

司会: 鍛治秀伍
KIM DOHYUN D, 中間発表 コンピューティング・アーキテクチャ 中島 康彦, 林 優一, 張 任遠, KAN Yirong, PHAM HOAI LUAN, Le Vu Trung Duong
title: Faster nearest neighbor search through more efficient memory allocation
abstract: In vector databases, nearest neighbor search is an essential operation to find vectors that are close to the query. However, the vector operation itself is computationally expensive, and loading the vectors to be searched is time-consuming. While research on speeding up the vector operation itself has progressed due to the increased demand for operations that handle matrix operations, there has been little research at the hardware level that focuses on improving the efficiency of loading the vectors to be searched. In this study, using IMAX3, we propose and verify a method that can dramatically reduce operation time during the first and second runs by storing the vectors to be searched in the local memory of the IMAX3 computation unit in advance.
language of the presentation: Japanese
発表題目: メモリ配置による最近傍探索のベクトル間計算の効率化
発表概要: ベクトルデータベースにおいて、最近傍探索はクエリに近いベクトルを探すためには必要不可欠な演算となる。しかし、ベクトル演算そのものも計算コストが高く、検索対象のベクトルをロードすることにも時間を要する。ベクトル演算そのものの高速化は、行列演算を扱う演算の需要が高まったことから研究が進んでいるが、探索対象のベクトルのロードの効率化に重きを置くハードウェアレベルでの研究は少なかった。本研究では、IMAX3を用いて、探索対象のベクトルを予めIMAX3の演算ユニットのローカルメモリに格納して置くことにより、初回実行時と二回目以降実行時の演算時間を画期的に削減できる手法を提案し、検証を行なっている。
 
村田 百合菜 M, 2回目発表 サイバーレジリエンス構成学 門林 雄基, 林 優一, 妙中 雄三
title: Distinguishing Natural and Adversarial Concept Drifts in Edge-Learning IoT Devices
abstract: With the advancement of smart cities, IoT devices equipped with edge AI are increasingly being deployed. Some of these devices are capable not only of performing inference but also of retraining models on the edge. While this enables adaptation to concept drift, it also introduces the risk that adversarial drift attacks—where malicious data induces undesired model updates—may steer the model in unintended directions. To address this issue, this study proposes a two-stage detection method using correlation coefficients and entropy to distinguish between natural and adversarial drifts before retraining. This enables the exclusion of adversarial data and facilitates appropriate and secure retraining on the edge.
language of the presentation: Japanese
発表題目: エッジで再学習を行うIoTにおける自然ドリフトと敵対的ドリフトの判別
発表概要: 現在、スマートシティの推進に伴い、エッジAIを搭載したIoTデバイスの導入が進んでいる。これらのデバイスの中には、エッジ側で推論だけでなく再学習を行うものも存在する。これは再学習によりコンセプトドリフトに適応できる一方、敵対的ドリフト攻撃(悪意あるドリフトをさせるデータの注入)を受けると、攻撃者の意図した方向にモデルが再学習されるリスクがある。そこで本研究では、再学習の前に敵対的ドリフトデータと自然ドリフトデータを、相関係数とエントロピーの2段階で判別する手法を提案する。これにより、敵対的データを除外し、適切な再学習を可能にすることを目指す。
 
荒木 亮介 D, 中間発表 サイバーレジリエンス構成学 門林 雄基, 笠原 正治, 林 優一, 妙中 雄三
title: Dynamic DDoS Threat Scoring Incorporates Periodic Trends and Social Attention
abstract: DDoS (Destributed Denial of Service) is still a serious threat to the network system. With the development and spread of high computational resources and IoT devices, it is becoming more serious. To address these problems, researchers are working on improving the attack traffic detection system, also the CSP (Cloud Service Provider) gives effort to resource provided the CDN (Content Delivery Networks) to manipulate the content from DDoS attack. For efficient mitigation, reactive resource management is developed as autoscaling. However, they are sometimes fail the service if the resource for one provider could bother other customers. Therefore, if we forcast the DDoS attack in advance, it will be useful for resource management for DDoD mitigation. To realize that, in this paper we focus on periodic trends of DDoS attack and social attention to the content provider or services. We utilized the Tranformer model to determine the priority for protection. Tranformer learns the periodic trends and from the past attack traffic and also learns latent social attention, which could lead to DDoS attack in the future.
language of the presentation: English
 

日時: 07月24日 (Thu) 2限目(11:00-12:30)


会場: L2

司会: 日永田智絵
古賀 荘翠 M, 2回目発表 光メディアインタフェース 向川 康博, 安本 慶一, 舩冨 卓哉, 藤村 友貴, 北野 和哉
タイトル:AniDepth: 深度画像に基づく線画変形を用いた条件付き拡散モデルによるアニメ中割手法 発表概要:アニメ制作の中割り作業は高い専門性と多大な作業コストを要するため,その自動化が望まれている.本稿では,動画拡散モデルを用いたアニメの中割りを支援する補間フレーム生成手法"AniDepth"を提案する.AniDepthでは,彩色後のキーフレーム補間に線画を用いることで実際の制作工程に近い条件設定で補間結果の品質の向上を行う.キーフレームを深度画像に変換し,アニメと実写のドメインギャップを埋めるモダリティに変換することで,動画拡散モデルの事前知識の活用を図る.実験により,事前学習済みの動画拡散モデルが明示的な深度画像による学習を介さずとも補間を実現し,定量的指標で既存手法を上回り視覚的破綻の低減が得られることを確認した. Title: AniDepth : Anime In-between Diffusion using Depth-guided Warped Line-art abstract: The task of dividing anime frames into smaller frames is highly specialized and expensive, and its automation is desired. This paper proposes “AniDepth,” a method for generating in-betwen to assist in animation production using a video diffusion model. The keyframes are converted to depth images, a modality that bridges the domain gap between anime and live action, thereby taking advantage of prior knowledge of the video diffusion model. Experimental results show that the pre-trained video diffusion model achieves interpolation without explicit learning through depth images, and that it outperforms existing methods in terms of quantitative measures to reduce visual breakdown. language of the presentation: Japanese
 
ALNAJJAR MOHAMAD M, 1回目発表 ソーシャル・コンピューティング 荒牧 英治, 安本 慶一, 若宮 翔子, Peng Shaowen
title: Improving Demographic-Specific Communication Effectiveness on Social Media Using Large Language Models
abstract: Social media platforms such as Twitter and Facebook offer powerful channels for governments and organizations to share important public information that can support people in their daily lives. These platforms are especially valuable during times of crisis or danger—such as public health emergencies—when rapid and clear communication is essential. However, such posts often fail to reach or engage all segments of society. Differences in age, educational background, and other demographic factors influence how individuals perceive and interact with these messages. Some users may skip the posts entirely, while others may read them but struggle to understand the content. In this research, we aim to address this communication gap by identifying the underlying reasons why certain groups do not engage with medical content on social media. We then leverage large language models (LLMs) to adapt and rewrite these posts, improving their clarity, accessibility, and relevance. Our goal is to enhance the directivity of social media communication—ensuring that critical information reaches and resonates with diverse audiences across society.
language of the presentation: English
 
ROJPAISARNKIT RUKSIT D, 中間発表 ソフトウェア工学 松本 健一, 安本 慶一, Raula Gaikovina Kula, 嶋利 一真, Fan Youmei
title: Towards Proficiency Assessment through Code
abstract: In an era increasingly shaped by generative AI, code understanding remains a critical foundation for ensuring software quality, reliability, and maintainability. While AI systems can accelerate code generation, developers still face substantial challenges in comprehending, debugging, and effectively integrating the resulting artifacts. Code proficiency, defined not only by comprehension but also by the ability to write efficient, idiomatic code, plays a central role in addressing these challenges. Existing tools, such as those assigning CEFR-based levels to code constructs, offer initial frameworks for assessing code difficulty. However, their manually derived classifications lack empirical validation and often diverge from pedagogical progressions found in computer science textbooks. This research seeks to address the absence of a standardized, data-driven metric for determining the proficiency levels required to understand specific programming elements. We propose an automated framework grounded in textbook analysis and clustering techniques to establish a scalable proficiency metric applicable to both human-written and AI-generated code. Preliminary findings reveal strong alignment with educational sequencing and suggest promising applications in AI-assisted software development, particularly in enhancing code review workflows and tailoring AI-generated code to developers’ proficiency levels.
language of the presentation: English
 

日時: 07月24日 (Thu) 2限目(11:00-12:30)


会場: L3

司会: 松井智一
RIERA MACHIN MARIA ANGELICA M, 2回目発表 自然言語処理学 渡辺 太郎, 荒牧 英治, 上垣外 英剛
title: Using the CEFR to Guide LLMs in Simplifying Spanish.
abstract: Lexical simplification involves modifying complex or difficult vocabulary and structures with simpler alternatives to make text more accessible while preserving its original meaning. This task is essential for several groups of people, including language learners. Large language models (LLMs) can play a significant role in this process due to their ability for generating simpler versions of text. Focusing on using the Common European Framework of Reference for Languages (CEFR) to guide LLMs in simplifying Spanish, a language with substantial resources but fewer compared to English, this study aims to align LLM text outputs with CEFR proficiency levels to improve Spanish language accessibility and support learners at different stages of language proficiency.
language of the presentation: English
 
清水 聖司 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, PENG SHAOWEN
title: RecordTwin: Towards Creating Safe Synthetic Clinical Corpora
abstract: The scarcity of publicly available clinical corpora hinders developing and applying NLP tools in clinical research. While existing work tackles this issue by utilizing generative models to create high-quality synthetic corpora, their methods require learning from the original in-hospital clinical documents, turning them unfeasible in practice. To address this problem, we introduce RecordTwin, a novel synthetic corpus creation method designed to generate synthetic documents from anonymized clinical entities. In this method, we first extract and anonymize entities from in-hospital documents to ensure the information contained in the synthetic corpus is restricted. Then, we use a large language model to fill the context between anonymized entities. To do so, we use a small, privacy-preserving subset of the original documents to mimic their formatting and writing style. This approach only requires anonymized entities and a small subset of original documents in the generation process, making it more feasible in practice. To evaluate the synthetic corpus created with our method, we conduct a proof-of-concept study using a publicly available clinical database. Our results demonstrate that the synthetic corpus has a utility comparable to the original data and a safety advantage over baselines, highlighting the potential of RecordTwin for privacy-preserving synthetic corpus creation.
language of the presentation: Japanese