| 垣内 龍 | M, 1回目発表 | 生体医用画像 | 佐藤 嘉伸, 向川 康博, 大竹 義人, SOUFI Mazen |
|
title: Quantitative Evaluation of Rehabilitation by Estimating Muscle Movement and Joint Load in Patients with Artificial Hip Implantation Using AnyBody
abstract: Hip replacement is the most frequent surgery performed for diseases such as osteoarthritis of the hip joint, osteonecrosis of the femoral head, and hip joint lesions of rheumatoid arthritis that do not function adequately. While appropriate evaluation of postoperative rehabilitation greatly contributes to improving patients' quality of life, quantitative evaluation has not been performed. This study aims to introduce a system to quantitatively evaluate patients' postoperative rehabilitation. We will present an overview of the project, our efforts to date and future plans. language of the presentation: Japanese 発表題目: AnyBodyを用いた人工股関節埋入患者の筋運動量・関節負荷推定によるリハビリテーション定量評価 発表概要:人工股関節置換術は変形性股関節症や大腿骨頭壊死症、関節リウマチの股関節病変など股関節が十分に機能しない疾患に対して行われる最も頻度の高い手術である。手術後のリハビリテーションの適切な評価は患者のQOL向上に大きく寄与する一方、定量的評価は行われていない。本研究は患者の術後リハビリテーションを定量的に評価するシステムの導入を狙うものである。プロジェクトの全体像およびこれまでの取り組み、今後の方針について発表する。 | |||
| 上谷 明日香 | M, 1回目発表 | 光メディアインタフェース | 向川 康博, 清川 清, 舩冨 卓哉, 藤村 友貴, 北野 和哉 |
|
Title: Visualization of Overwritten Text Using Hyperspectral Imaging and K-means
Abstract: Draft studies are important for clarifying the process of authors' thoughts and the generation of their writings. This study aims to visualize overwritten characters in the drafts of modern writers. As a preliminary experiment,we made hyperspectral imaging of modern writers' drafts using a microscope. In this presentation, we show the results of visualizing the overwritten characters by classifying the spectral distribution of the drafts by K-means. language of the presentation: Japanese 発表題目:分光計測とK-meansを用いた塗りつぶされた文字の可視化 発表概要:草稿研究は、作家の思考や文章の生成過程を明らかにするうえで重要である。本研究では、近代作家の草稿において、塗りつぶされた文字を可視化することを目的としている。予備実験として、顕微鏡を用いて近代作家の草稿を分光計測した。本発表では草稿の分光分布をK-meansによって分類することで塗りつぶされた文字を可視化した結果を示す。 | |||
| 村本 幸次郎 | M, 1回目発表 | 光メディアインタフェース | 向川 康博, 清川 清, 舩冨 卓哉, 藤村 友貴, 北野 和哉 |
|
title: Measurement and Modeling of Optical Properties of Staked Glass Panels
abstract: The purpose of this study is to separate the visibility of glass dry plates stuck by water damage such as tsunami and floods by nondestructive means. This study aims to use an image scanner to separate the visibility of dry glass panels. In this presentation, we will show the results of analyzing the light propagation characteristics of scanned glass plates. Based on these characteristics, a simulation based on a model of a sticking glass plate will be performed. language of the presentation:Japanese 発表題目: 張り付いたガラス乾板の光学特性の計測とモデル化 発表概要: 津波や洪水などの水害によって張り付いたガラス乾板の、非破壊による見えの分離を目的としている。本研究ではイメージスキャナを使ってガラス乾板の見えを分離することを目指す。この発表では、ガラス乾板のスキャンにおける光伝搬特性の分析結果を示す。またこの特性を基に、張り付いたガラス乾板のモデルに基づいて見え方をシミュレーションする。 | |||
| 松浦 光希 | M, 1回目発表 | 光メディアインタフェース | 向川 康博, 清川 清, 舩冨 卓哉, 藤村 友貴, 北野 和哉 |
|
title: Emphasis rendering of scratches on metal cutting surface
abstract: Unintentional scratches may occur on metal surfaces when metal is cut using NC machines. In this study, computer graphics is used to render the scratches on the metal cutting surface. As a proposed method, a method based on ray tracing algorithms was devised and implemented. Specifically, the method is a rendering method that emphasizes the case where a ray emitted from the camera is reflected in the specular reflection direction on the object surface in ray tracing, and the reflected ray further impacts the object. As a result, the overall uneven edges of the object were emphasized instead of the minute scratches, and the scratches could not be emphasized well. However, we believe that we were able to propose a method that provides a clue for emphasizing scratches, rather than focusing on photorealistic rendering. language of the presentation: Japanese 発表題目: 金属切削面の傷の強調レンダリング 発表概要: NC加工機を用いて金属の切削をした際，金属表面に意図せぬ傷が発生することがある．本研究では，金属切削面における傷を強調したレンダリングをコンピュータグラフィックスを用いて行う．提案手法として，レイトレーシングのアルゴリズムを基とした手法を考案し，実施した．具体的な手法は，レイトレーシングにおいてカメラから放たれたレイが物体表面で鏡面反射方向に反射し，その反射したレイが更に物体に衝突した場合を強調するようなレンダリング法である．結果としては，微細な傷ではなく物体の全体的な凹凸のエッジが強調され，傷をうまく強調することはできなかった．しかし，写実的なレンダリングにこだわるのではなく，傷の強調のための手がかりとなる手法を提案できたと考えており，今後は傷のないCADモデルを活用した傷の強調手法を探りたいと考えている． | |||
| HERNANDEZ RODRIGUEZ DIEGO | M, 1回目発表 | 光メディアインタフェース | 向川 康博☆, 川西 康友, 薗頭 元春, 舩冨 卓哉, 藤村 友貴, 北野 和哉 |
|
title: Event Camera Simulation from Pre-trained Dynamic Neural Radiance Fields
abstract: Event cameras are sensors which asynchronously measure light intensity changes and feature low-latency and high-dynamic range. Since these sensors are still very expensive, we aim to simulate their imaging capabilities. Current event camera simulation approaches lack the ability to simulate dynamic scenes from arbitrary viewing angles. In this study we propose a novel approach to simulating event cameras by leveraging the capabilities of a dynamic neural radiance field. language of the presentation: English | |||
| 中畔 彪雅 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, 吉野 幸一郎（客員教授）, 河野 誠也（客員助教） |
|
title: End-to-End Voice Dialogue System
abstract: Existing spoken dialogue systems are realized by using modules for speech recognition, dialogue generation, and speech synthesis in a cascade. However, such cascade processing makes natural dialogue difficult due to problems such as processing delays and text-based processing, which results in the loss of information that cannot be converted into text. Therefore, we aim to solve these problems by developing a system that generates responses directly from voice input without using text. language of the presentation: Japanese 発表題目: End-to-End音声対話システム 発表概要: 既存の音声対話システムは音声認識，対話生成，音声合成のモジュールをカスケードに用いることで，実現される．しかし，こうしたカスケードな処理では，処理の遅延やテキストによる処理のため，テキスト化できない情報は欠落するなどの問題から，自然な対話が困難になる．そこで，テキストを介さず，音声入力から直接応答を生成するシステムを開発することで，これら問題の解決を目指す． | |||
| 三輪 拓真 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, 吉野 幸一郎（客員教授）, 河野 誠也（客員助教） |
|
title: Proposal for a Language Model Using Quantum Computation
abstract: Large-scale language models are treated as the foundation for natural language processing, and it is important to increase the number of parameters while maintaining computational efficiency. One technique that contributes to improving computational efficiency is quantum computation, which maintains numerical values as quantum states. This study examines the use of quantum transformer with quantum gates for language model tasks to improve computational efficiency and accuracy. language of the presentation: Japanese | |||
| 日浦 隆博 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, 吉野 幸一郎（客員教授）, 河野 誠也（客員助教） |
|
title: Building a hypothesis generation and dialogue system to support research
abstract: In recent years, Large Language Models(LLMs) trained on large datasets have played an important roll in various natural language processing tasks. ChatGPT, a dialogue AI that utilizes LLM, has been used in a wide range of situations, including information gathering and educational support. Many graduate students and researchers in particular are likely to use it as a tool to support their own research. However, since LLM is a probabilistic model that learns from training datasets, it can be expected to be effective for tasks such as information gathering, but its ability to propose new research ideas and hypotheses that do not exist in the training datasets is limited. Therefore, I consider building a model that differs from LLMs and specializes in generating hypotheses through interacting with user. language of the presentation: Japanese 発表題目: 研究をサポートするための仮説生成対話システムの構築 発表概要: 近年、大規模なデータセットで学習された大規模言語モデル(LLM)が様々な自然言語処理タスクで重要な役割を果たしている。LLMを活用した対話型AIであるChatGPTは情報収集、教育的支援など幅広い場面で使用されており、特に大学院生や研究者は、自分の研究をサポートするツールとして使用している人も多いと思われる。しかしLLMは学習データからパターンを学習した確率モデルであるため、情報収集などでは効果が期待できるが、学習データには無い新規の研究アイデアを提案する能力には限界がある。そこでLLMとは別に、ユーザーとの対話を通じて仮説生成を行うことに特化したモデルを構築することを考える。 | |||
| 林 和樹 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, 池田 和司, 上垣外 英剛 |
|
title: Artwork Explanation in Vision Large Language Models
abstract: Vision Large Language Models (VLLM) are large language models that generate text based on images and instructions input by users. When utilizing these models for image creation support, they are often required to generate explanations based on deep artistic knowledge, including composition, innovation, comparison with other works, historical context, and the artist's intentions. However, the extent of VLLM's capability in these knowledge areas has not been quantitatively clarified. In this study, we provide an evaluation dataset and metrics to assess how well VLLMs can describe artworks based on artistic knowledge. Furthermore, we also aim to construct a training dataset to expand the artistic knowledge held by VLLMs. language of the presentation: Japanese | |||
| CAO ZHE | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, 池田 和司, 上垣外 英剛 |
|
title: Low Rank Language-specific Adapters in Multilingual Neural Machine Translation
abstract: Parameter Efficient Fine-tuning (PEFT) plays an important role based on the current model scale. Recent research about LoRA (Low Rank Adaptation) shows that through fine-tuning in a much smaller, low-rank subspace, we can still achieve performance comparable to full parameter fine-tuning. However, there is still a lack of research on how to apply LoRA in multilingual neural machine translation (MNMT). This study attempts to propose a more efficient LoRA approach through architecture learning, and tries to further reduce the size of language-specific parameters by introducing language family information into LoRA. language of the presentation: English | |||
| 鈴木 刀磨 | M, 1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: Investigation of Negation Learning in Instruction Tuning of Language Models Abstract: Negation plays a crucial role in altering the truth value of propositions within sentences, holding significant meaning in language comprehension. However, text data from the internet, used for training language models, often contains a multitude of affirmative expressions, highlighting insufficient training of language models in handling negation. In this study, with the aim of improving the task performance of the model through Instruction tuning, we explore how language models learn and process negation commands such as "do not." We comprehensively examine which aspects of negation the language model effectively learns and identify challenges it encounters. Specifically, we provide two types of instructions: regular commands and instructions containing numerous negations, analyzing the differences in the model's internal representation and output. Through this analysis, our goal is to unveil the mechanism by which language models effectively learn negation expressions. language of the presentation: Japanese | |||