| 宮本 佳子 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦 | 松本 健一 | 小野 直亮 | MD.Altaf-Ul-Amin |
|
title: Relationship between Influenza A Virus Inactivation and Lipid Bilayer Properties Induced by Fragrance Compound Combinations
abstract: Essential oils and their constituents have been reported to possess antiviral ability. Previous studies have indicated that the inactivation ability of influenza A virus was synergistically enhanced by appropriately combining specific fragrance compounds. However, the detailed mechanisms of the synergistic effect remain unclear. This study aims to elucidate the phenomena behind the synergistic effect, with the goal of leveraging the synergism to obtain compositions with higher antiviral ability. In this research, the focus was placed on changes in the physical properties of lipid bilayer membrane as the phenomenon underlying the synergistic effect. Specifically, the plan is to investigate (1) the relationship between antiviral activity and membrane properties, and investigations have already been conducted on (2) the relationship between membrane properties and compound structure, and (3) the interrelationship among various indices of membrane properties. In this presentation, we report on our examination of (2) and (3) using canonical correlation analysis. language of the presentation: Japanese 発表題目: 化合物組み合わせにおける抗ウイルス活性と脂質二重膜物性の関係 発表概要: エッセンシャルオイルとその構成成分はウイルス不活化能を有することが知られている。また、個々には不活化能が低い複数の化合物についても、適切に組み合わせることでインフルエンザA型ウイルスの不活化能が相乗的に向上することが報告されている。ウイルスの構成要素（脂質二重膜、スパイクたんぱく質、核酸）を破壊することでウイルスが不活化されることが知られているが、個々には不活化能の低い化合物を組み合わせると不活化能が向上する機構については不明な部分が多い。本検討では、抗ウイルス活性の高い組成物を得ることを目的として、相乗効果の背景にある現象の解明に取り組んでいる。本検討では、相乗効果の背景にある現象として、脂質二重膜の物性変化に着目した。そこで、①抗ウイルス活性と膜物性の関係、②膜物性と化合物構造の関係、③膜物性の各指標同士の関係について検討している。本発表では、②③について正準相関分析を用いて検討したので報告する。 | ||||||
| HUANG BIN | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 | 松本 健一 | 小野 直亮 | MD.Altaf-Ul-Amin |
|
title:Machine Learning for Predicting Epileptic Seizures Using EEG Signals: A Gabor Function-Based Approach
abstract:Predicting epileptic seizures using EEG signals remains a critical challenge with the potential to significantly enhance the quality of life for individuals living with epilepsy. While machine learning has made significant strides in this field, many approaches still struggle with extracting features that fully capture the intricate patterns in EEG data. To address these limitations, we propose a new method that combines the strengths of Gabor functions and convolutional neural networks (CNNs) to deliver better results. In this study we incorporate adaptive Gabor kernels instead of manually designed filters with the neural networks. these kernels are learned during training, allowing them to adapt to the unique time-frequency characteristics of EEG signals. These adaptive kernels then serve as a foundation for extracting features that are highly relevant for seizure prediction. The CNN layers that follow further refine these features and handle classification. What’s particularly exciting is our ability to analyze the resulting predictions to identify which Gabor kernels contribute the most, offering insights into the features that truly matter. This method bridges the gap between traditional handcrafted techniques and purely data-driven models, offering the dual advantages of improved feature representation and higher prediction accuracy, as demonstrated in our experimental results. Additionally, the computational efficiency of dynamically generating Gabor kernels makes this approach practical for real-time applications—a critical factor in seizure prediction. Beyond the technical improvements, the potential impact on patients is profound. With more accurate and timely predictions, individuals living with epilepsy could take preventive actions to avoid seizures, reducing both their physical risks and emotional stress. Through this combination of technological innovation and a focus on patient-centered outcomes, our study represents a meaningful step toward making reliable and accessible seizure prediction a reality. language of the presentation: English | ||||||
| 稲葉 真幸 | M, 1回目発表 | ソフトウェア工学 | 松本 健一 | 飯田 元 | Raula Gaikovina Kula | 嶋利 一真 |
|
title: Towards Detecting Textually Dissimilar Duplicate Bug Reports using Transformer models
abstract: In large-scale bug management systems, hundreds of bugs are reported daily, and it is not uncommon for multiple individuals to report the same bug, resulting in duplicate bug reports. Manually classifying duplicates places a burden on developers, prompting research into automated detection methods. However, since bug reports are written in natural language, differences in subjective interpretation, perceptions of the issue, and circumstances surrounding the bug often lead to variations in the descriptions used in reports. In this study, such cases are referred to as textually dissimilar duplicate bug reports. Detecting textually dissimilar duplicate bug reports is challenging with traditional natural language processing (NLP) or machine learning methods, as it requires accurately understanding the semantics of each report's description. This study aims to identify effective methods and models for detecting textually dissimilar duplicate bug reports. An evaluation experiment was conducted using Transformer-based large language models and Sentence Transformers, which specialize in comparing textual similarity. The results showed that Sentence Transformer-based models achieved high precision in terms of both F-score and AUC for detecting both textually similar and textually dissimilar duplicate bug reports, demonstrating their utility in detecting duplicate bug reports. language of the presentation: Japanese 発表題目: テキスト表現が異なる重複バグ報告の検出におけるTransformerモデルの性能評価 発表概要: 大規模なバグ管理システムでは毎日何百ものバグが報告されるため，複数の人が同じバグについて報告を行ってしまうことがあり，重複したバグ報告が発生する．重複を手動で分類することは開発者にとって負担となるため，自動的に検出する手法が研究されてきた．しかし，バグ報告は自然言語で構成されるため，報告者の主観や認識内容の相違，発生状況によって，報告の説明文に異なる表現が用いられることがある．本研究ではそれらをテキスト表現が異なる重複バグ報告と呼ぶ．テキスト表現が異なる重複バグ報告は，それぞれのバグ報告の説明文が表す意味を正しく理解しないと重複と判断することができないため，従来の自然言語処理や機械学習の手法では検出が難しい場合がある．本研究では，テキスト表現が異なる重複バグ報告を検出するための効果的な方法とモデルを特定することを目的とし，Transformerベースの大規模言語モデルと，文章の類似性比較に特化したSentence Transformersを用いて検出性能の比較を行う評価実験を行った．実験の結果，テキスト表現が類似している重複バグ報告，類似していない重複バグ報告のいずれにおいてもF値とAUCにおいてSentence Transformerベースのモデルが高い精度を示し，重複バグ報告の検出において有用であることが明らかとなった． | ||||||
| 田中 英武 | M, 1回目発表 | ソフトウェア工学 | 松本 健一 | 飯田 元 | Raula Gaikovina Kula | 嶋利 一真 |
|
title: Investigation of a Highly Accurate Code Generation Method Using Property-based Testing with LLM
abstract: In code generation using large language models (LLMs), an approach that iteratively tests and modifies the generated code to improve accuracy has been adopted in recent years. In this approach, example-based testing is mainly used as a testing method, but conventional methods suffer from insufficient code coverage and high computational cost. To overcome these challenges, we propose a new framework based on property-based testing, which verifies general properties that a code must satisfy, rather than relying on input examples, allowing for wider code coverage and more efficient bug detection. In our framework, we build an iterative process in which LLM verifies the generated code using property-based testing code generated by LLM, and LLM improves the code based on the results. This approach aims to significantly improve the accuracy and reliability of code generation while reducing the number of inferences. language of the presentation: Japanese 発表題目: LLMを用いたProperty-based Testingによる高精度なコード生成手法の検討 発表概要: 大規模言語モデル（LLM）を用いたコード生成において，近年精度向上を目的として生成したコードのテストと修正を反復的に行うアプローチが採用されつつある．このアプローチにおいて，主にExample-based testingがテスト手法として用いられるが，従来の手法ではコードカバレッジの不十分さと高い計算コストが課題となっている．本研究では，これらの課題を克服するために，Property-based testingに基づく新たなフレームワークを提案する．Property-based testingは，入力例に依存するのではなく，コードが満たすべき一般的な性質を検証する手法であり，より広範なコードカバレッジと効率的なバグ検出が可能となる．本研究のフレームワークでは，LLMが生成したProperty-based testingのコードを用いて生成コードを検証し，その結果を基にLLMがコードを改良する反復プロセスを構築する．この手法により，推論回数を抑えながらも，コード生成の精度と信頼性を大幅に向上させることを目指す． | ||||||
| 河瀬 浩毅 | M, 1回目発表 | ネットワークシステム学 | 岡田 実 | 林 優一 | 東野 武史 | |
|
title: Evaluation of allowable loss of radio section in A-RoF relay of LPWA signals
abstract: Although the multi-hop relay is a promising method for increasing radio coverage to reduce the radio dead zone of the wireless access of the Internet-of-Things (IoT) devices, a radio air interface is limited to being homogeneous systems in low power wide area (LPWA). Analog radio-on-fiber (A-RoF) network enables us to transmit heterogeneous radio signals and relay them, however, the link has constraints on the noise and non-linearity came from analog components in the network. This paper proposes a link budget design of a non-regenerative relay over an A-RoF. The paper shows several plans for deploying radio amplifiers and RF filters with considering gain and noise figures while maintaining required remote-cell coverage. language of the presentation: Japanese | ||||||
| 小山 穂菜美 | M, 1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 | 笠原 正治 | 諏訪 博彦 | |
|
title: Proposal of a Behavior Change Navigation System Considering Congestion and Satisfaction in Tourist Spots.
abstract:The concentration of congestion in tourist destinations significantly impacts the living environment of local residents and the satisfaction of tourists. This study proposes a behavior change navigation system that addresses this issue by considering both congestion and satisfaction. The system aims to achieve congestion dispersion and a comfortable tourist experience by presenting options for destinations and transportation methods while flexibly guiding tourists’ behavior. To facilitate behavior change, the system is designed to solve a multi-objective optimization problem, enabling tourists to choose actions that contribute to alleviating congestion while maintaining their satisfaction. Additionally, the system provides predicted satisfaction levels based on congestion and user preferences, creating a framework where recommended options are naturally accepted by users. At this stage, we are formulating a problem framework with congestion and satisfaction as objective functions and considering the development of an optimization framework based on this formulation. In this presentation, we will discuss the details of the proposed problem formulation, the basic design for facilitating behavior change, the significance of the research, and plans for future simulations and algorithm development. language of the presentation: Japanese 発表題目: 観光地の混雑度と満足度を考慮した行動変容ナビゲーションシステムの提案 発表概要: 観光地における混雑の集中は、地域住民の生活環境や観光客の満足度に深刻な影響を及ぼしている。本研究では、この課題を解決するため、混雑度と満足度の両側面を考慮した行動変容ナビゲーションシステムを提案する。本システムは、観光地や移動手段の選択肢を提示し、観光客の行動を柔軟に誘導することで、混雑の分散と快適な観光体験を実現することを目的とする。 行動変容の実現においては、多目的最適化問題を解き、各観光客が満足度を維持しながら混雑緩和に寄与する行動を選択できるよう設計する。また、混雑度やユーザの嗜好をもとに予想満足度を提示することで、推奨される選択肢をユーザーが自然に受け入れられる仕組みを構築する。 現段階では、混雑度と満足度を目的関数とした問題設定を行い、これに基づく最適化のフレームワークの構築を検討している。本発表では、提案する問題設定の詳細、行動変容を促すための基本設計、研究の意義、および今後のシミュレーションやアルゴリズム開発の計画について議論する。 | ||||||
| 長山 和樹 | M, 1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 | 笠原 正治 | 諏訪 博彦 | |
|
title: Semi-Automatic Detection and Evaluation of Tourist Spots Using Cyclists' Riding Log Data Analysis
abstract: In cycle tourism, where cyclists can freely adjust their pace while experiencing local attractions, it is presumed that there exist locations particularly favored by cyclists (Cyclist Points of Interest: CPoI). However, these CPoIs are often known only within cyclist communities and are not widely shared as general tourist information. This study proposes a semi-automatic method for detecting and evaluating CPoIs from cyclists' riding log data. To verify the effectiveness of the proposed method, detailed riding data was collected from 374 cyclists at the Michi-no-Eki Maebashi Akagi roadside station in Maebashi City, Gunma Prefecture. By applying stop position estimation and hierarchical clustering to the collected data, 125 CPoIs were identified. Subsequently, crowdsourcing was utilized for CPoI identification and evaluation, resulting in 109 locations (87.2%) matching the author's expectations, with positive correlations observed between evaluation results. Furthermore, through analyzing the relationship between CPoI evaluations and visitor numbers, the study demonstrated the potential to distinguish between widely recognized popular spots and lesser-known but highly rated hidden gems. These results indicate that the proposed method is effective for detecting and evaluating CPoIs. language of the presentation: Japanese 発表題目: サイクリストによる走行ログデータ解析を用いた観光スポットの半自動検出・評価手法の検討 発表概要: サイクリストが自由にペースを調整しながら地域の魅力を体験できるサイクルツーリズムにおいて，サイクリストが好む場所（Cyclist Point of Interest: CPoI）の存在が推察される．しかし，これらのCPoIはサイクリスト間でのみ知られており，一般的な観光情報として広く共有されていないという課題がある．本研究では，サイクリストの走行ログデータからCPoIを半自動的に検出・評価する手法を提案する．提案手法の有効性を検証するため，群馬県前橋市の道の駅まえばし赤城において，374名のサイクリストから詳細な走行データを収集した．収集したデータに対して停車位置の推定と階層的クラスタリングを適用し，125個のCPoIを特定した．次に，クラウドソーシングを活用してCPoIの同定と評価を行った結果，109箇所（87.2%）が筆者の想定と一致し，評価結果間にも正の相関が確認された．さらに，CPoIの評価と訪問者数の関係分析により，広く認知された人気スポットと，認知度は低いが評価の高い穴場スポットを識別できる可能性を示した．これらの結果は，提案手法がCPoIの効果的な検出と評価に有効であることを示している． | ||||||
| 廣瀬 桃花 | M, 1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 | 笠原 正治 | 諏訪 博彦 | |
|
title: Optimization of School District Reorganization with Extended Facility Location Models
abstract: School district reorganization resembles facility location problems aimed at optimizing resource allocation but includes unique constraints such as population distribution, commuting distances, and educational equity. This study proposes an extended facility location model designed to address the specific requirements of school district reorganization. The proposed approach aims to improve computational efficiency and provide practical solutions. Additionally, validation with real-world municipal data will be conducted to enable its practical use as a decision support tool. language of the presentation: Japanese 発表題目: 施設配置モデルによる校区再編の最適化 発表概要: 校区再編問題は、リソースの最適配置を目的とする施設配置問題に類似するが、児童数の分布、通学距離、教育的公平性といった特有の制約を含む複雑な問題である。本研究では、施設配置モデルを拡張し、校区再編の特有要件に対応した新たな最適化手法を提案する。この手法により計算効率を高めつつ、現実的な解を導出することを目指す。さらに、地方自治体の実データを用いて検証を行い、意思決定支援ツールとしての実用化に向けた取り組みを行う。 | ||||||
| 松本 一晟 | M, 1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 | 荒牧 英治 | 諏訪 博彦 | |
|
title: Enhancing News Understanding via Multi-Agent News Chatbot abstract: The rise of recommendation systems has increased the influence of filter bubbles and echo chambers, causing users to develop one-sided perspectives on certain news events. Addressing this challenge requires exposing users to a variety of viewpoints through news articles from differing perspectives. This system enables users to navigate specific news topics, read relevant articles, and ask contextual questions such as, "How are people reacting to this event?" In response, the chatbot generates a simulated debate between two agents with opposing stances, providing users with a third-party perspective on the discussion. Key contributions of this work include: (1) an automated framework for organizing news articles into topical chatrooms, (2) an interactive Q&A module for engaging with the chatbot, and (3) a method to foster broader comprehension of news events through simulated dialogues. This research evaluate the system through a usability study, demonstrating that users successfully engage in multi-turn interactions, gain deeper insights into news topics, and critically evaluate contrasting viewpoints. This work highlights the potential of conversational agents in promoting informed and balanced news consumption. language of the presentation: Japanese | ||||||
| 若山 ことみ | M, 1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 | 清川 清 | 諏訪 博彦 | 松井 智一 |
|
title : Proposal for a Behavioral Change System Using a Smart Mirror for Efficient Stress Management
abstract : An increasing number of workers experience anxiety and stress related to their work and daily lives, leading to rising mental and physical health issues. Preventive measures through feedback on daily habits have become a critical need. This study proposes a system that utilizes a smart mirror and smart devices to estimate the user’s stress levels through counterfactual machine learning. By providing behavioral recommendations based on feedback from the smart mirror, the system aims to efficiently manage stress. Currently, we are in the phase of developing the components of the system and collecting data, which forms the basis of this presentation. language of the presentation: Japanese | ||||||
| 柴田 洸希 | M, 1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 | 林 優一 | 諏訪 博彦 | 松井 智一 |
|
title: Proposal of a Swipe Gesture Recognition Interface System for Existing Furniture Using FPGA and Multi-Vibration Sensors
abstract: In modern households, there is a demand for intuitive and functional interfaces to operate a variety of electronic devices. However, traditional interfaces that rely on cameras and microphones present challenges such as privacy concerns and dependency on environmental conditions. This study proposes a system that leverages FPGA and multi-vibration sensors to utilize existing furniture as an interactive interface. Specifically, vibration sensors are attached to the backside of furniture, and the vibrations generated when swiping the furniture surface are analyzed to recognize the direction of the swipe. Using a dataset created in prior research, experiments were conducted to achieve a balance between reducing model size and maintaining classification accuracy. The results demonstrated that by employing separable convolutions and residual connections, the model size was reduced to 0.34 MB while achieving a classification accuracy of 74%. Future work will focus on exploring models with higher classification accuracy and implementing real-time classification through FPGA integration. language of the presentation: Japanese 発表題目: FPGAとマルチ振動センサを用いた既存家具向けスワイプジェスチャー認識インターフェースシステムの提案 発表概要: 現代の家庭では、多様な電子機器を直感的かつ機能的に操作するインターフェースが求められている。しかし、カメラやマイクを用いた従来のインターフェースにはプライバシーの懸念、環境条件への依存といった課題が残されている。本研究では、FPGAとマルチ振動センサを活用し、既存家具をインタラクティブなインターフェースとして利用するシステムを提案する。具体的には、家具の裏側に振動センサを取り付け、家具表面をスワイプした際に発生する振動を分析することでスワイプの方向を認識する。先行研究で作成されたデータセットを用いて、モデルサイズの縮小と分類精度の両立を目指した実験を行った。その結果、分離畳み込みと残差接続を活用することで、モデルサイズを0.34MBまで削減しながら、74%の分類精度を達成することができた。今後は、より高精度に分類できるモデルの探索と、FPGA上への実装を通じたリアルタイム分類の実現を目指す。 | ||||||
| 飯田 昌直 | M, 1回目発表 | 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 |
|
title: A Divide-and-Conquer Approach for Large-Scale QUBO Problems on Neuromorphic Hardware Using Spiking Representation Learning
abstract: This research proposes a novel representation learning method to efficiently solve large-scale QUBO problems on neuromorphic hardware, such as Intel’s Loihi. Existing approaches are constrained by the limited number of variables that can be processed per chip, making direct large-scale problem solving challenging. To address this, we focus on the traveling salesman problem (TSP) as a representative QUBO instance and employ a divide-and-conquer strategy. We partition the entire problem into smaller TSPs and leverage self-supervised spiking representation learning based on reconstruction to compress variable-length connectivity information into a fixed-length, low-dimensional feature space. We compare multiple representation learning strategies inspired by LCA-based methods, VAE-type generative models, and diffusion models. By clustering in this learned low-dimensional space, we effectively divide the problem. Solving these smaller TSPs on neuromorphic hardware enables scaling to large QUBO problems. In our evaluation, we measure computational resource consumption and the number of timesteps required for convergence, demonstrating the effectiveness of our approach. This research provides a promising foundation for harnessing neuromorphic hardware in large-scale QUBO problem solving. language of the presentation: Japanese 発表題目: スパイキング表現学習を用いたNeuromorphic Hardware向け大規模QUBO問題の分割統治的解法 発表概要: 本研究は、Neuromorphic Hardware（例：Loihi）上で大規模なQUBO問題を効率的に求解する新たな表現学習手法を提案する。既存手法では1チップあたりの処理可能変数数に制約があり、大規模問題を直接求解することは困難である。本研究では、QUBO問題の一例である巡回セールスマン問題（TSP）に着目し、問題全体を小規模TSPへと分割する分割統治的戦略を採用する。その際、再構成に基づく自己教師ありスパイキング表現学習により、可変長の接続情報を固定長かつ低次元な特徴空間へ圧縮する。LCAベース、VAE系生成モデル、Diffusionモデルに着想を得た複数の表現学習戦略を比較し、得られた低次元空間上でのクラスタリングを通じて問題を分割する。分割後の小規模TSPをNeuromorphic Hardware上で求解することで、大規模QUBO問題への対応を可能にする。評価では、計算資源消費や収束までの反復（Timestep）数を測定・比較し、本手法の有効性を示す。これにより、大規模QUBO問題求解におけるNeuromorphic Hardware活用の基盤技術として有望であることが示唆される。 | |||||||
| 衛藤 優 | M, 1回目発表 | 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 |
|
title: Implementing and Evaluating Llama on IMAX as an Energy-Efficient Alternative to GPUs and CPUs for LLM Acceleration
abstract: Recent advancements in AI, particularly large-scale language models (LLMs), have drastically increased the demand for computational resources, raising concerns about energy sustainability as data center energy consumption grows. Current CPU and GPU architectures offer limited improvements in energy efficiency, casting doubt on the sustainability of this approach. Edge devices face similar challenges with limited power, highlighting the need for energy-efficient on-device accelerators. To address these issues, our group developed IMAX (In-Memory Accelerator eXtension), a non-von Neumann architecture. The IMAX3 prototype, based on FPGAs, supports flexible configurations from high-performance computation to energy-efficient edge processing. Previous studies compared IMAX’s energy efficiency with CPUs using the GGML library, showing superior power-delay and energy-delay products. However, GPU comparisons were not included due to GGML's CPU-centric design. In this study, we port llama.cpp to IMAX, enabling direct comparisons with GPUs and refining IMAX’s architecture to evaluate its performance and energy efficiency as an LLM accelerator. language of the presentation: Japanese | |||||||
| 竹内 歩夢 | M, 1回目発表 | 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 |
|
title: Enhancing Power Efficiency in Attention Mechanisms Using IMAX
abstract: The Attention mechanism is central to machine learning models such as Transformers. However, it presents a significant challenge in terms of computational cost, as memory usage scales with O(N^2) with respect to the sequence length N. FlashAttention addresses this issue by performing calculations in block units, reducing memory usage to O(N) and improving power efficiency by minimizing I/O operations. However, existing hardware, such as GPUs, still has limitations in power efficiency. This study proposes leveraging IMAX, a Coarse-Grain Reconfigurable Architecture (CGRA), to achieve substantial improvements in power efficiency. language of the presentation: Japanese | |||||||
| 平野 太雅 | M, 1回目発表 | 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 |
|
title: Spiking Representation Learning towards Achieving Higher Energy Efficiency and More Interpretable Generative Models
abstract: In recent years, latent diffusion models have been one of the most promising options as generative models for various modalities such as image, speech, and language. However, existing high-performance latent diffusion models are computationally expensive and have low model interpretability. Therefore, this study aims to solve these problems by learning representations using the Locally Competitve Algorithm based on Spiking Neural Networks(S-LCA), instead of the Variable Autoencoder (VAE) conventionally used. S-LCA has been shown to provide very high efficiency and low latency seeking compared to classical methods on Neuromorphic Hardware designed specifically for SNN. Although previous studies have successfully achieved high accuracy and low latency image reconstruction using S-LCA, evaluation using the S-LCA encoder as a preprocessing module in classification tasks or in generative tasks has not yet been performed. In addition, since latent space models such as VAE and sparse modeling such as LCA are based on different theories, it is necessary to find a connection between them. I will explore ways to improve S-LCA based on those untested experiments and theories, and aim to realize more energy efficient and interpretable generation models. language of the presentation: Japanese 発表題目: 高効率かつ解釈可能な生成モデルの実現に向けたスパイキング表現学習 発表概要: 近年、潜在拡散モデルは画像・音声・言語など様々なモダリティに対する生成モデルとして最も有力な選択肢の一つである。しかしながら、既存の高性能な潜在拡散モデルは計算コストが大きく、モデルの解釈可能性も低い。そこで本研究では、従来用いられているVariable Autoencoder(VAE)ではなく、Spiking Neural NetworkをベースとしたLocally Competititve Algorithm(S-LCA)による表現学習を行うことでこれらの課題の解決を図る。S-LCAはスパースモデリングの一種であるため解釈可能性に優れ、また、SNN専用に設計されたNeuromorphic Hardware上において、古典的手法と比較して非常に高効率かつ低レイテンシの求解を提供できることが分かっている。先行研究ではS-LCAを用いた高精度・低レイテンシの画像再構成に成功しているが、S-LCAエンコーダを前処理モジュールとして利用した分類タスクによる評価や、生成タスクによる評価などはまだ行われていない。また、VAEのような潜在空間モデルとLCAのようなスパースモデリングはそれぞれ異なる理論に基づいているため、その関連性を見出す必要がある。私は、それら未検証の実験や理論からS-LCAの改善方法を探索し、よりエネルギー効率および解釈可能性の高い生成モデルの実現を目指す。 | |||||||
| 古谷 大介 | M, 1回目発表 | 光メディアインタフェース | 向川 康博 | 和田 隆広 | 舩冨 卓哉 | 藤村 友貴 | 北野 和哉 |
|
title: Restoration of Missing Parts in Ancient Tomb Wall Paintings Using Hyperspectral Imaging
abstract: Cultural heritage sites, such as ancient tombs, represent invaluable assets left by past generations, yet they are increasingly at risk of being lost due to aging and natural disasters. This study aims to restore the missing parts of tomb wall paintings using hyperspectral imaging to reveal what might have been depicted. However, challenges arise due to the deterioration of the wall paintings, which obscures information on pigment distribution and thickness, leaving uncertainty about the extent of recoverable data in the missing areas. By applying principal component analysis (PCA) and existing clustering methods, this study demonstrated the effectiveness of these approaches in classifying pigment thickness. Future work will focus on enhancing restoration accuracy by utilizing models trained on RGB data derived from separated spectral information. language of the presentation: Japanese 発表題目: 分光画像を用いた古墳壁画の消失部分の復元 発表概要: 古墳などの文化遺産は昔の人々が残した貴重な財産であり、経年劣化や災害によって失われつつある文化遺産を将来へ残していく必要がある。本研究では分光画像を用いて、消えてしまった部分に何が描かれていたかを復元することを目指す。しかし壁画の劣化により顔料の分布や厚みの情報が不明であり、消えてしまった部分に情報が残されているか分からないといった問題がある。主成分分析や既存のクラスタリング手法を用いた結果、顔料の厚みの分類に対しては一定の有効性を示すことができた。今後はスペクトル情報を分離してRGBで学習したモデルを用いることで、消失部分の復元を目指す。 | |||||||
| 鶴貝 拓海 | 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: Scratch enhancement rendering of metal cutting surfaces based on length of rays per reflection
abstract: There is a need to develop a technique to efficiently detect scratches on metal cutting surfaces and reduce the number of defective products. The goal of this research is to propose a rendering method that highlights scratches on metal cutting surfaces to improve the accuracy of scratch detection. In metal cutting, microscopic scratches may occur due to the effects of cutting conditions and equipment. This machined surface can be obtained by simulation as data that includes scratches, but these scratches are difficult to see both on the real machined part and on CG, and there is a high risk of overlooking them. To solve this problem, we proposed a rendering method that measures the length of rays for each reflection on the CG, identifies the location of scratches based on the threshold value for determining scratches, and visually enhances the scratches by increasing the brightness of the area. The proposed method improves the visibility of the scratches, and it is confirmed that the type of scratches to be emphasized can be flexibly changed by adjusting the threshold value. In the future, we aim to develop a technique to visualize only specific scratches by devising the threshold setting and to expand the range of expression by combining various types of data. language of the presentation: Japanese 発表題目: 反射ごとの光線の長さに基づいた金属切削面の傷強調レンダリング 発表概要: 金属切削加工面の傷を効率的に検出し,不良品の発生を減らす技術の開発が求められている. 本研究の目標は,金属切削面の傷を強調表示するレンダリング手法を提案し,傷の検出精度を向上させることである. 金属切削加工では,切削条件や機材の影響によって微細な傷が生じることがある.この加工面はシミュレーションによって傷を含んだデータとして取得することが可能だが,これらの傷は現実の加工品でもCG上でも視認が難しく,見落としのリスクが高いという課題がある. この問題を解決するために,CG上で反射ごとの光線の長さを計測し,傷を判定する閾値を基に傷の箇所を特定し,その箇所の明るさを上げることで視覚的に強調させるレンダリング手法を提案した. 提案手法により,傷の可視性が向上し,さらに閾値を調整することで強調する傷の種類を柔軟に変更できることが確認された. 今後は,閾値設定を工夫することで特定の傷のみを可視化する技術の開発や,データの多様な組み合わせによる表現の幅を広げることを目指す. | |||||||