| 押尾 怜穏 | 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 |
| DUONG THI SANG | 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 |
| GOL BABAEI BABAK | 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 |
|
title: *** Efficient Deep Neural Network Implementation for Embedded Devices in Resource-Constrained Environments ***
abstract: *** As the demand for intelligent applications grows, the deployment of deep neural networks (DNNs) on embedded devices and in resource-constrained environments has become increasingly critical. This research focuses on developing and optimizing algorithms to enable the efficient implementation of DNNs on platforms with limited computational power, memory, and energy resources. By exploring hardware-aware neural network architectures, quantization techniques, sparse computation and other aprroximate methods, we aim to enhance the performance and reduce the resource consumption of DNNs without comprising much of their accuracy. The outcomes of this research will facilitate the integration of advanced machine learning capabilities into a wide range of applications, including IoT devices, mobile platforms, and edge computing systems, paving the way for more intelligent and autonomous systems in various fields. *** language of the presentation: *** English *** | |||||||
| WANG ZIHANG | M, 2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: Drug Repurposing for Non-Alcoholic Fatty Liver Disease Based On Relations Among Drugs,Disease and Genes abstract: Faced with complex diseases such as non-alcoholic fatty liver disease (NAFLD), traditional drug development pathways are time-consuming and costly. Drug repurposing is an efficient strategy that accelerates the discovery and development of treatments by reevaluating the potential of existing drugs for new diseases. In this study, we explored new uses for existing drugs in the treatment of NAFLD through biclustering analysis of disease-disease relationships and drug-target gene interactions. Using the BioSNAP database, we constructed a bipartite network of drugs and their target genes and applied the BiClusO algorithm to discover high-density clusters. Analysis of these clusters revealed statistically significant groups containing NAFLD risk genes, forming the basis for drug repurposing. Additionally, association analysis identified gene groups that are potential candidates for NAFLD treatment. To comprehensively evaluate and rank the drugs discovered through these strategies, we employed a novel methodology, resulting in the selection of highly promising drugs supported by literature review. This study not only demonstrates the potential of drug repurposing to accelerate therapeutic research but also provides new perspectives and strategies for treating complex diseases like NAFLD, offering valuable reference material for future drug development and disease treatment research. language of the presentation: English | ||||||
| MUHAMMAD HENDRICK SEDAYU | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦, | 松本 健一, | MD.Altaf-Ul-Amin, | 小野 直亮 |
|
title:Extraction of Molecular Features for Classification of Metabolic Pathways of Natural Organic Compounds abstract:Natural compounds have gained popularity as potential sources for the discovery and development of new drugs and therapies. However, the complex chemical structures of these compounds often lead to complex biological activities, thereby posing great challenges in their study. Understanding the biological activity, specifically how each compound interacts and is metabolized in the body, is critical to unlocking its therapeutic potential. By understanding metabolic pathways, we can gain a detailed picture of the chemical transformation stages that occur in a compound, understand the enzymes involved, and identify the final products produced. This understanding is key to advancing the development of effective and safe therapeutic agents derived from natural compounds. We will build a machine-learning model that automatically classifies organic compounds according to their biosynthetic pathway based on the molecular structures using deep-learning models. We construct a model to extract molecular features using molecular fingerprinting, SMILES, molecular graphs, and other descriptors and train classifiers based on deep learning models to identify their metabolic pathways automatically. span class="Apple-converted-space"> language of the presentation: *** English *** | ||||||
| KPALEMON ABENA SAMUEL | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦, | 松本 健一, | MD.Altaf-Ul-Amin, | 小野 直亮 |
|
title: *** Enhanced Estimation of Bone Mineral Density from an X-ray Image Using Denoising Diffusion Probabilistic Model to Quantify Prediction Uncertainty for Osteoporosis Diagnosis *** abstract: *** Osteoporosis is a chronic bone disease that weakens bones and increases their vulnerability to fractures, leading to a significant reduction in mobility. Dual-energy X-ray absorptiometry (DXA), the currently available method for diagnosis of osteoporosis, is expensive and inaccessible. In this study, we aim to estimate bone mineral density (BMD) from a single X-ray image and precisely measure the uncertainty associated with these predictions to improve the accuracy and reliability of the results. Using a dataset of X-ray images and Digitally Reconstructed Radiographs (DRRs), we will apply a Denoising Diffusion Probabilistic Model to reconstruct the DRR of the femur bone region from an X-ray image for assessing bone mineral density and predicting the associated uncertainty. The findings from this research could facilitate low-cost screening and early diagnosis of osteoporosis, thereby reducing expenses and making these services accessible to patients everywhere and at any time.*** language of the presentation: *** English *** | ||||||
| 梶原 隆太郎 | M, 1回目発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 鶴峯 義久, | 佐々木 光 | |
|
title: Hierarchical imitation learning for Cable bundling task
abstract: Servers and communication systems are composed of many devices connected by cables, and the cables need to be bundled to improve maintainability. Currently, this cable bundling work is done manually, and there is a growing demand for robotic automation to improve efficiency. Since cable bundling work is performed in industry, imitation learning, in which a robot learns control policies through human operation, is suitable. However, automating cable bundling using imitation learning is difficult because there are many possible combinations of deformable cable states and corresponding robot actions.In this study, we propose a framework for automating cable bundling work by extracting cable features using topological representation and hierarchical imitation learning. language of the presentation: Japanese | ||||||||
| NGUYEN NGOC HUY | M, 1回目発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 鶴峯 義久, | 佐々木 光, | Kuo Cheng-yu |
|
title: *** Exploring the Limitations of the Neural Acceleration Estimator for Predicting Trajectories of Uneven and Deformable Objects *** abstract: *** Mobile robot object capture is a challenging task, especially in-flight object catching. This becomes even more difficult when capturing uneven and deformable objects because their trajectories are affected by aerodynamics, making their paths unstable and difficult to predict. To accurately capture these objects, two key factors are required: precise object trajectory prediction and an accurate robot control. In this study, we focus on the prediction of object trajectories for a robot to catch in-flight objects. Prediction of object trajectories can be divided into two categories: model-based and model-free approaches. Existing model-based approaches assume simple dynamics of point mass systems, making them inapplicable to objects with uneven shapes or mass distributions. Meanwhile, existing model-free approaches can be applied to these uneven objects by training LSTM on the trajectory of the object's motion state. However, they are limited to rigid bodies and have not been validated for deformable objects with more complex dynamics. The aim of this colloquium is to clarify the limitations of a model-free baseline method, the Neural Acceleration Estimator (NAE). To achieve this, we conduct comprehensive simulations using NAE to predict the trajectories of various objects, including uneven and deformable objects. These simulations will be conducted using the Isaac Sim program. *** language of the presentation: *** English *** | ||||||||
| 本間 天譲 | M, 1回目発表 | ロボットラーニング | 松原 崇充, | 安本 慶一, | 柴田 一騎, | 鶴峯 義久, | 佐々木 光, | 角川 勇貴 |
|
title: Sim-to-Real Reinforcement Learning for Neurochip-Driven Edge Robots
abstract: Neurochips are computational devices suitable for function approximation of Spiking Neural Networks (SNNs), and their power-saving nature makes them promising for use as computational devices for control policies in edge robot tasks where battery capacity is limited. To achieve this, reinforcement learning (RL) methods that learn SNN policies from the interaction between the real environment and the robot have been studied. However, a challenge is that it is extremely difficult to collect sufficient learning samples in a realistic amount of time for edge robot tasks, which involve a huge number of possible state-action transition patterns. In this study, we therefore extend the conventional method to a Sim-to-Real RL framework to solve the data collection problem. As a result, by learning control policies from the interaction between the simulation environment and the robot, sufficient learning samples can be collected in a short amount of time. As a preliminary verification of the proposed framework, we verified that a quadruped robot can navigate a maze in a small-scale maze environment with few state-action transition patterns. In the future, we plan to verify the method in a complex maze environment with many state-action transition patterns and expand the method to achieve the task. language of the presentation: Japanese 発表題目: ニューロチップ駆動エッジロボットのためのSim-to-Real強化学習 発表概要: ニューロチップはSpiking Neural Network (SNN)の関数近似に適した計算機器であり，その省電力性からバッテリー容量が限られるエッジロボットタスクの制御方策の計算機器としての活用が期待されている．その実現のために，実機環境とロボットの相互作用からSNN方策を学習する強化学習(RL)手法が研究されてきた．しかし課題として，膨大な状態行動遷移パターンが考えられるエッジロボットタスクにおいては，十分な学習サンプルを現実的な時間で収集するのは極めて困難である．そこで本研究では，データ収集の課題を解決するために従来手法をSim-to-Real RLフレームワークに拡張する．これにより，シミュレーション環境とロボットの相互作用から制御方策を学習することで，十分な学習サンプルを短時間に収集できる．提案フレームワークの事前検証として，状態行動遷移パターンが少ない小規模迷路環境において，四足歩行ロボットが迷路走破できることを検証した．今後は状態行動遷移パターンが多い，複雑迷路環境での検証とそのタスク達成のための手法拡張を行う予定である． | ||||||||
| 福田 竜平 | M, 1回目発表 | ロボットラーニング | 松原 崇充, | 安本 慶一, | 柴田 一騎, | 鶴峯 義久, | 佐々木 光 | |
|
title: Imitation learning for chemical experiments with skill decomposition
abstract: As the demand for chemical experiments increases, chemical experiment automation by robots is highly expected. To realize the automation of chemical experiments, it is important to imitate the motions of a chemist to a robot by learning the experimental skills by imitation learning. However, motion demonstration in chemical experiments includes multiple skills, and the multiple skills make the learning process complicated. In this study, we propose an imitation learning framework that separates specific skill information from chemists’ motion demonstrations and enables robots to learn skills efficiently. As a result, robots are expected to imitate each skill accurately and automate chemical experiments more efficiently. language of the presentation: Japanese | ||||||||