| CHOONHAKLAI PAPON | M, 1回目発表 | ソフトウェア設計学 | 飯田 元, | 藤川 和利, | 市川 昊平 | |
|
title: A Study on GPU Sharing with KubeRay for Machine Learning Workloads
abstract: The increasing complexity of machine learning (ML) and artificial intelligence (AI) applications necessitates efficient GPU resource management in distributed environments such as Kubernetes. Conventional one-to-one GPU mapping, which allocates a single GPU to a single container, often results in the underutilization of these critical resources. This study introduces an approach that leverages KubeRay and time slicing to enable dynamic GPU sharing among multiple concurrent workloads, significantly improving memory utilization and overall response times. The results show that while memory efficiency is notably enhanced, the proposed method incurs longer task completion times due to the overhead associated with managing distributed tasks. Specifically, an average increase in task completion times of approximately 74.43% with two parallel workloads. For three parallel workloads, the average increase in completion times was approximately 158.4%. This study reveals the trade-offs between improved resource utilization and execution time, highlighting the need for future research to optimize these mechanisms in Kubernetes-based ML operations. language of the presentation: English | ||||||
| WEI ZHEYUAN | M, 2回目発表 | ソフトウェア設計学 | 飯田 元, | 藤川 和利, | 市川 昊平 | |
|
title: Optimizing Multipath QUIC Performance through Segment Routing over IPv6 (SRv6) for Efficient Path Selection
abstract: In an era of diverse network environments ranging from WiFi to cellular data networks and satellite communications like Starlink, optimizing data transmission paths is crucial. Multipath QUIC (MP-QUIC) protocol enhances data transport by utilizing multiple network paths but often struggles with efficient path selection. I proposes the integration of Segment Routing over IPv6 (SRv6) with MP-QUIC to refine path selection, ensuring efficient and reliable data flow across complex networks. SRv6 allows for explicit routing control, embedding path information directly into IPv6 headers, enabling dynamic adaptation to network conditions and avoiding congested routes. The research evaluates the performance benefits of SRv6-enhanced MP-QUIC through simulations and real-world testing, demonstrating improvements in throughput, latency, and network resilience. This integration promises to advance multipath transport protocols, optimizing internet communications across heterogeneous networks. language of the presentation: English | ||||||
| MARIS LUCAS | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 藤川 和利, | 諏訪 博彦, | 松田 裕貴 |
|
title: Visual Privacy for Smart Cities: Impact of Image Differential Privacy on Person Re-identification & Demographic Predictions
abstract: Video cameras are prevalent in large cities but their use outside of public safety remains limited due to legitimate privacy concerns. Nevertheless, the rich information they can capture appears incredibly promising for large-scale smart city applications, as they can function as very powerful and versatile sensors. This ambivalence naturally leads to the question of whether such image data can be used in a privacy-responsible manner. For most data types, encryption is the de facto privacy standard, yet encryption assumes the end server can be trusted with keeping data safe; data leaks show us this assumption does not necessarily hold true. As such, privacy approaches have been introduced to offer some level of protection directly on the data itself, regardless of potential leaks. We here extend two of these notions, differential privacy and 𝑘-anonymity, to image data, and extensively evaluate the resulting privacy-utility trade-off on cross-camera person re-identification and attribute recognition data. Our results show that using our proposed approach, image data can be made significantly less privacy-sensitive at source while retaining decent utility for both of these tasks. language of the presentation: English | ||||||
| LIM BRIAN GODWIN SY | D, 中間発表 | 数理情報学 | 池田 和司, | 安本 慶一, | 久保 孝富, | 日永田 智絵, | Li Yuzhe |
|
title: A Recursive Framework for Evaluating Moments Using Zero-Suppressed Binary Decision Diagrams
abstract: The zero-suppressed binary decision diagram (ZDD) is a compact data structure widely used for the efficient representation of families of sparse subsets. Its inherent recursive structure also facilitates easy diagram manipulation and family operations. Practical applications generally fall under discrete optimization, such as combinatorial problems and graph theory. Given its utility, summarizing the subsets represented in the diagram using key metrics is of great value as this provides valuable insights into the characteristics of the family. A recursive algorithm to extract information on moments from families represented as ZDDs is proposed. Given a value for every element in the universe, the value of a subset is first formulated as the sum of the values of its elements. The moments of a family are then calculated as the mean of the exponentiated subset values, akin to the method of moments. Leveraging the structure of ZDDs, the proposed algorithm recursively traverses a given diagram for efficient moments evaluation via multinomial expansion. Its utility is then demonstrated with three classical problems - power sets, the knapsack problem, and paths in graphs - offering orders of magnitude increase in computational efficiency relative to conventional methods. Overall, the proposed algorithm enhances the functionality of the ZDD by introducing an efficient family operation to uncover the distribution of subset values in a represented family. language of the presentation: English | |||||||
| CAO SHIYI | M, 1回目発表 | 数理情報学 | 池田 和司, | 川鍋 一晃, | 杉本 徳和, | 田中 沙織 | |
|
title: Impact of Environment Statistics on Metacognition in
Detection & Discrimination.
abstract: Previous work explores how environmental statistics affect metacognitive judgements in detection and discrimination tasks, sought to use several statistical approaches to understand metacognition, especially in the context of confidence or evidence. This project investigates metacognitive sensitivity, suggesting that environmental statistics influence metacognitive judgment asymmetry, with unbalanced sampling altering decision strategies and reversing asymmetry patterns.To address this, we used a deep neural network to develop a model of decision condifence that operates directly over high-dimensional, naturalistic stimuli. The model accounts for a number of puzzling dissociations between decisions and confidence, with future work focusing on detection tasks and model validation. language of the presentation: English | |||||||
| RAMOS FERNANDEZ ALONSO | M, 2回目発表 | ロボットラーニング | 松原 崇充, | 池田 和司, | 鶴峯 義久, | 佐々木 光, | 角川 勇貴 |
|
title: Deep Reinforcement Learning with FPNN-to-SNN Policy Distillation for Neurochip-driven Robots
abstract: This paper presents a novel deep reinforcement learning (DRL) framework designed to acquire spiking neural network (SNN) policies suitable for neurochip implementation. DRL typically requires complex function approximations, which are challenging to estimate directly with SNNs due to their low approximation accuracy. To address this issue, we propose a method for learning SNN policies by first updating a floating-point neural network (FPNN), then distilling this policy into a quantized neural network (QNN), and finally converting the QNN to an SNN. This approach mitigates the severe performance degradation inherent to SNNs by leveraging the accuracy of FPNNs. Our method was evaluated on two simulation tasks and one real robot task. In all cases, we observed improved sample efficiency compared to previous works. In the future, we plan to analyze the reasons for the sample efficiency improvement of the proposed framework in detail. language of the presentation: English | |||||||
| CHUANG ZHONG | M, 2回目発表 | 数理情報学 | 池田 和司, | 岩田 具治, | 田中 佑典 | ||
|
title: Graph Neural Networks for Meta-learning from Heterogeneous Attribute Graphs
abstract: Prediction of edges between nodes in graph data is useful for many applications. However, it is challenging to train the existing graph neural network-based approaches with little data available. Although some meta-learning approaches for graph neural networks were introduced to make better predictions when little data is available, they require the nodes of training graphs and test graphs to be in common attribute space. This paper proposes a graph neural network-based model for edge prediction that can meta-learn from graphs with nodes in heterogeneous attribute spaces. The proposed model can share the parameters across heterogeneous graphs and be trained effectively without graph-wised iterative procedures to adapt to each graph. Our model outperforms existing approaches in experiments involving edge predictions with sparse edge information across 14 real-world graph datasets. language of the presentation: Japanese | |||||||