| 荒木 駿佑 | M, 1回目発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 鶴峯 義久, 佐々木 光 |
|
title: Automation of Grinding Tasks Using Imitation Learning Based on Human Force Data
abstract: In recent years, robots have been used to automate human tasks. In the grinding task, which involves shape manipulation by removing unnecessary parts, it is necessary to grasp the changing shape and plan movements that consider the grinding resistance that occurs during machining. In grinding operations performed by humans, a large percentage of grinding operations are based on force information transmitted to the fingertips to determine the state of machining and the magnitude of removal resistance. Therefore, the grinding operation is automated based on the force information during the grinding process by humans. To achieve this objective, we extract the force information used by a human to perform the grinding operation using bilateral control and perform imitation learning. In a preliminary experiment, we verified the imitation performance under bilateral control by performing the task of writing sin waves while pressing against a piece of paper. language of the presentation:Japanese 発表題目:人間の力情報に基づく模倣学習を用いた研削タスクの自動化 発表概要近年，ロボットによる人の作業の自動化が行われている．その中で，不要な部分を削り取ることによって形状操作を行う研削タスクでは，変化する形状の把握と加工時に生じる研削抵抗を考慮した動作計画を行う必要がある．人間が行う研削作業では，手先に伝わる力情報によって，加工具合や除去抵抗の大きさを把握する割合が多い．そのため，人間が研削加工を実施する際の力情報に基づいて，研削動作の自動化を行う．本研究では，この目的を達成するためにバイラテラル制御を用いて，人間が研削動作を行う際に用いる力情報を抽出し，模倣学習を行う．予備実験において，紙に押し付けながらsin波を書くタスクを行い，バイラテラル制御下の模倣性能の検証した． | |||
| 林 純子 | M, 1回目発表 | 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, 矢田 竣太郎 |
|
title: Estimating Happiness from Workplace Daily Reports abstract: In recent years, well-being has garnered attention as a crucial indicator for enhancing people's lives, and investigating individuals' happiness is essential for improving well-being. Currently, the prevalent method to measure happiness involves respondents providing numerical ratings in response to questions. However, this method poses issues such as respondents falsely presenting themselves in a more positive light or the surveys not accommodating frequent short-term assessments. Therefore, there has been a call for alternative approaches to measure happiness. This study proposes a method to estimate happiness from short texts. Specifically, it collects diary entries along with happiness ratings (on a scale of 0 to 10) for each day and constructs a classifier to gauge happiness from the text. language of the presentation: Japanese | |||
| 辻本 陵 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎, 安本 慶一, 大内 啓樹 |
|
title: Generating Explanations for Temporal Changes in Remote Sensing Images
abstract: Satellite images are essential for acquiring geographic spatial information and monitoring changes in land use, with a growing demand for detailed explanations of these changes. This study aims to input satellite image datasets into a multimodal large-scale language model and automatically generate detailed explanations about temporal changes. In this presentation, we will highlight current issues and challenges, proposing improvements in accuracy through dataset expansion and the creation of evaluation metrics. language of the presentation: Japanese 発表題目: 衛星画像を活用した時系列的な変化の説明生成 発表概要: 衛星画像は、地理空間情報の取得や土地利用の変化のモニタリングに不可欠であり、その変化の詳細な説明に対する需要が高まっている。本研究では、衛星画像データセットをマルチモーダル大規模言語モデルに入力し、時間的変化に関する詳細な説明を自動生成することを目指す。本発表では、現在の問題点と課題を明らかにし、データセットの拡張による精度の向上と評価指標の作成を提案する。 | |||
| 田口 穂鷹 | M, 1回目発表 | 大規模システム管理 | 笠原 正治, 藤川 和利, 原 崇徳 |
|
title: On Network Learning/Inference in eBPF Data Plane
abstract: The fusion of machine learning and networking holds the potential to promote intelligent networking. Programmable network devices achieve in-network inference by offloading CPU-intensive tasks to programmable data planes on dedicated hardware. exteneded Berkeley Packet Filter (eBPF) is an abstract virtual machine with an instruction set that can execute arbitrary programs in a Linux kernel, which realizes efficient packet processing. This research aims to explore the potential of in-network learning/inference on generic hardware by integrating neural networks with the eBPF data plane. In this presentation, we delve several machine learning and eBPF techniques to clarify the viability of performing network inference within the kernel space and learning the latest traffic trends in the user space in an online manner. language of the presentation: Japanese | |||
| 灘井 美樹 | M, 1回目発表 | 大規模システム管理 | 笠原 正治, 池田 和司, 井上 美智子 |
|
title: Nudge: The Scheduling Algorithm Stochastically Improving upon FCFS for Light-tailed Job Size Distributions
abstract: The First-Come First-Served (FCFS) scheduling policy is one of the simplest scheduling algorithm used in practice. Furthermore, its usage is theoretically validated. For light-tailed job size distributions, FCFS has weakly optimal asymptotic tail of response time. But outside of the asymptotic regime, optimality was open. Recently it was shown that the FCFS scheduling algorithm can be stochastically improved upon by a scheduling algorithm called Nudge for light-tailed job size distributions. Nudge divides jobs into four regions based on their job sizes i.e. small, medium, large and very large. The scheduling by Nudge is basically the same as that of FCFS, but when a small job arrives and a large job is found at the end of the queue, the small job and the large job are swapped unless the large job was already involved in an earlier swap. In this presentation, we introduce the sufficient conditions for Nudge stochastically improve upon FCFS. language of the presentation: Japanese | |||