| 長谷 洋斗 | M, 2回目発表 | 大規模システム管理 | 笠原 正治, | 井上 美智子 | 笹部 昌弘, | 川原 純（客員准教授） |
|
title: Novel beam search-based algorithms for variable ordering accelerating frontier-based search
abstract: Enumerating subgraphs on a graph is a useful approach for graph optimization problems. Frontier-based search is often used for enumerating subgraphs. The efficiency of frontier-based search depends on a given variable order. Among many heuristics for variable ordering, it is known that a beam search-based algorithm tends to find a good variable order. However, frontier-based search following the variable order obtained from the method failed to run on some graphs due to mainly the memory shortage. To overcome the problem, we propose novel beam search based algorithms. We show frontier-based search with the variable order the proposed methods output more efficiently performs than that with the variable order the existing method outputs for almost all instances. language of the presentation: Japanese | ||||||
| 馬場 瑛義 | M, 2回目発表 | 大規模システム管理 | 笠原 正治, | 井上 美智子 | 笹部 昌弘, | 川原 純（客員准教授） |
|
title: Algorithms on constructing ZDD representing Meyniel graphs and crossing chordal graphs
abstract: It is typical to enumerate substructures that belong to certain graph classes in graph problems. We consider to apply a data structure called ZDD, which represents data in a compressed way to reprensenting a set of certain subgraphs. In this study, we use colored ZDD, which is extended from conventional ZDD, and design algorithms to construct ZDD of subgraphs belonging to Meyniel graphs and (k, 2)-crossing chordal graphs, which are a variety of choral graphs. language of the presentation: Japanese | ||||||
| 滝下 雄太 | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 井上 美智子, | 中田 尚, | TRAN THI HONG, 張任遠 |
|
title: Image recognition by simulation of using xbar memristor
abstract: In recent years, a neuromorphic computer has been attracting attention as a computer replacing Neumann computers. The neuromorphic computer is a systemthat imitates the neural circuit of the human brain and isa specialized computer for realizing artificial intelligence. There are approaches using digital elements such as CMOS and those using analog elements that use physical phenomena. In this study, we have developed a hybrid system byimplementation of visual cortex as a digital element and xbar memristor as an analog element. The visual cortex layer works as a feature extraction layer and it does notrequire training. This layer imitates visual cortex of living organisms. On the other hand, the oxide semiconductor memristor has a characteristic that the resistance value changes depending on the times and value of voltage application. This oxide semiconductor memristor can be formed in an xbar shape and has a very simple structure. By usingthe structure, it is possible to highly integrate the synapses. Therefore, we have developed a training algorithm using the oxide semiconductor memristor for the training of neural networks without back propagation and evaluated its performance for image recognition. As a result, a maximum of 70% of correct answers were obtained in the simulation using the MNIST datasets. In addition, it was confirmed that there was a suitable training rate depending onthe datasets. Although the recognition accuracy is not so high, it is meaningful that it is confirmed that the completely novel concept mentioned above can operate correctly, and the recognition accuracy will be improved by the further optimization. language of the presentation: Japanese | ||||||
| 本田 卓 | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 井上 美智子, | 中田 尚, | TRAN THI HONG, 張任遠 |
|
title: Speeding Up VBGMM By Using Logsumexp With the Approximate Exp-function
abstract: Gaussian Mixture Models (GMMs) are a representative method to realize clustering and are used in many applications such as probability density modeling and soft clustering. There are several methods for parameter estimation of the GMM, among which the Variational Bayesian Gaussian Mixture Model (VBGMM) is known that they are hardly to overfitting than other methods. In the VBGMM, the logsumexp takes up most of the computation time. This is due to the heavy use of the exponential function, which is computationally intensive. In this paper, we will explain a method to speed up the logsumexp by using an approximate exponential function. As a result, the logsumexp is accelerated by 1.79x, and the total speed of the VBGMM is improved by 1.08x. language of the presentation: Japanese | ||||||
| 鶴尾 明大 | M, 2回目発表 | 数理情報学 | 池田 和司, | 杉本 謙二, | 吉本 潤一郎, | 久保 孝富（特任准教授）, | 福嶋 誠, | 日永田智絵 |
|
title: Mathematical Model of Jumping Horse
and Its Rider
abstract: Horse riding is an interaction between a horse and its rider. To understand its interaction, their movements at trot are successfully modeled by spring-damper-mass (SDM) models. The SDM models show that a rider’s technique decreases the mechanical work of the horse. In contrast, the rider’s policy at jumping, which is not a periodical movement, is unclear in terms of the mechanical perspective. To solve this problem, firsty, we examined whether the SDM models can apply to jumping. As a result, one of the SDM models succeeded to represent trajectories matched with observed ones at jumping. Additionally, monitoring the spring stiffness from the model implies that the rider changes their body stiffness drastically. Currently, we are investigating the rider’s effect on the horse in terms of the mechanical perspective. language of the presentation:Japanese | ||||||||
| 龍田 侑弥 | M, 2回目発表 | ロボティクス | 小笠原 司, | 杉本 謙二, | 高松 淳, | GARCIA RICARDEZ GUSTAVO ALFONSO | ||
|
title: Manipulation Using a Spherical Jamming Gripper - Tumbling Operation without Force Control -
abstract: My reseach purpose is to explore the possibility of assembly tasks by grasp-less manipulation with a jamming gripper's trajectory keeping point contact in order to achieve tumbling operation of a pin by a spherical jamming gripper. The experiments with a physical robot arm show the success rate of the tumbling operation by the proposed trajectory is higher than a straight trajectory. The proposed trajectory can achieve the tumbling operation with a quasi-static process unlike a straight trajectory. The effectiveness of the proposed method were also evaluated using three types of objects, a plastic bottle, a sandwich and a rice ball. language of the presentation: Japanese | ||||||||
| 関 直哉 | M, 2回目発表 | 知能コミュニケーション | 中村 哲☆, | 荒牧 英治, | 鳥澤 健太郎（客員教授）, | 飯田 龍（客員准教授） | ||
|
title: Answer Selection in Factoid Question Answering using Sentiment Classification
abstract: Factoid question answering is a task of answering what-, who-, when-, and where-type questions. Users often ask a question such as “What should diabetics eat?”, which asks positive aspect of something. Existing question answering methods, however, cannot handle questions with such positive or negative aspect appropriately. Sentiment classification is a task of classifying sentiment polarity of a sentence or a document, but our preliminary experiment shows that just applying the sentiment classification method to factoid question answering doesn’t work well. Therefore, we propose methods (1) to identify clues for sentiment classification by span extraction, (2) to collect clues for that by asking what-happens-if type questions (e.g., What happens if global warming worsens?). language of the presentation: Japanese | ||||||||
| 二宮 大空 | M, 2回目発表 | 知能コミュニケーション | 中村 哲☆, | 荒牧 英治, | 宮尾 知幸, | 鳥澤 健太郎（客員教授）, | 飯田 龍（客員准教授） | |
|
title: Answer Passage Ranking in Conditional Why-Question Answering
abstract: Conditional Why-Question Answering (why-QA) is a task to retrieve answers from a given text archive for a conditional why-question such as "Why does the sea level rises as global warming progresses?, which can provide the reasoning for the causality between"A: the sea level rises" and "B: global warming progresses." Conditional why-question contributes to increasing the reliability of causality. However, the previous works seldom consider the causality between A and B, and our preliminary experiments show that existing methods have a poor performance on such questions. Therefore, we assume that combining various causal knowledge contributes to good performance for Conditional why-QA, and we propose the following four methods; (1)additional pre-training, (2)multi-task learning, (3)combining some BERT, (4)retrieving causality related to answers, with a large amount of causal knowledge for Answer Passage Ranking. language of the presentation: Japanese | ||||||||