| 小原 一貴 | 1651031: M, 2回目発表 | ネットワークシステム学 岡田 実, 安本 慶一, 東野 武史, Duong Quang Thang |
|
title: Rainfall prediction using particle filter
abstract: In recent years localized heavy rain called guerilla heavy rain is a problem. It is necessary to predict the occurrence of disasters caused by localized heavy rain in advance and minimize the damage by accurately grasping the situation. As one of countermeasures, phased array radars capable of measuring three-dimensional raindrop distribution have been developed, and it is becoming possible to observe high spatiotemporal resolution. However, the amount of data acquired by the phased array radar is enormous and it is not realistic to send the observed raw data to the required place. Therefore, in this report, I propose a method to drastically reduce rainfall data by modeling observed data with 3-D volume model. By using this method, it is possible to transmit data on an inexpensive internet line with a transmission speed of about several [Mbps]. In addition, we propose a system that predicts transmitted data using a particle filter. In the proposed system, a particle model is constructed by using the moving direction and increasing or decreasing direction of each rain clump in the 3-D volume model as state vectors. In this paper, we evaluate the change of the prediction accuracy by the particle number and the sampling period of the particle filter. language of the presentation: Japanese (choose one) | ||
| 河端 洋人 | 1651039: M, 2回目発表 | ネットワークシステム学 岡田 実, 杉本 謙二, 東野 武史, Duong Quang Thang |
|
title: Theoretical analysis on maximum efficiency of multiple-transmitter inductive power transfer system abstract: This study investigates the maximum power transfer efficiency (RFtoRF efficiency) of inductive power transfer (IPT) system using an arbitrary number of transmitters under an assumption that there is significant cross-coupling among transmitters. We model the system by Z-parameter and derive the maximum efficiency, the optimum load impedance and input currents based on the first order necessary condition. We confirm the validity of analytical solution by electromagnetic field simulation on IPT systems with 4 transmitters. language of the presentation: Japanese | ||
| 杉山 玲央奈 | 1651062: M, 2回目発表 | ネットワークシステム学 岡田 実, 杉本 謙二, 東野 武史, Duong Quang Thang |
|
title: Analysis of Maximum Achievable Efficiency in Multiple-Receiver Inductive Power Transfer with Cross-Coupling abstract: This research investigates maximum achievable efficiency of inductive power transfer (IPT) system with arbitrary number of receivers. We derive the formulas of optimal loads and maximum efficiency based on N-port network model. Via analysis and computer simulations, we show that the cross- coupling among receivers does not affect system efficiency if load resistances and reactances are jointly optimized. We also prove that the maximum efficiency is dominated by a so-called system kQ-product, whose square is interestingly, equal to the sum of squares of kQ-products of individual transmitter-receiver links. language of the presentation: Japanese | ||
| RATIH HIKMAH PUSPITA | 1651130: M, 2回目発表 | ネットワークシステム学 岡田 実, 杉本 謙二, 東野 武史, Duong Quang Thang |
|
title: Compressed Sensing based Channel Estimation Algorithm for MIMO-OFDM System
abstract: A multiple-input multiple-output (MIMO) wireless communication system combined with the orthogonal frequency division multiplexing (OFDM) modulation technique can mitigate the effects of multipath fading and achieve reliable high data rate transmission over broadband wireless channels. The channel estimation determines the quality of the data transmission from transmitter to receiver. Channel estimation using Compressed Sensing (CS) algorithms can achieve higher correctness using small number of pilots than that of conventional interpolation based methods. This research proposed to reduce the complexity in sensing matrix algorithm for CS based channel estimation in MIMO-OFDM system. For STBC MIMO, the pilot matrixes of multiple antennas are same. So we can use the same measurement matrix for CS algorithm process to reduce its computational complexity. In addition, for multiple antenna system, we can assume the position of CIR in time domain are almost same. Therefore we can use average process to reduce the AWGN noise component to find the correct positions of CIR which its vital for CS based channel estimation. Our proposed method can reduce the complexity and achieve good BER performance. language of the presentation: English | ||
| 森元 彩華 | 1661018: D, 中間発表 | 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之, 能地 宏 |
|
title: Extraction of verb phrase relationships considering Multiword-Verb
abstract: I extract relationships between verbs of two sentences connected by conjunction as expression of cause and effect using expression learning. I propose to evaluate the obtained relation vector using RTE task. language of the presentation: Japanese | ||
| 真鍋 陽俊 | 1651100: M, 2回目発表 | 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之, 能地 宏 |
|
title: Data-dependent Learning of Symmetric/Antisymmetric Relations
for Knowledge Base Completion
abstract: Embedding-based methods for knowledge base completion (KBC) learn representations of entities and relations in a vector space, along with the scoring function that estimates the likelihood of relations between entities. In these methods, the learnable class of scoring functions is designed to be expressive enough to cope with a variety of real-world relations. Although this expressiveness is desirable, they come at a cost of increased number of parameters, leading to the difficulty in learning suitable representations. In particular, they are often superfluous for symmetric and antisymmetric relations that form a vast majority in a knowledge base. To overcome this difficulty, we propose a new L1 regularization for Complex Embeddings, one of the state-of-the-art embedding-based methods for KBC. Our method promotes the learned representation of relations to be more symmetric or antisymmetric in a data-dependent manner. Our empirical evaluation shows that the proposed method outperforms the original Complex Embeddings and other baseline methods. language of the presentation: Japanese | ||
| ZUO YONG | 1651134: M, 2回目発表 | 自然言語処理学 松本 裕治, 中村 哲, 新保 仁, 進藤 裕之, 能地 宏 |
|
title: Multi-Sense Embeddings for Semantic Role Labeling
abstract: Semantic Role Labeling(SRL) is a fundamental natural language processing task to discover the predict-argument structure of a predicate in a given sentence. In traditional methods, each word type has only one vector representation, which means that polysemy are ignored. We propose applying the multi-sense embedding for the task that could lead to more fine-grained model. From the result of the experiment, while using multi-sense embeddings is helpful to some degree, the advantage of multi-sense embeddings disappears as model becomes more sophisticated. We guess this is because complicated model like LSTM can handle better to seperate the information from the one-word-one-vector embeddings and are trying different pipeline system to resolve the problem. language of the presentation: Japanese | ||
| 福岡 久和 | 1651091: 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: Distributed neural network for video recognition using compression and aggregation
abstract: In recent years, DeepNeuralNetwork (DNN), which attracts attention, has been proposed for use in various applications. The main execution environment is the cloud, and an increase in the number of devices is one of the causes of traffic congestion and power consumption on servers and routers. We propose a distributed neural network which reduces the computational complexity by taking advantage of the characteristics of CNN, and finishes the processing in the edge devices, for high-workload video recognition DNN. language of the presentation: Japanese 発表題目: 圧縮・集約を用いた動画認識を対象とする分散ニューラルネットワーク 発表概要:近年、注目を浴びているDeepNeuralNetwork(DNN)は様々な用途での利用が提案されている。主な実行環境はクラウドであり、デバイス数増加がトラフィック輻輳やサーバ・ルータ消費電力の問題の一因となっている。私たちは、高負荷な動画認識DNNを対象とし、CNNの特徴を活かした計算量削減、エッジ群で処理を完結させる分散ニューラルネットワークを提案する。 | ||
| 山野 龍佑 | 1651118: 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: Time-division multiplexed execution of CGRA
abstract: FPGA, Systolic array and CGRA are being studied as the general-purpose accelerator which made up for the consumption electricity that was a bottleneck of the Neumann type architecture. It's reported that accelerating of DNN by an accelerator of a systolic array base shows the performance beyond GPU recently. We developed EMAXV (Energy-aware Multimode Accelerator eXtension) which was CGRA accelerator. However, as a feature of CGRA, there was a problem that the scalability is low due to the wiring complexity. So, we propose EMAXVI which is a systolic array based accelerator which solves wiring congestion. Currently, performance evaluation by simulator is carried out. language of the presentation: Japanese | ||