| 熊谷 一騎 | 1351037: M, 2回目発表 | 伊藤 実, | 井上 美智子, | 楫 勇一, | 関 浩之 |
|
title: A Flash Code Utilizing Dynamic Segment Allocation
abstract: A novel ash code is proposed, and shown to have good average-case performance through computer simulation. The proposed code utilizes dynamic segments which are allocated in an erase block of a fash memory. A segment consists of one or more slices which have 2 cells. A segment is dynamic in the sense that it can be expanded to accommodate as many write operations as possible, and the expansion contributes to assign more number of cells to more frequently written bits. The average write de ciency of the code is evaluated by computer simulation, and shown to have better performance than ILIFC which was proposed by Mahdavifar et al. This result suggests that we can accommodate more data in a xed-size erase block, and, allow longer life-time cycle of fash memory products. language of the presentation: Japanese 発表題目：動的なセグメントを用いたフラッシュ符号の提案 発表概要：フラッシュ符号は，フラッシュメモリの長寿命化を目的とした符号化方式である．本研究では，従来の符号化の仕組みとは異なる記録原理に基づくフラッシュ符号を提案し，その性能を評価する．提案手法では，2 つのセ ルにより1 つのスライスを構成し，１つ以上のスライスで構成される動的なセグメントをフラッシュメモリの消去ブロックに配置することで，データの記録を行う．提案方式では，セグメントを伸縮したり，未使用の位置に新しいセ グメントを配置したりすることが可能であり，データ操作に応じて柔軟に，より多くのwrite 操作を実施することができる．提案手法における平均的書き換え不足量を計算機で評価したところ，既存の符号よりも優れた性能を有することが明らかとなった． | |||||
| 小池 和正 | 1351041: M, 2回目発表 | 中島 康彦, | 井上 美智子, | 姚 駿, | 高前田 伸也 |
|
title: Instruction Execution Method towards Error Reduction of Neural Network
abstract: Approximate Computing has attracted attention as a method for reducing power consumption in varies applications, such as machine learning and image processing．Approximate Computing architecture was proposed in preceding study which indicates that the more approximate instructions are increased，the more power consumption can be reduced． In this paper，we applied approximate computation to memory load address calculation that capable to increase amount of approximated instruction．Furthermore，we proposed a method of raising error tolerance by focusing on characteristic of Neural Network processing． language of the presentation: Japanese 発表題目: ニューラルネットワーク処理のエラー削減に向けた命令実行手法 発表概要: 画像処理や機械学習といったアプリケーションを低消費電力に実行する手法として，Approximate Computing が注目されている．先行研究として，Approximate Computing をサポートするハードウェアが提案されて おり，approximate に実行する命令数が多いほど，消費電力が削減できることが示されている．本論文では，ロード アドレス演算に関する命令を approximate に実行することで，approximate に実行できる命令数を増やした．さらに， ニューラルネットワーク処理の特性に注目した命令実行手法により，少ない precise な命令でエラー耐性を効率よく 上げる手法を提案する． | |||||
| 田ノ元 正和 | 1351066: M, 2回目発表 | 中島 康彦, | 井上 美智子, | 姚 駿, | 高前田 伸也 |
|
title: Convolutional Neural Network Processing on An Accelerator based on Manymemory Network
abstract:Recently, Convolutional Neural Network (CNN) is widely used for image recognition. GPU is generally preferred to accelerate CNN. However, CNN contains preprocessing before its matrix multiplication kernel. GPU is inefficient for this preprocessing part in power consumption. In this work, we proposed an accelerator based on distributed many-memory network, named “EMAX”, to execute CNN effectively. This presentation will introduce the present evaluation of CNN on EMAX. language of the presentation: Japanese 発表題目: メモリネットワークベースアクセラレータを用いた畳み込みニューラルネットワーク処理 発表概要: 近年画像認識の分野において，畳み込みニューラルネットワーク（CNN）が盛んに用いられている． 一般にCNNの高速化にはGPUが用いられるが，行列積における前処理などによってハードウェアの持つ性能を十分に引き出せず非効率的である． 一方我々が提案するメモリネットワークベースアクセラレータEMAXでは畳み込み演算のもつデータの再利用部分を最大限利用し，効率的な実行が可能となる． 本発表ではEMAXを用いたCNNの実行方法と現在の評価について述べる. | |||||
| 都築 匠 | 1351069: M, 2回目発表 | 中島 康彦, | 井上 美智子, | 姚 駿, | 高前田 伸也 |
|
title: Evaluation of optimizing insertion of LUT in PPC
abstract:Digital circuits using state-of-the-art devices are expected to have high failure rate. Thus, efficient techniques to enhance fault tolerance in such circuits is demanded. Recently, a novel fault-tolerant circuit model, Partially-Programmable Circuits (PPC) has been proposed. A PPC replaces parts of a combinational circuit with LUTs and connects a redundant wire from a gate to an LUT. By only configuring the LUTs appropriately, PPCs can bypass some stuck-at-faults. We show that fault tolerance was improved by 10% and 34% when using 3-input and 4-input LUTs, respectively, by optimizing the insertion place of LUTs in PPCs. In addition, we reduce the PPC generation time by limiting the design space to be explored based on the observations we obtained from the experimental results. language of the presentation: Japanese 発表題目: PPCにおけるLUT挿入位置最適化の定量的評価 発表概要: 新素材を用いた回路などの製造故障率が高い回路においては，単純に回路を冗長化する（DMR）だけではなく，より面積効率のよい高信頼化手法が求められ る．そこで，近年，新たな高信頼論理回路モデルとしてPartially-Programmable Circuits (PPC) が提案された．PPCは組み合わせ回路を対象とし，設計時にASICの一部をLUTで置き換え，冗長結線を追加したもので， LUTのコンフィギュレーションの書き換えのみで故障を回避できる．本発表では，LUTの挿入位置を変えることで，従来手法に対して3入力LUTで合成し たPPCでは平均約10%，4入力LUTで合成したPPCでは平均約34%の故障耐性が向上できることを示す．また，実験により得られた知見をもとに，探 索する設計空間を限定することで，平均約70%の探索時間を削減した． | |||||
| 下山 冬馬 | 1351057: M, 2回目発表 | 萩田 紀博, | 向川 康博, | 浮田 宗伯 | ||
|
title: Evaluation of a physical condition from gait motions using a depth sensor abstract: We evaluate a physical condition of a subject based on a human body pose estimated by a depth sensor. In this work, daily gait motions are observed and analyzed. In this talk, the results of 3D pose estimation, how to investigate the difference of gait motions between disabled and able-bodied persons, and future work are reported. language of the presentation: Japanese | ||||||
| 久保 和隆 | 1351036: M, 2回目発表 | 中村 哲, | 向川 康博, | 戸田 智基, | Sakriani Sakti, | Graham Neubig |
|
title: Voice quality control based on perceptual axes to achieve high controllability
abstract: A voice quality control technique using pairs of voice-quality expression words has been proposed in order to intuitively control voice quality to manipulate voice quality scores related to those word pairs in statistical voice conversion. However, its controllability is still insufficient because the conventional expression word pairs were designed for speech analysis rather than voice conversion. To improve controllability, independence among the expression word pairs and sensitivity of each word pair to voice quality control will be essential. In this research, we will investigate new expression word pairs satisfying these demands and improve controllability of voice quality control. language of the presentation:Japanese 発表題目: 操作性の高い知覚軸に沿った声質制御 発表概要: 所望の声質での音声合成を実現するために，声質を人間の主観的な尺度で表現する声質表現語対を用いた統計的な声質変換法が提案されている．しかしながら，その声質制御性能は十分なものとは言い難い．この原因として，従来の声質表現語対は声質分析を目的として選定されたものであり，声質変換における声質制御において最適であるとは限らないことが挙げられる．本研究では，声質変換において最適な声質表現語対の選定を行い，声質変換においてより操作性の高い声質制御法の実現を目指す． | ||||||
| 鶴田 さくら | 1351071: M, 2回目発表 | 中村 哲, | 向川 康博, | 戸田 智基, | Sakriani Sakti, | Graham Neubig |
|
title: An Evaluation of Target Speech for a Nonaudible Murmur Enhancement System in Noisy Environments
abstract: Recently, silent speech interfaces have attracted attention in speech communication area. One of the typical types of silent speech is Non-Audible Murmur (NAM). NAM is a soft whispered voice recorded with NAM microphone through body conduction. NAM allows for silent speech communication as it makes it possible for the speaker to convey their message in a nonaudible voice. However, its intelligibility and naturalness are significantly degraded compared to those of natural speech owing to acoustic changes caused by body conduction. To address this issue, a statistical NAM enhancement method converting NAM into normal speech (NAM-to-Speech) or whisper (NAM-to-Whisper) was proposed. It has been reported that these NAM enhancement methods significantly improve speech quality and intelligibility of NAM, and NAM-to-Whisper is more effective than NAM-to-Speech. However, it is still not obvious which method is more effective if a listener listens to the enhanced speech in noisy environments, a situation that often happens in silent speech communication. In this presentation, assuming a typical situation in which NAM is uttered by a speaker in a quiet environment and conveyed to a listener in noisy environments, we investigate what kinds of target speech are more effective for NAM enhancement.We also propose NAM enhancement methods for converting NAM to other types of target voiced speech. Experiments show that the conversion process into voiced speech is more effective than that into unvoiced speech for generating more intelligible speech in noisy environments. language of the presentation: Japanese | ||||||
| 笹倉 隆史 | 1351049: M, 2回目発表 | 中村 哲, | 向川 康博, | 戸田 智基, | Sakriani Sakti, | Graham Neubig |
|
title: Unknown Word Detection based on Event-Related Brain Desynchronization Responses
abstract: The appearance of unknown words often disturbs communication. Most spoken dialog systems deal with unknown words that are uttered by the user, but which are not covered by the system's vocabulary. In this paper, we focus on detecting unknown words from the user side in which the system utterance is unknown to the user. In particular, we develop a classifier based on Event-Related Desynchronization (ERD) features from user's brain waves, or Electroencephalography (EEG) signal. The results show that we could detect the characteristics of brain waves at the time of unknown word perception significantly better than the chance rate. language of the presentation: Japanese | ||||||