Search

In this study, we propose a method for generating a sequence of notes of a human-like performance. The human-like performance here refers to a session-like improvised jazz performance given a chord progression, rather than a classical music performance with a score prepared in advance. This study attempts to reproduce session-like performances by learning features from session-like bass performances. Based on the proposed method, we learned the performance of bassist James Jamerson using his performance as the teacher data. Using the learned features of the bass performance, we generated a sequence of notes using a hidden Markov model and neural networks such as LSTM and Bi-LSTM and confirmed the effectiveness of the proposed method.

We propose a hidden Markov model (HMM) based method to categorize eye movement pattern in terms of machine learning technique, which was applied to the cognitive process of impression evaluation of paintings. I n our experiment, we tracked the gaze transition pattern of subjects while they evaluated their personal impressions of traditional Japanese Ukiyo-e paintings. For each class of positive/negative impressions, we estimated the HMM parameters from the training samples of eye movements. For the eye movement data of the test samples, we conducted a classification test based on the differences of the log likelihood values obtained from each HMM. For the cases of receiving negative impression, we identified a common pattern of eye movement from a broader perspective, but no such holistic patterns of eye movement were found for the cases of receiving positive impression.

教育の最も難しい問題は，教師は学習者に教えすぎても，教えなさすぎても学習者の十分な発達は望めないということである．そのために，教師は個々の学習者の理解度や最適な支援の度合いを予測することが重要な課題となっている．足場がけによる学習者のパフォーマンスを予測するために，項目反応理論を用いて最適な予測正答確率になるようにヒントを提示する足場がけシステムが開発されている．しかし，従来の項目反応理論では，学習者の能力変化がモデルに考慮されておらず，正確な正答確率を予測できないために，最適なヒント数を予測できていない可能性がある．本研究では，学習者の能力が時間変化していくプロセスを項目反応理論に組み込み，能力が隠れマルコフ過程に従って変化すると仮定した新しい項目反応モデルを提案する．提案モデルでは，能力値が継続する時間(課題数) を表すウィンドウサイズと能力の変動の程度を反映する変動パラメータをもち，これらの最適値がデータから推定されるために，学習者の真の能力変化を反映でき，予測精度を向上させることが期待される．実データを用いて，本提案の有効性を示す．, To scaffold a learner efficiently, a teacher should predict the optimal degree of assistance to support learner's development. However, conventional Item Response Theory (IRT) model does not consider learner's ability changes during his/her studying, therefore the IRT model might cause over-assistance or lack of assistance. We propose a new IRT model that incorporates learner's ability change according to a Hidden Markov process. The proposed model has the following two new parameters: the degree of the ability changes and the period of time that the learner's ability does not change. The experiments result shows that the proposed model improves the prediction accuracy of learner's performances.

Sign Language are an important comminucation method for learning impaired people. To Communicate to learning impared people, mprmal people also have to learn the gesure of meanings. In order to achieve smooth communication between learning impaired people. automatic sighn recognition system is required. In this paper, we applied. The combination of simple 2D camera convoluitional neural network(CNN) Classification method is applied. The combination of simple 2D camera with CNN enable real-time recognition is achieved by using GPU based mixraocomtroller board Through preliminary experiments, we can classify three different types of gestures with accuracy of 70％.

We examined the relationship between the gaze movement and the recognition result in the face recognition work focusing on the state transition of the gaze movement. For analysis, Hidden Markov Model (HMM) widely used for speech recognition and handwritten character recognition, etc. was adopted. The reappraisal result was evaluated by the Hit rate which is the rate at which the face image was correctly recognized. Results showed that HMM recognition rate was higher for people with lower Hit rate and HMM recognition rate was lower for people with higher Hit rate. From the results, it was suggested that people with higher reapproval results do not have a fixed pattern for transitions of viewpoints, while those with lower recognition results may have a fixed pattern for transitions of viewpoints.

本研究では，旅行中の観光客の目的の変化を隠れマルコフモデルを用いてモデル化する．さらに，観光客の目的を統計的決定理論に基づいて推定する方法を提案する．観光目的の推定に関する提案方法では，推定誤り率をベイズ基準のもとで最小化する．また，観光客が次に訪問する観光施設を統計的決定理論に基づいて推薦する方法を提案する．推薦に関する提案方法では，推薦誤り率をベイズ基準のもとで最小化する．提案モデルおよび提案方法の有効性を数値計算例によって確認する．, In this research we use hidden Markov model in order to represent changes of tourist purposes. We propose an estimation method for tourist purposes based on statistical decision theory. The proposed estimation method minimizes the estimation error rate with reference to the Bayes criterion. We proρose a recommendation method for visiting tourist facilities based on stαtistical decision theory. The proposed recommendation method minimizes the recommendation error rate with reference to the Bayes criterion. We show the effectiveness of the proposed model and methods by some numerical calculation examples.

application/pdf

本論文では、論文からプレゼンテーションシート自動生成を目指し、隠れマルコフモデル（HMM)を用いた論文とプレゼンテーションシート（シート）との対応付けを試みる。対応付けには、JingのHMMを用いた対応付け手法をもとに、シートごとに論文の章・節を特定することを行った。しかし、論文とシートを対象とした場合では、新聞記事と比べ一テキスト中の文数と一文当たりの長さが大きく、また本文の他に図表が含まれるため、単純にJingの手法を適用しても十分な対応付けができない。そこで、本論文では、論文とシートの対応位置やレイアウトなどの情報を利用した対応付け手法を提案する。評価実験では、人手で作成した正解データをもとに７７％の対応付け精度が得られた。また、提案した改善手法も有効性を確認した。 : We have been studying towards automatic generation of presentation sheets from a technical paper. In this paper, we propose an alignment method between sections in a technical paper and presentation sheets. Our method is based on the HMM alignment method by Jing. Thought her method is effective to align with short sentences in a newspaper, it is difficult to align with a paper including chart and long sentences. So, we proposed and adapted an alignment approach using feature with a paper and sheets, such as information from text appearance. In experiments, our method showed 77% precision and our strategies were efficiency.

application/pdf

application/pdf, The purpose of our research is to improve the recognition rate of off-line character recognition systems using the HMM (Hidden Markov Model). Some 1-dimensional (1-D) HMM recognition systems have been proposed to increase the recognition rate of the conventional recognition systems. However, the 1-D HMM recognition systems have not achieved a sufficient recognition rate. On the other hand, 2-D HMM character recognition systems have been proposed to increase the representational power. However, the complex structure of 2-D HMM could not produce enough representational power with the actual small number of learning samples. To overcome these problems, we proposed a method using the multiple 1-dimensional HMM approach. In our experiment, the recognition rate of our multiple 1-D HMM method performed 97.1%, which was about 3% better than the conventional 1-D HMM systems, and as good as the 2-D HMM (HMM-MRF) system. Thus our method has better performance in practical use.

本稿は、大語彙の連続音声認識を実時間で処理するための、HMM-LR連続音声認識装置の設計、処理性能について述べる。HMM-LR法は、一般化LR構文解析により入力音声データ中の音素を予測し、予測された音素の存在確率をHMM音素照合により調べることで、音声認識と言語処理を同時進行させる方式であり、高精度で効率的な処理系を構成することができる。処理量が極めて大きな継続時間制御付きトレリス計算を伴うHMM-LR連続音声認識を実時間で実行するために、本装置は種々の高速化手法を33個のDSPを用いて実現した。その結果、連続文節発声の入力音声に対し、入力文の長さに依らずに、発声終了後から2〜3秒ですべての認識処理を終了する処理速度を達成した。

近年ロボットは技術の発達や低価格化に伴い、日常生活により身近になってきている。しかし未だ十分な自律性とそのための学習機構を持つには遠く、ロボットの思考や行動は逐一設計者やユーザが編集、作成しなければならない。行動学習において動作の獲得や認識、再生、合成を扱った研究は多いものの、動作から概念形成に役立つ情報を得ようというのを直接ねらいとする研究は少ない。動作データ間に存在する類似性とその理由を見出して整理することは、柔軟な認識や思考に役立つ情報と考えられる。本論文では隠れマルコフモデルを用いた行動認識に基づき、観測した複数の動作間の類似性と、その原因となっている体の部位を抽出する手法を提案する。 検証実験の結果、擬似データに対しては正解率100%、人間の動作に対しては、7動作中"蹴る", "パンチ", "手を振る", "寝る"の動作において、人間の主観に近い結果を得ることに成功した。, 卒業論文

1 Introduction||2 Fundamental theory||3 Methodology||4 Experimental results||5 Conclusions, 九州工業大学博士学位論文 学位記番号:情工博甲第346号 学位授与年月日:令和2年3月25日, 令和元年度