| 遠藤 栄典 | D, 中間発表 | 数理情報学 | 池田 和司☆, 佐藤 嘉伸, 川人 光男(客員）, 森本 淳(客員） |
|
Title: Evaluating resting spatio-temporal dynamics of thalamo-cortical neural-mass network model using resting fMRI connectivity, EEG amplitude intensity distribution and EEG microstate
Abstract: Recently, researches on resting-state network (RSNs) attract a lot of attention in human neuroimaging such as functional Magnetic Resonance Imaging (fMRI) and Electroencephalogram (EEG). RSNs is characterized by resting-state functional connectivity (rsFC) that show spatial correlation based on slow blood fluctuation measured with fMRI and microstate that show non-stationary switching of spatial brain wave based on fast electrical neuronal activity measured with EEG. However, rsFC and microstate explain mainly phenomenological characteristics and the mechanistic origin of RSNs is still ambiguous. Thus, our previous research attempted to reveal rsFC and microstates simultaneously by differential equations that represent dynamics between excitatory and inhibitory neurons and include empirical cortico-cortical structural connectivity. As a result, simulated cortical neuronal dynamics explained characteristics of rsFC and microstates modestly. However, power intensity distributed occipital strongly for empirical EEG but frontal strongly for simulated EEG. Therefore, it is necessary to introduce the mechanism that alpha wave generated by Lateral geniculate nucleus in thalamus transmit to visual cortex and diffuse to other regions. In this study, we calculate thalamo-cortical structural connectivity and simulate thalamic network model to simulate thalamo-cortical neural-mass network model based on empirical structural connectivity. As a result, thalamus connected somato motor and sensory cortex strongly and thalamic network model exhibit entrainment and slow fluctuation for alpha band periodic stimulation. Language of presentation: Japanese | |||
| 趙 崇貴 | D, 中間発表 | ロボティクス | 小笠原 司, 佐藤 嘉伸, 高松 淳, 丁 明, Gustavo Garcia |
|
title: Upper Limb Motion Estimation Based on Skin Deformation Measured with a Distance Sensor Array
Abstract: In various applications, studies of upper limb motion estimation play an important role. In these studies, the methods based on biosignals have gained in popularity because biosignals can be measured using only wearable devices. Among the various biosignals, we focus on skin deformation. The skin deformation provides the motion information about the surface and deep layer muscles, tendons, and bones. In this study, we propose an upper limb motion estimation method for joint angles level based on the skin deformation measured with a distance sensor array. Support Vector Regression (SVR) is used for joint angle estimation. We also develope two types of distance sensor array for measurement of the forearm and upper arm deformation. In addition, we perform motion estimation experiments to verify the accuracy of our proposed method. language of the presentation: Japanese | |||
| 武田 悠佑 | M, 2回目発表 | 自然言語処理学 | 松本 裕治☆, 中村 哲, 澤田 宏(客員）, 岩田 具治(客員）, 新保 仁 |
|
title: Robust Clustering for Relational Data with Various Noises
abstract: We propose a probabilistic model for clustering noisy relational data, such as a document-word network and friend links on social networking services. Relational data often contain noise objects, which do not have meaningful interaction patterns with other objects, such as stop words in document-word networks and spam accounts in a friend network. Existing clustering methods for relational data suffer from these noises because the noise objects are assigned into clusters by assuming that there exist meaningful interaction patterns. The proposed model contains noise clusters, which do not exhibit any patterns to other clusters, as well as normal clusters. The noise clusters enable us to separate the noise objects from the other objects, and help to extract patterns from noisy relational data. To handle various noises, the proposed method assumes an infinite number of noise clusters. Using Dirichlet processes, the number of noise and normal clusters are automatically estimated from the given data. We also present an efficient variable inference for the proposed model based on a collapsed Gibbs sampling. We show the proposed model can more precisely predict the existence of unobserved relations and extract various noise objects than conventional models through the experiments using synthetic data and real-world data sets. language of the presentation: Japanese | |||
| 平岡 達也 | 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 |
|
title:
Unsupervised Word Segmentation with Neural N-gram Language Models
abstract: We propose a new architecture of unsupervised word segmentation with neural language models. Our model is able to segment text data which causes localized segmen- tations by a previous method with a discrete language model. Word segmentation and representations gained by our model are evaluated on sentence classification tasks in Japanese, Chinese, and English. It is confirmed that segmentations and representations by our model have a good effect on the results of subsequent tasks. language of the presentation: Japanese | |||
| 本多 右京 | 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 |
|
title: Natural Language Inference with Image*
abstract: Natural Language Inderence (NLI) is a task to classify the relation of the given two sentences into entailment, contradiction and neutral. Although NLI is designed as a natural language processing task, some instences require implicit world knowledge, such as "in a bar" entails "be inside of a building". To compensate the world knowledge, we exploit images which correspond to the given sentences. Firstly we select the parts of the image to which the words in the each sentence correspond, and then compare the parts as well as the sentences themselves. Our experimental results show that our method outperforms the sentence-only baseline. language of the presentation: Japaese | |||
| 松野 智紀 | 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 |
|
title:Reduction of Parameter Redundancy in Biaffine Classifiers with Symmetric and Circulant Weight Matrices:
abstract: Currently, the biaffine classifier has been attracting attention as a method to introduce an attention mechanism into the modeling of binary relations. For instance, in the field of dependency parsing, the Deep Biaffine Parser by Dozat and Manning has achieved state-of-the-art performance as a graph-based dependency parser on the English Penn Treebank and CoNLL 2017 shared task. On the other hand, it is reported that parameter redundancy in the weight matrix in biaffine classifiers, which has O(n^2) parameters, results in overfitting (n is the number of dimensions). In this paper, we attempted to reduce the parameter redundancy by assuming either symmetry or circularity of weight matrices. In our experiments on the CoNLL 2017 shared task dataset, our model achieved better or comparable accuracy on most of the treebanks with more than 16% parameter reduction. language of the presentation: Japanese | |||