ページの先頭です

ニュース&トピックス

情報数理学専攻 平成29年度情報数理学セミナー

今後の開催予定

第4回

日時: 6月1日(木) 14:40~16:10
会場: 情報科学研究科A棟 A109講義室
内容: 講演会
講演者: 田嶋達裕氏(ジュネーブ大学研究員)
講演題目: Consciousness and Cross-Embedding (意識と相互埋め込み)
概要: Brain-wide, complex neural dynamics are considered crucial for cognitive functions and consciousness in the human and animals. The system complexity, however, often defies conventional mechanistic analyses via reductionist modelings. I propose that such a dilemma could be resolved partially by using theorems in general dynamical systems. In this lecture, I introduce Takens' embedding theorem in general dynamical systems and its application to neural data analyses. The method enables us to characterize the dynamical structures and causal interactions among variables without relying on equation-based models. Applying this method to wide-field electrophysiological data reveals a universal hierarchy of cross-area interactions and dynamical complexity in the conscious brain, demonstrating its potentially wide application to deciphering complex biological systems. (*The lecture will be given in Japanese with English slides. The discussion could be in either English or Japanese.)
References:
[1] Tajima, S., Yanagawa, T., Fujii, N., & Toyoizumi, T. Untangling brain-wide dynamics in consciousness by cross-embedding. PLOS Computational Biology, 11(11), e1004537:1-28, (2015).
[2] Tajima, S., & Kanai, R. Integrated information and dimensionality in continuous attractor dynamics. Neuroscience of Consciousness, (in press). Preprint: arXiv:1701.05157 [q-bio.NC]

第5回

日時: 6月8日(木) 14:40~16:10
会場: 情報科学研究科A棟 A109講義室
内容: 講演会
講演者: 宮本裕一郎先生(上智大学 理工学部 情報科学科 准教授)
講演題目: ブラックボックス最適化とその手法紹介
概要: 数理最適化問題のうち,目的関数がブラックボックス関数であるものをブラックボックス最適化問題という. ブラックボックス最適化はシミュレーションのパラメーター調整などに使われ,近年その重要性が増している. 今回は,ブラックボックス最適化問題に対する様々な手法(解法)を紹介し,その特徴を概観する. 簡単な計算実験結果なども紹介する.

過去の情報数理学セミナー

第1回

日時: 4月27日(木) 14:40~16:10
会場: 大阪大学 吹田キャンパス 応用物理学 講義棟 P1-311
内容: 安全教育講演Ⅰ
司会: 吉川 裕之 助教
概要:

博士前期・後期課程の教育・研究において、またその後の社会活動においても重要な事柄である「安全」について、その概略を講義する。

  1. 14:40~15:10 板崎 徳禎 「情報セキュリティ」
  2. 15:10~15:40 吉川 裕之 「安全全般に関して、化学実験・生物実験」

第2回

日時: 5月11日(木) 14:40~16:10
会場: 大阪大学 吹田キャンパス 応用物理学 講義棟 P1-311
内容: 安全教育講演ⅠI
概要:

博士前期・後期課程の教育・研究において、またその後の社会活動においても重要な事柄である「安全」について、その概略を講義する。

  1. 14:40~15:10 小西  毅 「電気、電子機器、電磁波、赤外・紫外光、レーザー光」
  2. 15:10~15:40 木村 吉秀 「各種機械・工具・工作機械、放射線、粒子線」

第3回

日時: 5月18日(木) 13:00~15:40
会場: 情報科学研究科A棟 A109講義室
内容: 博士論文中間発表会
講演者: Md Sohel Ahmed(森田研) 13:00~13:30
講演題目: Earthquake Disaster Management Analysis in Dhaka
概要:  A massive earthquake has been forecast for Bangladesh in the near future due to its position at the junction of three continental plates. Should it occur, the earthquake is expected to be devastating, particularly in the capital city of Dhaka, due to the large number of high-rise buildings constructed on relatively unstable ground and the high population density of the metropolitan area. In this paper, we use statistics-based methodology to identify correlations among the occupancy classifications, emergency response, and population density. Multiple regression analysis indicates that the level of satisfaction regarding disaster management is fairly low with respect to emergency response and medical care.
講演者: 滝本直也(森田研) 13:30〜14:00
講演題目: The Black-box Optimization via Multi-points Criterion
概要: The optimization of expensive-to-evaluate functions generally relies on meta model based exploration strategies. Many global optimization algorithms used in the field of computer experiments are based on Kriging method (Gaussian process regression). They proceed by choosing iteratively the point maximizing a criterion which balances the predicted value and the uncertainty. Then, multi-points sampling methods can achieve an effective search by less iteration. Finally, we investigate the multi-point criterion which aims at distributing sampling points depending on the number of processors which enable us to evaluate the different settings of simulations simultaneously.
講演者: Jian Yang (鈴木研) 14:00〜14:30
講演題目: A Study on A Framework for A Self-Organizing Agent Model Tackling Multi-Target Detection and Gravitation
概要: Swarm intelligence is an emergent property that arises from a group of locally interacting individuals. It is often observed in self-organized "intelligent" systems where autonomy, emergence and distributed functions replace control, pre-programming and centralization. Examples of such systems range from bacterial colonies to insect colonies, fish schools and bird flocks, which demonstrate intelligent behaviors including collective foraging, nest building and pattern formation (e.g., biofilm formation by bacteria). Inspired by how a group of individuals in nature exhibit such intelligent behavior, in this research we investigate the design of a framework for such self-organizing agent model for multiple possible applications, including drone clusters' autonomic controlling, nanoscale drug delivery by bio-nanomachines and even load balancing for networks. Here we define the rules by which agents interact to meet the application goals. We focus on a swarm of agents that interact by secreting signals called attractants and repellents to detect and localize to targets that may exist in the domain of our concern. This multi-target detection and gravitation problem was investigated in our previous work, in which, agents release both attractants and repellents to induce gravitational forces toward target locations. In the ongoing work, first we examine three modified approaches by which agents coordinate their behavior through releasing (1) attractants, (2) repellents or (3) both attractants and repellents. We examine the primary behavior of agents in these approaches and evaluate the performance. Second, under different initial spatial distributions of the agents over the entire domain, interesting behaviors of the agents emerge including repetitive oscillations alternately among the targets' locations. These intrigue us to study the model using methodology of non-linearity. Third, we will step further to demonstrate the validity and quality of this framework when applied to multiple practical problems, for example, the drone clusters' target detecting process. This research is also expected to shed light on performance optimization process of the model according to different requests in different applications. A comprehensively exhaustive method over parameter space should be averted.
講演者: Juan Lorenzo Hagad(沼尾研) 14:40~15:10
講演題目: Learning Robust Representations of Mental Stress Using Spatio-Temporal Features of Physiological Signals and Deep Learning
概要: Driven by advancements in wearable technology, physiological signals are proving to be invaluable tools able to unobtrusively gather insights about a wide range of physical and mental states. For instance, consumer-grade sensors can now track physiological signs associated with stress and potentially use these observations towards enhancing quality of life. Automated models stand to benefit the most due to their ability to learn from and process the vast amount of data that will result from the proliferation of wearables. However, the dynamics of physiological data has been proven to be inherently complex, and existing models still struggle to provide reliable performance especially across multiple users and practical application domains. Deep learning could theoretically close this gap. However, existing approaches are still unable to make full use of the capacity of certain deep learning structures such as convolutional neural networks (CNN). This is primarily due to the nature of the traditional parameters used in these models, which typically lack meaningful local spatial correlations. It would therefore be beneficial to construct models using data representations with meaningful spatial characteristics, provided that they draw from the same data sources as established features. This work proposes a more robust representation of authentic mental stress using a deep network structure trained on spatio-temporal representations of traditional physiological data features. Specifically, it applies CNNs to discover latent features from spectral and fractal representations of physiological signals. This is in contrast to the flat statistical feature vectors typically seen in existing methods. The final model will be trained on a combination of multiple physiological signals; however, priority will be placed on optimizing certain individual modalities. In the case of heart rate, the temporal nature of heart rate dynamics and psychological states should entail that CNNs should be able to discover salient features from heart rate time-frequency spectrograms. These raw spectrograms, while inherently more complex, should be more expressive than the statistical measures used to summarize entire windows of data, provided that data noise can be minimized. An additional challenge is the use of a small data-set, a common problem with many medical datasets, which necessitates the use of regularization methods to avoid over-fitting. Furthermore, in an effort to build towards a practical real-world model, the data collection methodology in this work prioritizes the elicitation of authentic mental stress through mentally fatiguing activities, as opposed to the frustration and physical stressors used by many traditional methodologies. Wearable sensors such as a chest-worn electrocardiogram (ECG) and a wristband sensor will be used to track the subject's physiological states while subjective user annotations will establish the ground truth stress levels. Finally, class activation maps (CAM) will be used to analyze important feature regions in the spectra which are most relevant to different stress-related states. This will be used to validate the model vis-a-vis known mechanisms about physiological stress signals or could potentially be used to improve the learning stages of the model.
講演者: Wasin Kalintha(沼尾研) 15:10~15:40
講演題目: Step Toward the Next Generation of Evolutionary Distance Metric Learning
概要: Many machine learning and data mining algorithms, such as classification, clustering and regression tasks, heavily rely on the distance metric for the input data patterns. Distance metric learning (DML) attempts to optimize a metric to improve data mining task. Many research studies on DML reiterate that the definition of distance between two data points substantially affects clustering tasks. Most of recent DML methods have been proposed based on Mahalanobis distance to improve the accuracy of clustering by formulating a penalty function for constraints into an objective function to learn a distance metric. However, these techniques sometimes lead to decreasing in the cluster validity index. Evolutionary distance metric learning (EDML) has been proposed to address this problem by directly improve cluster validity index as an objective function. Although EDML provides outstanding results over other semi-supervised clustering in many data set, it can only perform a linear transformation, like most of the DML techniques, which yields insignificant to non-linear separable data. Moreover, this unifies view of evolutionary algorithm and data mining has only few work recently, this will be a challenging problem to solve. This study proposes a non-linear Evolutionary Distance Metric learning. First, we address the non-linear separable by Kernelized Evolutionary Distance MetricLearning. Specifically, DML method which provides an integration of kernelization technique with Mahalanobis-based DML and EDML. Therefore, the non-linear transformation of the distance metric can be performed while maintain the optimized cluster validity index by an evolutionary algorithm. Furthermore, rather than just apply to clustering, I would like to expand this work to all other data mining technique, e.g., classification, regression and so on.