Zhumatiy Viktor 　D	神経計算学	池田和司	杉本謙二	銅谷賢治	吉本潤一郎
題目： Adaptive Resolution Reinforcement Learning for Mobile Robot Control
概要： Reinforcement learning with function approximator is a promising approach to realizing automatic robot control in continuous state-space domains. Adaptive resolution reinforcement learning (e.g., Bernstein et al, 2010) is a subclass in it, relying on finite support kernel functions derived from experienced instances. While the theoretical convergence was proven, there is a practical issue still remaining: precise approximation to guarantee the theoretical convergence requires a prohibitive number of kernel functions with tiny supports. Indeed, it is difficult for the method to solve even cart-pole balancing problem with four-dimensional state space. To remedy this, we additionally assume the state space to be locally Euclidean to make locally linear state transition approximation possible and design a value-updating multi-step rule. We demonstrate the efficiency of our approach experimentally by solving the cart-pole balancing problem.

RODRIGUES ALAN DE SOUZA 　D	神経計算学	池田和司	湊小太郎	銅谷賢治	吉本潤一郎
題目：Neural mechanisms for model-free and model-based reinforcement strategies in humans performing a multi-step navigation task
概要： Humans can learn actions from scratch, or by using knowledge from past experiences. Reinforcement Learning (RL), a computational theory of adaptive optimal control, suggests two methods that resemble real human behavior: Model-Free (MF) method uses action value functions to predict future rewards based on current states, and Model-Based (MB) method uses a forward model to predict the future states reached by hypothetical actions. We tested whether humans utilize MF and MB strategies by having subjects perform a grid-sailing task whose goal was to move a cursor from a start position to a target by sequentially pressing three keys, each moved the cursor in a different direction. After one day of training, subjects were tested inside the fMRI scanner under three task conditions: (1) learning of new key-map and start-goal positions, (2) use of learned key-map for new start-goal positions, (3) well-learned action sequences. In half of the trials a response started immediately after the go signal, or after a delay period of 4~6 seconds. Analysis of reward acquisition revealed high performance in condition 3; condition 2 was significantly better in trials with a delay; condition 1 had the lowest performance. By calculating the posterior probability of whether MF or MB algorithms generate subjects' action selection, we showed their actions in conditions 1 and 2 could be explained by MF and MB models, respectively. The fMRI analysis of this delay period using condition 3 as the control task showed that condition 2 activated the left DLPFC, ventral premotor cortex, anterior basal ganglia and right posterior cerebellum, whereas task condition 1 activated the left dorsal premotor cortex, parietal and visual areas bilaterally. Analysis of the time course and signal intensity of activation in these areas showed the strongest activation in condition 2 in this anterior prefrontal-basal-ganglia-cerebellum network, a candidate to implement MB strategy.

三宅　正夫 M　1回目	像情報処理学
題目：視覚障がい者のための洋服等に関する色及び模様の情報提示システム
概要：近年、カラーユニバーサルデザインの概念が普及し、デザイン段階で色の見えをシミュレーションするなど、色弱者のための色情報提示に関する研究が盛んである。しかし、視覚障がい者、特に全盲の人に色情報を提示するシステムは少ない。既存のソフトウェアでは、画素ごとの部分的な色判定を可能としたものはあるが、対象物全体としてどのような色に見えるかを判定できるものはない。本研究では、視覚障がい者が日常生活でカメラを用いて、洋服などの色と模様に関する情報を同時に調べることができるシステムを提案する。 色情報はHSL表色系をもとに表し、撮影画像から得られる画素ごとの数値を平均化して、対象範囲全体としての色名を出力する。テクスチャ情報は空間フィルタリングを用いて取得し、洋服によく現われる模様の種類を出力する。 実験では撮影した試料画像について、提案手法を用いることによって、色と模様に関する情報を出力することができた。

米澤　智 M　1回目	ロボティクス
題目：指先接触面制御を用いた重さと滑りやすさ提示装置の開発
概要： 近年，ハプティックデバイスの遠隔操作やVR， エンターテイメントなどへ応用が期待されている．特に物体把持における触覚提示では重さや滑りやすさといった感覚の提示が重要になると考えられる．しかし，これまでのハプティックデバイスの制御では物体把持における人間の反応まで十分に考慮されていなかった．本研究では，物体把持において重要とされている初期滑りを接触面画像を用いて計測しながら，重さや滑りやすさを提示する手法を提案する．本発表では提案手法を用いたハプティックシステムの概要と試作機について述べる．

Alfattany, Sami Abdul Ghani A M 1回目	インタラクティブメディア設計学
題目：Query Optimization Using GPU
概要： in this presentation I will introduce my graduate research topic which is optimizing database query using GPU. I will show the growth of GPU vs CPU. and then I will explain how can we use GPU to optimize database queries. and then I will show my final goal of my research. I hope you will enjoy the seminar and hopefully it will be useful for developing systems.

CHAIKIJWATANA　PIRIYA M 1回目	応用システム科学
題目：論文紹介“Cooperative Network Coding-aware Routing for Multi-Rate Wireless Networks,” Proc. of IEEE INFOCOM 2009
概要：Network coding is an excellent technique to improve network throughput in wireless networks. To fully exploit the performance gain brought by network coding, coding-aware routing has been studied. With this method, a transmission route that creates more coding opportunities is determined. However, more coding opportunities may not be a wise decision for multi-rate wireless networks. This paper proposes cooperative network coding (CNC) that exploits cooperative communication for improving coding opportunity. It also provides theoretical formulation for calculating the maximal throughput that can be achieved with CNC in multi-rate wireless networks. Moreover, CNC-aware routing is discussed in this paper. The performance evaluation shows that CNC-aware routing improves end-to-end throughput by comparing with existing method. In the final of the presentation, I will explain about my future work.

趙　凱 M　1回目	情報基礎学
題目：A Mutual Authentication Using Visual Secret Sharing
概要： This reasch investigates a mutual authentication scheme by making use of the visual secret sharing (VSS) scheme. The main concern of the investigated scheme is that it is easy for novice users to use the system. Novice users are seriously threatened by recently increasing phishing fraud. There are many technical countermeasures against phishing attacks, but those means are often too difficult for novice users to understand, set-up and utilize. In this reasch, a scheme is investigated which does not require special hardware, software, plug-ins and so on. Thanks to the characteristics of the VSS scheme, users are able to obtain minimum but practical security by using their accustomed web browsers only. This reasch discusses protocols which allow novice users protect themselves from phishing attacks.