生命体工学セミナー

The brain of new-born babies has random connections for its neuronal system. Every neuron is connected with any other neurons. There is no obvious synaptic rule. Such randomness keeps increasing until the brain owner becomes several years old.

The randomness (synapse density) eventually reaches a peak. Thereafter, it begins to decrease and finally settle on a plateau.

The interpretation follows. Initially, connections are prepared as many as possible. However, through everyday life and experiences, it will be learned that not all such connections are necessary. Unnecessary and redundant ones are eliminated. The remaining ones are selected neural pathways that form a pruned slimy network.

In this talk, I will try to explain that such “skeleton brain” has a high ability of abstraction and generalization for cognitive science. This is called “Principle of Redundancy Reduction (PRR).　Furthermore, applications of PRR to artificial intelligence will be described, where a neural ?network algorithm automatically leans PRR adaptively, so as to deal with changing environments.

Rhythmic synchrony between different individuals often brings various emotions in our mind. We found that we sometimes feel phantom agency, if our motor rhythm synchronize with external mechanical rhythm of non-animate objects (e.g. robot). To verify this illusion, we have performed several behavioral and fMRI experiments. In this presentation, I would like to introduce our recent findings and discuss how rhythmic synchrony is transformed to the sense of agency in our brain.

The efficient emulation of complex Spiking Neural Networks (SNNs) presents two main challenges: a) Massive parallelism of neurons and synapses; b) Complex connectivity among them. In order to efficiently emulate neurons and synapses, a parallel multiprocessor architecture based on simple custom processors operating on a Single-Instruction Multiple-Data (SIMD) basis is proposed. The main properties of such architecture, namely, software-programmable algorithms, parallel emulation of neural parameters, serial emulation of synapses and local parameter memory, support any spiking model and real-time execution with excellent scalability. To cope with the interconnect complexity, a serial pipelined ring topology based on Address Event Representation (AER) is proposed. The AER transmission allows channel multiplexing while still providing real-time operation, due to the low-frequency of biological spikes. The serial ring topology presents a much better scalability than parallel or serial common buses, because it only uses local, point-to-point lines.

10:00-11:15: 講演1

Model Selection for Latent Variable Models via Asymptotic Bayesian Inference

Dr. Kouhei Hayashi (NII, Japan)

講演：日本語 (Slides in English)

世話人：柴田智広教授

11:20-12:00: 講演2

東京製鐵（株）のセミナー

多賀谷一也　総務部長代理 （東京製鐵株式会社）

講演：日本語

世話人：田中啓文教授

講師

Dr. Kouhei Hayashi (Project researcher at Global Research Center for Big Data Mathematics, NII, Japan)

題目

Model Selection for Latent Variable Models via Asymptotic Bayesian Inference

概要

Latent variable models, such as mixture models and hidden Markov models, represent high-dimensional observed data by low-dimensional latent variables. The latent space captures the hidden structure of data and thus the determination of its dimensionality, i.e. model selection is an important problem. For example, if the dimensionality is misspecified, predictive performance and/or interpretability would be degraded. In this talk, from a Bayesian perspective, a few standard methods for model selection will be introduced. In addition, a recently-developed asymptotic approximation technique will be explained.

略歴

Kohei Hayashi is a project researcher at Global Research Center for Big Data Mathematics, National Institute of Informatics. He received the B.Eng degree from Ritsumeikan University in 2007, and M.Eng and Ph.D degrees from Nara Institute of Science and Technology in 2009 and 2012, respectively. Before the current position, he was a JSPS Postdoc at the University of Tokyo. He is interested in relational data analysis and Bayesian probabilistic modeling.

講師

多賀谷一也　総務部長代理　（東京製鐵株式会社）

講演内容

• 地元若松区二島に工場を持つ東京製鐵（株）が環境都市北九州で環境を重視した地元とのかかわり
• 高炉法と電気炉法の違い
• 電炉鋼による製造プロセス
• スクラップに含まれる微量金属元素の有効活用や高度な生産・設備管理により高性能な電炉鋼材を開発に取り組む意義

略歴

1985年東京製鐵入社。 製鋼担当時には、大型の直流電気炉の導入、 新工場の建設計画やホットランに携わり、生産部長代理、 管理部長代理を経て、2014年より現職に至る。

I believe that only if it is possible to measure human's full body Motions and Emotions then we can naturally interact with robots and understand deeply human being. In this presentation, I will briefly introduce the research conducted at Waseda University, Takanishi-laboratory. In particular, I will focus my discussion on the Waseda Bioinstrumentation systems developed on purpose for the objective evaluation of motion and emotion in medical and robotics applications. I will show several medical applications in which the objective evaluation of motion is fundamental such as surgical gesture evaluation and assessment of older adults mobility. Furthermore, I will explain how motion analysis and gesture recognition can improve and promote a natural human-robot interaction showing several examples of interactions between the anthropomorphic musical robots developed in Waseda University and their human partners. Waseda Bioinstrumentation systems has been also used for emotion measurement and detection. In this presentation, I will focus on "laugh" because is a direct and universal form of social interaction associated with positive emotion and physical well-being but sometimes it is a symptom of behavioral or neurological disorders. Objectively measure physiological signals that play the most important role in laughter and integrate several non-invasive sensors as a unique tool specifically designed for medical application and robot interaction is a big challenge.

強化学習は環境からのフィードバックに基づき制御則を学習するための、また意思決定の脳内メカニズムを理解するための計算論的枠組みである。この講演では強化学習に関連する次の三つの重要な進捗である (1) 深層強化学習、(2) 報酬の設計論、(3) 制御と推論の相同性について概説する。最初のトピックは深層強化学習で、これは（順）強化学習と深層学習を組み合わせたもので、Google DeepMindが今年Natureに論文を発表して以来、高次元のセンサ空間から直接制御則を学習するための方法として注目を集めている。彼らはAtari 2600というコンピュータゲームにおいて人間よりもうまくゲームをプレイできることを示すなど非常に注目すべき成功を成し遂げたが、彼らが使用している技術は畳み込みネットワーク、経験の再利用、目標値の固定といった20年前のものである。また学習に時間がかかる、パラメータ設定に敏感であるという問題も報告されている。これに対し、我々は新しい単調増加でない非線形な活性度関数を導入したアプローチを紹介し、実験による従来の物よりも簡単かつ高速に学習できることを示す。次に、ロボットをどのように訓練するかという報酬の設計論について議論する。良い報酬を発見するために、本セミナーでは報酬に関する二つのアプローチを紹介する。一つは逆強化学習と呼ばれる枠組みで、観察された行動が最適であると仮定し、そこから報酬を推定する。ただし逆強化学習は逆問題であるため、一般には唯一解をもたない。この問題に対し、我々は最適制御則からのサンプルだけでなく、ベースラインとなる制御則からのサンプルを利用するように問題を修正したとき、問題を効率よく説くことのできる密度比推定を用いた逆強化学習を提案している。もう一つは内発的動機付けに基づいた強化学習で、タスク非依存の内発的報酬を用いて環境からのフィードバックを増強する枠組みであるが、内発的報酬は必ずしもタスクに有効なものとは限らない。我々は外発的報酬による制約を導入した強化学習法を提案している。我々の方法が報酬の設計に関して有用であることを示す。最後に強化学習分野における最適制御と最適推論の相同性について簡単に触れる。最適制御則を求めるためには一般に非線形偏微分方程式であるHamilton-Jacobi-Bellman (HJB)方程式を解く必要があるが、相同性に関する研究の進捗により、報酬関数に情報理論的な制約を導入することでHJB方程式が簡単になることが示された。結果として、いくつかの確率的最適制御問題を最適推論問題に変換することができ、最先端の強化学習アルゴリズムであるKL制御や経路積分強化学習などが導出された。時間の許す限り、我々のアプローチを紹介したい。

The presentation will cover the robotics research at IIT Delhi in two different laboratories, namely, Mechatronics Laboratory which Prof. Saha started in 2001, and the Programme for Autonomous Robotics Laboratory (a sponsored inter-disciplinary activity by three departments) started in 2010. The presentation will include a vision-based guidance of an industrial robot, an immersive environment for the tele-operation of an industrial robot, peg in tube operation using a force control algorithm, etc. The speaker will also explain the development of a truck simulator and a haptics device for medical simulation. Prof. Saha will then introduce a concept called RoCK-BEE (Robotics Competition Knowledge-based Education in Engineering). It is to encourage the students of engineering to take part in robotic competitions, e.g., ABU Robocon, to enjoy not only the competition but also the learning of the subject in a fun and effective way. It will be shown in the later part of his presentation that how the knowledge of robotics can be applied for the design optimization of many devices and processes used by the rural people, e.g., a carpet loom or a machine to clean it. This aspect is pursued by Prof. Saha under the banner called MuDRA or Multibody Dynamics for Rural Applications.

山田先生は走査トンネル顕微鏡を用いたナノ観察、ナノファブリケーション、ナノ磁性、単分子磁性などの精密測定で、新進気鋭の世界的に活躍されている研究者です。主に世界最小・単一分子磁気センサーの開発など、環境に配慮した次世代超省エネ・省資源デバイス創成にむけた新規材料の物性研究を展開されています。