トップ   一覧 単語検索 最終更新   ヘルプ   最終更新のRSS
***** 専攻談話会(セミナー) [#t67d9871]

集中講義などでいらっしゃる先生など、お心あたりのある方は専攻幹事 [[Tuan:http://www.is.titech.ac.jp/~tuan/]] (tuan at is.titech.ac.jp) までお知らせください。


**2013年度第7回 [#gdaa1d28]

日時: ''2013年12月16日(月)13:20 - 14:30 その後,簡単なお茶会をします''

会場: ''西8号館W棟10F1008コラボレーションルーム''

講演者: ''Ravi Montenegro (U. Mass, Lowell)''

題目: ''The Kruskal Count: how a card trick can help us break codes''

The Kruskal Count is a card trick popularized by Martin
Gardner. John Pollard independently used the same idea to construct a
tool that can be used to break some types of encryption. The Birthday
Paradox can explain why your ipod seems to play too many songs from the
same albums. Pollard again used similar intuition on cryptographic
problems. We discuss these curious results and explain how they are
examples of random walks, also known as Markov Chains.

**2013年度第6回 [#gdaa1d28]

日時: ''2013年12月4日(水) 16:30〜''

会場: ''大岡山西8号館 W棟 W1008室''

講演者: ''石崎 一明 (IBM東京基礎研究所)''

題目: ''静的型付き言語用Just-In-Timeコンパイラの再利用による、動的型付き言語用コンパイラの実装と最適化''

Web関連のソフトウェア開発において、JavaScript, PHP, Python,Rubyなどの動的型付き言語の人気が高い。これらの言語では、最初インタプリタによって実装され、その後言語の人気の高さとともに大きくなる高速化の要求に応えるためにコンパイラが開発される、という過程をたどることが多い。新たな言語のコンパイラ開発のために、言語毎にコンパイラを一から作成するのは実装にかかる労力が非常に大きい。我々は、この労力を削減し効率的な開発のために、既存の静的型付き言語Just-In-Timeコンパイラの再利用Repurpose-JIT compiler)というアプローチを用いて、Python言語のコンパイラを実装した。Repurpose-JIT compilerの実装においては、以下の2点が問題となり、高性能を得ることが難しかった。1)コンパイル対象とする言語の命令と再利用しようとするコンパイラの中間言語の命令の粒度の違い、2)高品質なコード生成にとって重要な型情報をソースコードから得ることの困難さ。本発表では、これらの問題点を解決する最適化を提案し、特に実行時情報を用いたコードの特殊化が有効であることを実験結果に基づいて紹介する。さらに、今回の経験から得られた、コンパイラ以外の性能に影響を与える要因についても紹介する。
This is joint work with Jose Castanos, David Edelsohn, Priya
Nagpurkar, Toshio Nakatani, Takeshi Ogasawara, and Peng Wu.


[1] Kazuaki Ishizaki, Takeshi Ogasawara, Jose Castanos, Priya
Nagpurkar, David Edelsohn, and Toshio Nakatani. Adding
dynamically-typed language support to a statically-typed language
compiler: performance evaluation, analysis, and tradeoffs, VEE 2012,

[2] Jose Castanos, David Edelsohn, Kazuaki Ishizaki, Priya Nagpurkar,
Toshio Nakatani, Takeshi Ogasawara, and Peng Wu. On the Benefits and
Pitfalls of Extending a Statically Typed Language JIT Compiler for
Dynamic Scripting Languages, OOPSLA 2012, 2012

**2013年度第5回 [#gdaa1d28]

日時: ''2013年11月7日(水) 13:20-14:50''

会場: ''大岡山西8号館W棟 W809号室''

講演者: ''鈴木大慈 氏(東京工業大学 数理・計算科学専攻)''

題目: ''マルチプルカーネル学習およびガウス過程事前分布を用いたスパース加法モデル推定''

Multiple Kernel Learning (MKL) およびガウス過程事前分布を用いたMKLのベイズ的変種についてその統計的性質を議論する.MKLは画像認識の分野などで成功を収めているが,オリジナルのL1正則化を用いるよりもL1とL2の中間的な正則化を用いたものが多くの問題で高い精度を出すことが実験的に報告されている.そこで,L1とL2の中間的な正則化としてelasticnet型の正則化やLpノルムを考えた場合に,それら中間的な正則化がどのように汎化性能に影響を与えるかを理論的に考察する.一方,MKLのベイズ的変種を用いれば,デザインへの条件を課さないでも速い収束レートを導出することができることも紹介する.


[1] Taiji Suzuki, and Masashi Sugiyama: Fast learning rate of multiple
kernel learning: trade-off between sparsity and smoothness.
The Annals of Statistics, vol. 41, number 3, pp. 1381-1405, 2013.

[2] Taiji Suzuki: PAC-Bayesian Bound for Gaussian Process Regression and
Multiple Kernel Additive Model.
Conference on Learning Theory (COLT2012), 2012, JMLR Workshop and Conference
Proceedings 23: 8.1--8.20, 2012.

[3] Taiji Suzuki: Unifying Framework for Fast Learning Rate of Non-Sparse
Multiple Kernel Learning.
Advances in Neural Information Processing Systems 24 (NIPS2011).

**2013年度第4回 [#gdaa1d28]

日時: ''2013年7月24日(水) 13:20-14:50''

会場: ''大岡山西8号館W棟 W1008号室''

講演者: ''寺嶋 郁二 氏(東京工業大学 数理・計算科学専攻)''

題目: ''クラスター変換とブレイド群''

クラスター変換の理論は,2002年に二人の数学者 S.Fomin と A.Zelevinsky によって導入されました.この講演ではクラスター変換を用いて,ブレイド群の表現を具体的かつ幾何的に構成できることを説明します.

**2013年度第3回 [#gdaa1d28]

日時: ''2013年7月18日(木) 13:20-14:50''

会場: ''大岡山西8号館W棟8F W809号室''

講演者: '' 野村 俊一 氏 (東京工業大学 数理・計算科学専攻)''

題目: ''アクチュアリーの業務、試験、課題について''


**2013年度第2回 [#gdaa1d28]

日時: ''2013年5月29日(水) 15:30-''

会場: ''大岡山西8号館W棟10F コラボレーションルーム(W1008)''

講演者: '' Ian Piumarta氏 (Viewpoints Research Institute)''

題目: ''To trap a better mouse''

Edsger Dijkstra once said: "The tools we use have a profound and
devious influence on our thinking habits, and therefore on our thinking
abilities." When we write programs, the single most important tool we're
using is the programming language. When we become fluent in a
programming language, we tend to think of our solutions directly in
terms of its data structures and algorithms. Unfortunately there is no
such thing as a programming language that is good for everything. In our
lives as programmers we will inevitably encounter problems for which a
natural and elegant language has not yet been invented in which to
express a solution. We then have a choice: either think poor thoughts in
a 'general purpose' programming language, or invent new languages in
which to think much better thoughts. Which option we choose can mean the
difference between a successful project and a failed project. In this
talk I will review some of the reasons why language is important, show
which parts of both natural and artificial languages tend to be
malleable, and suggest some ways of thinking about languages that might
lead to better ways of modelling and implementing the languages in which
we program -- and, therefore, think.

**2013年度第1回 [#gdaa1d28]

日時: ''2013年4月18日(木) 13:20-14:50''

会場: ''西8号館W棟8階W809号室''

講演者: '' Anne Bouillard 氏 (ENS Paris, 東京工業大学 数理・計算科学専攻)''

題目: ''Exact Worst-case Delay for FIFO-multiplexing Tandems''

This talk is about computing the actual worst-case end-to end
delay for a flow in a tandem of network under some general assumptions
using the network calculus theory, that is originally based on the
(min,plus) algebra. Unfortunately, using those algebraic methods lead to
very pessimistic bounds. We will see that we can indeed model this
network as a mixed interger linear program and drastically improve those

**2012年度第4回 [#j4f0fd9f]

日時: ''2013年3月21日(木) 15:00-16:30''

会場: ''西8号館E棟10階E1011号室''

講演者: '' Martin Mu"ller 氏''

題目: ''Move Quality in Monte Carlo Simulation:
A Case Study using the Fuego Go Program''

Despite half a dozen years of intense research, designing effective
simulation policies for Monte Carlo Tree search in Go is still
considered something of a black art, and driven largely by trial and
error. Important ideas that have evolved include pattern- and
tactics-based playouts, simulation balancing, and several schemes to
dynamically modify simulation policies online. In this study, we take
an in-depth look at what happens when the Go program Fuego runs its
playouts. We develop several methods to evaluate the quality of moves
played in these simulations, and we evaluate the contribution of the
different components of Fuego's playout policy. We study the
distribution of both the number of blunders, or result-changing moves,
and the absolute loss - in terms of number of points - for many
variations of the Fuego playout policy. We use this study to identify
an improvement to the Fuego default policy.

(Joint work with Sumudu Fernando)

**2012年度第3回 [#j4f0fd9f]

日時: ''2013年3月8日(金) 16:00-17:00''

会場: ''西8号館W棟10階1008号室''

講演者: ''Benjamin Burton 氏(The University of Queensland)''

題目: ''Untangling knots using combinatorial optimisation''

Unknot recognition is the algorithmic problem of determining whether a
knot (i.e., a closed loop) in 3-dimensional space can be untangled.
It is a major unsolved problem as to whether this problem has a fast
polynomial time solution.  Here we present the first conclusive
algorithm for unknot recognition which, although exponential time in
theory, exhibits a clear polynomial time behaviour in exhaustive
practical experiments.  The algorithm makes significant use of
techniques from combinatorial optimisation, and uses a
branch-and-bound framework with linear programming steps.  We also
introduce the topological software package Regina (in which the
algorithm was implemented), and discuss the relevance of our results
to other difficult problems in computational topology.

This is joint work with Melih Ozlen.

講演後には Burton 氏との懇親会を予定しています。奮ってご参加ください。

**2012年度第2回 [#j4f0fd9f]

日時: ''2012年11月12日(月) 15:00-16:30''

会場: ''西8号館E棟10階1011号室''

講演者: ''S.R.K. Branavan 氏''

題目: ''Grounding Linguistic Analysis in Control Applications''

Natural languages are the medium in which the majority of humanity's collective knowledge is recorded and communicated.  If machines were able to automatically access and leverage this information, they could effectively perform many tasks that are currently considered to be computationally intractable, thus requiring human involvement.  Today, the only way to infuse human knowledge into computational algorithms is to have humans in the loop - i.e., to manually encode the knowledge into heuristics, through annotations, or directly into the model structure itself.

Our ultimate goal is to automate this process, so that machines can access required knowledge directly from text.  One path to this goal is to perform a semantic interpretation of text by grounding the textual information in the objects, actions and dynamics of the physical world.  From a linguistic viewpoint, this grounding of language in control applications presents a very natural notion of language semantics.  I.e., by allowing us to define semantics with respect to the control application, and avoid imposing subjective human notions of correctness.

In this talk, I will explore two aspects of the connection between language and control applications: first, how the semantic analysis of language can be driven by control performance; and second, how information from text can be leveraged to improve performance in complex control systems.  These two aspects are in fact complementary, and language analysis is central to them both.  Addressing these aspects jointly is particularly challenging.  However, as our empirical results show, connecting the semantic analysis of language to control applications allow us to leverage the synergy between the two to simultaneously learn both language analysis and application control with little or no prior knowledge.

**2012年度第1回 [#j4f0fd9f]

日時: ''2012年4月19日(木) 13:20-14:50''

会場: ''西8号館W棟8階W809号室''

講演者: ''Tuan Phung-Duc 氏 (東京工業大学 数理・計算科学専攻)''

題目: ''Markov Chain and Queue with State-dependent Transition Rates''

In this talk, we consider continuous time Markov chains with
state-dependent transition rates arising from either retrial queues or
infinite server queues. First, for multiserver retrial queues, we show
that in the case of three and four servers, the joint stationary
distribution of the number of busy servers and the number of retrying
customers can be expressed in terms of continued fractions.
Furthermore, we formulate the general case with any number of servers
by a level-dependent QBD process for which a memory saving algorithm
based on matrix continued fractions is presented. Next, we briefly
present a topic on infinite server queues with synchronized
abandonment, where we develop a new methodology to obtain either
closed form solution or recursion for the stationary distribution of
the number of busy servers.

**過去の談話会 [#db336cc6]