hns - 日記自動生成システム - Version 2.10-pl1

先月 2011年01月 来月
1
2 3 4 5 6 7 8
9 10 11 12 13 14 15
16 17 18 19 20 21 22
23 24 25 26 27 28 29
30 31
HNS logo

2011年01月20日(木)

#1 統数研Talk終了

統数研の 第2回統計的機械学習セミナー のTalkが終わって, 京都に帰ってきました。スライドを 論文のページ に置いておきましたので, 興味のある方はどうぞ。
最後のFuture Workが大事だったのに言わなかったので, 少し悔やまれます。

当日, 部屋は後ろまで満杯で, 第1回よりも多かったらしく, Yee Whye 効果は凄いなぁ, と。
冨岡さんの質問が相変わらず鋭いなと思っていたのですが, 自分の番の場合は もともと鋭い質問を英語で受け答えしないといけないので, 日本語で言われれば 上手に説明できる場合でも英語に頭のリソースを割かれたりして, それが残念でした。

そういう意味では留学している人は有利だなというか, 一度も海外に留学していない僕 がこうやって英語で講演できる方が珍しいような気も。僕がNTTにいる限りは留学は (正社員優先なので)できないのですが..。 *1

当日は先日できたばかりの 赤池ゲストハウス に泊まってよいとの池田先生からの連絡だったので, 夜は歓迎会があるかもと思い(実際, 立川に行った), 泊まらせて もらうことにしました。
Akaike Guest Houseは超モダンで, 思ったよりもずっと広く快適でした。 僕は東京の出身なので, もし東京に来たらまず泊まることはないだろうと思うため, 貴重な経験をさせてもらってラッキーでした。以下写真。
基本的に共同研究者の滞在用なので, 簡単な台所があったりします。

部屋の入口部屋から見える
統数研の風景。
部屋の内部IH調理器具のある
台所付き。


*1: そういう意味では, 2008年に10日間だけとは言え, 僕を先にMITに行かせてくれたのは 非常に有難いなと思っています。

2011年01月26日(水)

#1 Yee Whye Teh Talk @ CS研

下のTalkは先ほど確定しました。
日時: 1月28日(金) 11:00-12:00
場所: NTTコミュニケーション科学基礎研究所 大会議室 (3F) (アクセス地図)
備考: NTTだけでなく, どなたでも参加できます。ご興味のある方はぜひおいで下さい。
正面玄関の受付で名前を書いていただき, 3Fへ上がって下さい。
Title: Modelling Genetic Recombination with Fragmentation-Coagulation Processes

Yee Whye Teh (Gatsby Computational Neuroscience Unit, UCL)

abstract:

Hudson's model of coalescent with recombination (also known as
ancestral recombination graph (ARG)) is a standard and well accepted
model of genetic recombination.  With the growing amount of population
genetics data, demand for probabilistic models to analyse such data is
strong, and the ARG is a very natural candidate.  Unfortunately, the
ARG posits very complex latent structures, and the observed data
cannot possibly restrict the posterior to a small space of latent
structures.  As a result posterior inference in the ARG is highly
intractable, and a large number of approximations and alternatives
have been proposed in the literature.

A large class of alternatives are based on hidden Markov models
(HMMs), which can be understood as approximating the tree-structure
genealogies at each point of the chromosome with a partition of the
observed haplotypes.  However due to the way HMMs parameterize
partitions using latent states, they suffer from significant
non-identifiability issues which can affect the quality of posterior
inference.

In this talk I will propose a new model for genetic recombination
based on Markov processes over partitions called
fragmentation-coagulation processes.  This model does not suffer from
the non-identifiability issues of HMMs though posterior inference is
somewhat more complex and requires development of a new auxiliary
variable Gibbs sampler for Markov jump processes.  I will discuss
issues of inference as well as various interesting properties of these
processes, and report some preliminary results on applying the model
to population genetic data.

ここ1~2日連絡がつかないので確定ではありませんが, 今週金曜日, 28日(金)に Yee Whye Teh にCS研に来ていただいて, 東京の話とは違う下の講演をしていただく 予定です。
NTTだけでなく, ぜひ外の方にも来ていただけるように現在調整しています。
場所はCS研大会議室, 15:00から(多分)です。
確定次第, すぐにここに情報を載せようと思います。


2011年01月29日()

#1 Yee Whye Talk終了

Talkは無事終了しました。
Yee Whyeの他の多くのTalkと違い, 今回は partition の分布としての 厳密さを妥協しない話で, 個人的には非常に良かったんじゃないかと思いました。
最前列で鋭い質問を連発されていた方が, 最初だけ一瞬誰だかわからなかったの ですが(普段と違って眼鏡を掛けてらしたので), 田中先生だった件。w

NTT外からも沢山来ていただき, 多分25〜30人くらいはいたのではないか と思います。僕も含めて, 刺激になれば良かったのではないかと思います。
現在進行中の話だったのでスライドは公開されないと思いますが, ポイントは

などでしょうか。
昼食時に聞いた話では, 実装には2ヶ月掛かっているのだそうです。凄い。

最後に帰る前に所長室で, Gatsbyで以前一緒だった上田さんと研究の話をして いましたが, 正直写真を撮りたいくらいでした。


以上、3 日分です。
タイトル一覧
カテゴリ分類
Powered by hns-2.10-pl1, HyperNikkiSystem Project