ゼロから始める深層強化学習(NLP2018講演資料)/ Introduction of Deep Reinforcement LearningPreferred Networks
Introduction of Deep Reinforcement Learning, which was presented at domestic NLP conference.
言語処理学会第24回年次大会(NLP2018) での講演資料です。
http://www.anlp.jp/nlp2018/#tutorial
ゼロから始める深層強化学習(NLP2018講演資料)/ Introduction of Deep Reinforcement LearningPreferred Networks
Introduction of Deep Reinforcement Learning, which was presented at domestic NLP conference.
言語処理学会第24回年次大会(NLP2018) での講演資料です。
http://www.anlp.jp/nlp2018/#tutorial
* Satoshi Hara and Kohei Hayashi. Making Tree Ensembles Interpretable: A Bayesian Model Selection Approach. AISTATS'18 (to appear).
arXiv ver.: https://arxiv.org/abs/1606.09066#
* GitHub
https://github.com/sato9hara/defragTrees
* Satoshi Hara and Kohei Hayashi. Making Tree Ensembles Interpretable: A Bayesian Model Selection Approach. AISTATS'18 (to appear).
arXiv ver.: https://arxiv.org/abs/1606.09066#
* GitHub
https://github.com/sato9hara/defragTrees
Robust Vehicle Localization in Urban Environments Using Probabilistic MapsKitsukawa Yuki
研究室ゼミでの論文紹介資料です。
Robust Vehicle Localization in Urban Environments Using Probabilistic Maps
Jesse Levinson, Sebastian Thrun
International Conference on Robotics and Automation (ICRA), 2010
FARIS: Fast and Memory-efficient URL Filter by Domain Specific MachineYuuki Takano
http://ytakano.github.io/
http://ieeexplore.ieee.org/document/7740332/
Uniform resource locator (URL) filtering is a fundamental technology for intrusion detection, HTTP proxies, content distribution networks, content-centric networks, and many other application areas. Some applications adopt URL filtering to protect user privacy from malicious or insecure websites. Some web browser extensions, such as AdBlock Plus, provide a URL-filtering mechanism for sites that intend to steal sensitive information.
Unfortunately, these extensions are implemented inefficiently, resulting in a slow application that consumes much memory. Although it provides a domain-specific language (DSL) to represent URLs, it internally uses regular expressions and does not take advantage of the benefits of the DSL. In addition, the number of filter rules become large, which makes matters worse.
In this paper, we propose the fast uniform resource identifier- specific filter, which is a domain-specific pseudo-machine for the DSL, to dramatically improve the performance of some browser extensions. Compared with a conventional implementation that internally adopts regular expressions, our proof-of-concept implementation is fast and small memory footprint.
This is a tutorial for implementing application level traffic analyzer by using SF-TAP flow abstractor.
http://sf-tap.github.io/
https://github.com/SF-TAP/
https://github.com/SF-TAP/flow-abstractor
https://www.usenix.org/conference/lisa15/conference-program/presentation/takano
http://ytakano.github.io/
10. 元論文の説明(1)
リーマン多様体 M は,すべての点 x M で
接ベクトル空間を持つ
リーマン多様体上の a と b を結ぶ
連続曲線 γ: [a, b] → M の長さは
length(γ) =
b
a
¦¦γ(t)¦¦dt
となり,任意の点 a, b 間を最小とする曲線の長さが
点 a, b 間の距離と定義され d(a, b) と表す
10
11. 元論文の説明(2)
1. 全ての y ∈M に対して τ−1
x (τx(y)) = y
2. ¦¦τx(y)¦¦ = d(x, y )
3. τx は原点に対する角度を維持する
4. Range(τx) = M
5. Range(τ−1
x ) = Tx M
ここで,すべての点 x M でリーマン多様体と
接ベクトル空間への写像を
τx: M → Tx M,τx
-1: Tx M → Mと定義する.
ただし,τx とτx
-1 は以下の条件を満たすとする.
11
12. 元論文の説明(3)
generateinitial layout(G)
while not donedo
for n ∈G do
x := position[n]
G := τx(G)
x := force directedplacement(n, G )
position[n] := τ−1
x (x )
end
end
すると,リーマン多様体上における
バネモデルレイアウトのアルゴリズムは以下のようになる
12
24. 球面幾何上へのグラフレイアウト(6)
直線 b b
点 b を通りベクトル a と
平行な直線 b b を考えると
この直線 b b は
x
y
z
a
b
b
x −bx
ax
=
y −by
ay
=
z −bz
az
と表せ,媒介変数 t を使うと
x = bx + ax t
y = by + ay t
z = bz + az t
となる
aと平行な直線
24
26. 球面幾何上へのグラフレイアウト(8)
b の導出(A)
x
y
z
a
b
b
ax (x - ax) + ay (y - ay) + az (z - az)
x = bx + ax t
y = by + ay t
z = bz + az t
を
へ代入すると t が求まる
t を
x = bx + ax t
y = by + ay t
z = bz + az t
へ代入すると b が求まる
26
27. 球面幾何上へのグラフレイアウト(9)
b の導出(B)
x
y
z
a
b
b
ax (bx + ax t −ax ) + ay (by + ay t −ay ) + az (bz + az t −az ) = 0
(a2
x + a2
y + a2
z )t = a2
x + a2
y + a2
z −ax bx −ay by −az bz
t =
a2
x + a2
y + a2
z −ax bx −ay by −az bz
a2
x + a2
y + a2
z
すなわち t は
となる.ただし単位球面なので
t = 1 −ax bx −ay by −az bz
となり,b つまり力の向きが求まる
27
29. 球面幾何上へのグラフレイアウト(11)
点 a, b のなす角ψ
点 a, b のなす角ψは
球面三角法の余弦定理より
となる
x
y
z a
b
N
ψ
cosψ= cosθa cosθb+ sin θa sinθbcos(φa −φb)
ψ= cos−1
(cosθacosθb+ sinθa sinθb cos(φa −φb))
29