Submit Search
自然言語処理のためのDeep Learning
318 likes
90,708 views
Yuta Kikuchi
1 of 143
Download now
Downloaded 1,292 times
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
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
More Related Content
PPTX
[DL輪読会]Graph R-CNN for Scene Graph Generation
Deep Learning JP
PPTX
TalkingData AdTracking Fraud Detection Challenge (1st place solution)
Takanori Hayashi
PPTX
Counterfaual Machine Learning(CFML)のサーベイ
ARISE analytics
PDF
方策勾配型強化学習の基礎と応用
Ryo Iwaki
PDF
LSTM (Long short-term memory) 概要
Kenji Urai
PPTX
データサイエンス概論第一 5 時系列データの解析
Seiichi Uchida
PPTX
【DL輪読会】The Forward-Forward Algorithm: Some Preliminary
Deep Learning JP
PDF
[DL輪読会]Understanding Black-box Predictions via Influence Functions
Deep Learning JP
[DL輪読会]Graph R-CNN for Scene Graph Generation
Deep Learning JP
TalkingData AdTracking Fraud Detection Challenge (1st place solution)
Takanori Hayashi
Counterfaual Machine Learning(CFML)のサーベイ
ARISE analytics
方策勾配型強化学習の基礎と応用
Ryo Iwaki
LSTM (Long short-term memory) 概要
Kenji Urai
データサイエンス概論第一 5 時系列データの解析
Seiichi Uchida
【DL輪読会】The Forward-Forward Algorithm: Some Preliminary
Deep Learning JP
[DL輪読会]Understanding Black-box Predictions via Influence Functions
Deep Learning JP
What's hot
(20)
PDF
TensorFlowで逆強化学習
Mitsuhisa Ohta
PPTX
[DL輪読会]Set Transformer: A Framework for Attention-based Permutation-Invariant...
Deep Learning JP
PDF
Rでisomap(多様体学習のはなし)
Kohta Ishikawa
PDF
深層生成モデルと世界モデル(2020/11/20版)
Masahiro Suzuki
PDF
最近のDeep Learning (NLP) 界隈におけるAttention事情
Yuta Kikuchi
PDF
最適輸送入門
joisino
PDF
研究室における研究・実装ノウハウの共有
Naoaki Okazaki
PDF
GAN(と強化学習との関係)
Masahiro Suzuki
PDF
Tensorflow Liteの量子化アーキテクチャ
HitoshiSHINABE1
PPT
AlphaGo Zero 解説
suckgeun lee
PDF
バンディットアルゴリズム入門と実践
智之 村上
PDF
Active Learning の基礎と最近の研究
Fumihiko Takahashi
PDF
[DL輪読会]Wasserstein GAN/Towards Principled Methods for Training Generative Adv...
Deep Learning JP
PDF
NLP2023 緊急パネル:ChatGPTで自然言語処理は終わるのか? 説明スライド
JunSuzuki21
PDF
実装レベルで学ぶVQVAE
ぱんいち すみもと
PDF
最近のKaggleに学ぶテーブルデータの特徴量エンジニアリング
mlm_kansai
PPTX
backbone としての timm 入門
Takuji Tahara
PPT
大学3年生の僕に伝えたいことをつらつらと
Toshinori Sato
PDF
時系列予測にTransformerを使うのは有効か?
Fumihiko Takahashi
PDF
深層学習の不確実性 - Uncertainty in Deep Neural Networks -
tmtm otm
TensorFlowで逆強化学習
Mitsuhisa Ohta
[DL輪読会]Set Transformer: A Framework for Attention-based Permutation-Invariant...
Deep Learning JP
Rでisomap(多様体学習のはなし)
Kohta Ishikawa
深層生成モデルと世界モデル(2020/11/20版)
Masahiro Suzuki
最近のDeep Learning (NLP) 界隈におけるAttention事情
Yuta Kikuchi
最適輸送入門
joisino
研究室における研究・実装ノウハウの共有
Naoaki Okazaki
GAN(と強化学習との関係)
Masahiro Suzuki
Tensorflow Liteの量子化アーキテクチャ
HitoshiSHINABE1
AlphaGo Zero 解説
suckgeun lee
バンディットアルゴリズム入門と実践
智之 村上
Active Learning の基礎と最近の研究
Fumihiko Takahashi
[DL輪読会]Wasserstein GAN/Towards Principled Methods for Training Generative Adv...
Deep Learning JP
NLP2023 緊急パネル:ChatGPTで自然言語処理は終わるのか? 説明スライド
JunSuzuki21
実装レベルで学ぶVQVAE
ぱんいち すみもと
最近のKaggleに学ぶテーブルデータの特徴量エンジニアリング
mlm_kansai
backbone としての timm 入門
Takuji Tahara
大学3年生の僕に伝えたいことをつらつらと
Toshinori Sato
時系列予測にTransformerを使うのは有効か?
Fumihiko Takahashi
深層学習の不確実性 - Uncertainty in Deep Neural Networks -
tmtm otm
Ad
Viewers also liked
(20)
PPTX
DL Hackerson 2
yyammt
PDF
言語モデル入門 (第二版)
Yoshinari Fujinuma
PDF
自然言語処理における意味解析と意味理解
Kanji Takahashi
PDF
言語モデル入門
Yoshinari Fujinuma
PPTX
TensorFlowで会話AIを作ってみた。
tak9029
PPTX
Chainer Meetup LT (Alpaca)
Jun-ya Norimatsu
PPTX
Chainer meetup
kikusu
PDF
Towards Chainer v1.5
Seiya Tokui
PDF
Chainer meetup20151014
Jiro Nishitoba
PDF
A Chainer MeetUp Talk
Yusuke Oda
PDF
LT@Chainer Meetup
Shunta Saito
PDF
Chainer Update v1.8.0 -> v1.10.0+
Seiya Tokui
PPTX
Chainerを使って細胞を数えてみた
samacoba1983
PDF
NVIDIA 更新情報: Tesla P100 PCIe/cuDNN 5.1
NVIDIA Japan
PDF
深層学習ライブラリの環境問題Chainer Meetup2016 07-02
Yuta Kashino
PDF
On the benchmark of Chainer
Kenta Oono
PDF
俺のtensorが全然flowしないのでみんなchainer使おう by DEEPstation
Yusuke HIDESHIMA
PDF
Tensor flow勉強会 (ayashiminagaranotensorflow)
tak9029
PDF
ヤフー音声認識サービスでのディープラーニングとGPU利用事例
Yahoo!デベロッパーネットワーク
PDF
Chainer Development Plan 2015/12
Seiya Tokui
DL Hackerson 2
yyammt
言語モデル入門 (第二版)
Yoshinari Fujinuma
自然言語処理における意味解析と意味理解
Kanji Takahashi
言語モデル入門
Yoshinari Fujinuma
TensorFlowで会話AIを作ってみた。
tak9029
Chainer Meetup LT (Alpaca)
Jun-ya Norimatsu
Chainer meetup
kikusu
Towards Chainer v1.5
Seiya Tokui
Chainer meetup20151014
Jiro Nishitoba
A Chainer MeetUp Talk
Yusuke Oda
LT@Chainer Meetup
Shunta Saito
Chainer Update v1.8.0 -> v1.10.0+
Seiya Tokui
Chainerを使って細胞を数えてみた
samacoba1983
NVIDIA 更新情報: Tesla P100 PCIe/cuDNN 5.1
NVIDIA Japan
深層学習ライブラリの環境問題Chainer Meetup2016 07-02
Yuta Kashino
On the benchmark of Chainer
Kenta Oono
俺のtensorが全然flowしないのでみんなchainer使おう by DEEPstation
Yusuke HIDESHIMA
Tensor flow勉強会 (ayashiminagaranotensorflow)
tak9029
ヤフー音声認識サービスでのディープラーニングとGPU利用事例
Yahoo!デベロッパーネットワーク
Chainer Development Plan 2015/12
Seiya Tokui
Ad
自然言語処理のためのDeep Learning
1.
自然言語処理のための Deep Learning 東京工業大学 奥村・高村研究室 D1
菊池悠太 @kiyukuta at 2013/09/11 Deep Learning for Natural Language Processing 13年9月28日土曜日
2.
13年9月28日土曜日
3.
2つのモチベーション - NLPでニューラルネットを - 言語の意味的な特徴を NN→多層×→pretraining→breakthrough
!! 焦って早口過ぎてたら 教えて下さい A yet another brief introduction to neural networks http://www.slideshare.net/yutakikuchi927/a-yet-another-brief-introduction-to-neural- networks-26023639 13年9月28日土曜日
4.
Neural networkベースの話 RBMとか苦しい 13年9月28日土曜日
5.
Deep Learning for NLP 13年9月28日土曜日
6.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN for NLP 13年9月28日土曜日
7.
Deep Learning for NLP 13年9月28日土曜日
8.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN 13年9月28日土曜日
9.
Deep Learning Unsupervised Representation
Learning 生データ 生データ 特徴抽出 学習器- 特徴抽出器 - 人手設計 答え! 答え!Deep Learning 従来 Deep Learning 13年9月28日土曜日
10.
Deep Learning 結論からいうと Deep Learningとは 良い初期値を(手に入れる方法を) 手に入れた 多層Neural
Networkです 13年9月28日土曜日
11.
Deep Learning ⽣生画像から階層毎に階層的な特徴を ラベル無しデータから教師なしで学習 13年9月28日土曜日
12.
生画像 高次な特徴は,より低次な特徴 の組み合わせで表現 13年9月28日土曜日
13.
= = = 低次レベルの特徴は共有可能 将来のタスクが未知でも 起こる世界は今と同じ 13年9月28日土曜日
14.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN 13年9月28日土曜日
15.
A yet another brief
introduction to Neural Networks 菊池 悠太 13年9月28日土曜日
16.
Neural Network 入力層x 隠れ層z 出力層y 13年9月28日土曜日
17.
入力層x 隠れ層z 出力層y 生データ,抽出した素性 予測 Neural Network 13年9月28日土曜日
18.
入力層x 隠れ層z 出力層y 例えば,手書き数字認識 784次元 10次元 MNIST (28*28の画像) 3!! [0.05, 0.05,
0.05, 0.40, 0.05, 0.05, 0.15, 0.05, 0.15, 0.05] 10次元の確率分布 (左から,入力画像が, 0である確率, 1である確率 ... 9である確率) 28*28= 784次元の数値ベクトル 13年9月28日土曜日
19.
入力層x 隠れ層z 出力層y Neuron 隠れユニットjの 入力層に対する重み W1 隠れユニットj 13年9月28日土曜日
20.
隠れユニットjの 入力層に対する重み W1 隠れユニットj 出力ユニットk 出力ユニットkの 隠れ層に対する重みW2 入力層x 隠れ層z 出力層y Neuron 13年9月28日土曜日
21.
入力層x 隠れ層z 出力層y W1 W2 行列で表現 層間の重みを行列で表現 13年9月28日土曜日
22.
Neural Networkの処理 - Forward
propagation - Back propagation - Parameter update 13年9月28日土曜日
23.
Neural Networkの処理 - Forward
propagation - Back propagation - Parameter update 13年9月28日土曜日
24.
Forward Propagation 入力層x 隠れ層z 出力層y 入力に対し出力を出す input x output
y 13年9月28日土曜日
25.
入力層x 隠れ層z 出力層y z = f(W1x
+ b1) 入力層から隠れ層への情報の伝播 非線形活性化関数f( ) tanh とか sigmoid とか f(x0) f(x1) f(x2) f(x3) f(x) = 13年9月28日土曜日
26.
入力層x 隠れ層z 出力層y z = f(W1x
+ b1) 入力層から隠れ層への情報の伝播 非線形活性化関数f( ) tanh,sigmoid reLU, maxout... f(x0) f(x1) f(x2) f(x3) f(x) = f( ) 13年9月28日土曜日
27.
入力層x 隠れ層z 出力層y z = f(W1x
+ b1) f( ) 入力層から隠れ層への情報の伝播 ⼊入⼒力力の情報を 重み付きで受け取る 隠れユニットが出す 出⼒力力値が決まる 13年9月28日土曜日
28.
入力層x 隠れ層z 出力層y z = f(W1x
+ b1) 入力層から隠れ層への情報の伝播 f( ) 非線形活性化関数f( ) tanh,sigmoid reLU, maxout... f(x0) f(x1) f(x2) f(x3) f(x) = 13年9月28日土曜日
29.
入力層x 隠れ層z 出力層y 出⼒力力層⽤用の 非線形活性化関数σ( ) タスク依存 隠れ層から出力層への情報の伝播 y =
(W2z + b2) 13年9月28日土曜日
30.
入力層x 隠れ層z 出力層y 出⼒力力層⽤用の 非線形活性化関数σ( ) タスク依存 隠れ層から出力層への情報の伝播 y =
(W2z + b2) 13年9月28日土曜日
31.
入力層x 隠れ層z 出力層y タスク依存の出力層 解きたいタスクによって σが変わる - 回帰 - 二値分類 -
多値分類 - マルチラベリング y = (W2z + b2) 13年9月28日土曜日
32.
実数 入力層x 隠れ層z 出力層y 回帰のケース 出力に値域はいらない 恒等写像でそのまま出力 (a) = a y
= (W2z + b2) 13年9月28日土曜日
33.
[0:1] 二値分類のケース 出力層は確率 σは0.0~1.0であって欲しい (a) = 1 1+exp(
a) Sigmoid関数入力層x 隠れ層z 出力層y y = (W2z + b2) 13年9月28日土曜日
34.
入力層x 隠れ層z 出力層y 多値分類のケース 出力は確率分布 各ノード0以上,総和が1 Softmax関数 sum( 0.2 0.7
0.1 )=1.0 (a) = exp(a) exp(a) y = (W2z + b2) 13年9月28日土曜日
35.
入力層x 隠れ層z 出力層y マルチラベリングのケース 各々が独立に二値分類 (a) = 1 1+exp(
a) element-wiseで Sigmoid関数 [0:1] [0:1] [0:1] y = (W2z + b2) 13年9月28日土曜日
36.
ちなみに多層になった場合 ... 出力層だけタスク依存 隠れ層はぜんぶ同じ 出力層 隠れ層1 隠れ層N ... 13年9月28日土曜日
37.
Forward Propagation 入力層x 隠れ層z 出力層y z =
f(W1x + b1) y = (W2z + b2) 13年9月28日土曜日
38.
Neural Networkの処理 - Forward
propagation - Back propagation - Parameter update 13年9月28日土曜日
39.
Back Propagation 入力層x 隠れ層z 出力層y 正解t NNが入力に対する出力の 予測を間違えた場合 正解するように修正したい 13年9月28日土曜日
40.
Back Propagation 入力層x 隠れ層z 出力層y 正解t NNが入力に対する出力の 予測を間違えた場合 正解するように修正したい 修正対象: 層間の重み z
= f(W1x + b1) ↑と,バイアス y = (W2z + b2) 13年9月28日土曜日
41.
Back Propagation 入力層x 隠れ層z 出力層y 正解t 誤差関数を最⼩小化するよう修正 E( )
= 1 2 y( ) t 2 E = K k=1 tk log yk E = t log y (1 t) log(1 y) E = K k=1 t log y + (1 t) log(1 y) いずれも予測と正解が 違うほど⼤大きくなる 13年9月28日土曜日
42.
Back Propagation 入力層x 隠れ層z 出力層y 正解t 出力ラベルと正解の差 ノードの誤差を計算 y =
y t 13年9月28日土曜日
43.
Back Propagation 入力層x 隠れ層z 出力層y 正解t ノードの誤差を計算 出力ラベルと正解の差 自分が情報を伝えた先の 誤差が伝播してくる z =
WT 2 yf (az) y = y t 13年9月28日土曜日
44.
Back Propagation 入力層x 隠れ層z 出力層y 正解t y =
y t z = WT 2 yf (az) 出力ラベルと正解の差 ノードの誤差を計算 自分の影響で上で発生した誤差 13年9月28日土曜日
45.
Back Propagation 入力層x 隠れ層z 出力層y 正解t 重みの勾配を計算 ⾃自分が上に伝えた 情報で発⽣生した誤差 En W2 = yzT En W1 =
zxT 13年9月28日土曜日
46.
Back Propagation 入力層x 隠れ層z 出力層y 正解t 重みの勾配を計算 ⾃自分が上に伝えた 情報で発⽣生した誤差 En W2 = yzT En W1 =
zxT 13年9月28日土曜日
47.
Neural Networkの処理 - Forward
propagation - Back propagation - Parameter update 13年9月28日土曜日
48.
Update parameters 入力層x 隠れ層z 出力層y 正解t 重みの更新 W1
= W1 En W1 W2 = W2 En W2 13年9月28日土曜日
49.
Update parameters 入力層x 隠れ層z 出力層y 正解t 重みの更新 W1
= W1 En W1 W2 = W2 En W2 -Gradient Descent -Stochastic Gradient Descent -SGD with mini-batch 修正するタイミングの違い 13年9月28日土曜日
50.
Neural Network の処理まとめ 13年9月28日土曜日
51.
Forward Propagation 入力層x 隠れ層z 出力層y z =
f(W1x + b1) y = (W2z + b2) 入力から予測 input x output y 13年9月28日土曜日
52.
Back Propagation 入力層x 隠れ層z 出力層y 正解t 誤差と勾配を計算 z
= WT 2 yf (az) y = y t En W2 = yzT En W1 = zxT 13年9月28日土曜日
53.
Update parameters 入力層x 隠れ層z 出力層y 正解t 勾配方向へ重み更新 W1
= W1 En W1 W2 = W2 En W2 13年9月28日土曜日
54.
ちなみにAutoencoder Neural Networkの特殊系 1. 入力と出力の次元が同じ 2.
教師信号が入力そのもの 入力を圧縮※1して復元 ※1 圧縮(隠れ層が入力層より少ない)でなくても,適切に正則化すればうまくいく 13年9月28日土曜日
55.
Neural Networkの特殊系 Autoencoder z =
f(W1x + b1) y = (W2z + b2) 13年9月28日土曜日
56.
Neural Networkの特殊系 Autoencoder y =
(W2z + b2) (a) = 1 1+exp( a) element-wiseで Sigmoid関数 マルチラベリングのケースに該当 画像の場合,各画素(ユニット)ごとに 明るさ(0.0:黒, 1.0:白)を判定するため 13年9月28日土曜日
57.
Autoencoderの学習するもの 13年9月28日土曜日
58.
Denoising Autoencoder add noise denoise 正則化法の一つ,再構築+ノイズの除去 13年9月28日土曜日
59.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN 13年9月28日土曜日
60.
Deepになると? many figures from http://www.cs.toronto.edu/~fleet/courses/cifarSchool09/slidesBengio.pdf 13年9月28日土曜日
61.
仕組み的には同じ 隠れ層が増えただけ 13年9月28日土曜日
62.
問題は初期化 NNのパラメータ 初期値は乱数 多層(Deep)になってもOK? 13年9月28日土曜日
63.
乱数だとうまくいかない NNはかなり複雑な変化をする関数なので 悪い局所解にいっちゃう Learning Deep Architectures
for AI (2009) 13年9月28日土曜日
64.
乱数だとうまくいかない NN自体が表現力高いので 上位二層分のNNだけで訓練データを 再現するには事足りちゃう ただしそれは汎化能力なし 過学習 inputのランダムな写像だが, inputの情報は保存している Greedy Layer-Wise Training
of Deep Networks [Bengio+, 2007] 13年9月28日土曜日
65.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN 13年9月28日土曜日
66.
2006年,ブレークスルー(Hinton+,2006) Greedy Layer-wise unsupervised pretraining 13年9月28日土曜日
67.
層ごとにまずパラメータを更新 層ごとに学習 13年9月28日土曜日
68.
層ごとにまずパラメータを更新 どうやって? Autoencoder !! RBMも [Bengio,2007] [Hinton,2006] 13年9月28日土曜日
69.
層ごとにまずパラメータを更新 どうなるの? 良い初期値を 得られるようになりました! Why does Unsupervised
Pre-training Help Deep Learning ? [Erhan+, 2010] [Bengio+, 2007] なぜpre-trainingが良いのか,諸説あり 13年9月28日土曜日
70.
Deep Learningとは 良い初期値を(手に入れる方法を) 手に入れた※1 Neural Network※2 つまり ※1
諸説あり Why does Unsupervised Pre-training Help Deep Learning ? [Erhan+, 2010] ※2 stacked autoencoderの場合 13年9月28日土曜日
71.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN 13年9月28日土曜日
72.
Autoencoder Neural Networkの特殊系 1. 入力と出力の次元が同じ 2.
教師信号が入力そのもの 訓練データ中の 本質的な情報を捉える 入力を圧縮して復元 13年9月28日土曜日
73.
Neural Networkの特殊系 圧縮ということは隠れ層は 少なくないといけないの? そうでなくても, 正則化などでうまくいく Autoencoder 13年9月28日土曜日
74.
Autoencoderの学習するもの これは,正確にはdenoising autoencoderの図 http://kiyukuta.github.io/2013/08/20/hello_autoencoder.html 13年9月28日土曜日
75.
Stacked Autoencoder 13年9月28日土曜日
76.
このNNの各層を, その層への⼊入⼒力力を再構築するAutoencoder として,事前学習 Stacked Autoencoder 13年9月28日土曜日
77.
Deep Learning Deep Learning概要 Neural
Networkふんわり Deepへの難しさ Pretrainingの光 Stacked Autoencoder , DBN 13年9月28日土曜日
78.
Deep Learning for NLP 13年9月28日土曜日
79.
画像処理のように Deeeeeeepって感じではない Neural Network-based くらいのつもりで 13年9月28日土曜日
80.
Deep Learning for NLP 13年9月28日土曜日
81.
Hello world. My name
is Tom. 2 4 MNIST 784 (28 x 28) 28 x 28= ??? size Input size ... ... ...... Image Sentence ... ... ... ... 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? 13年9月28日土曜日
82.
Hello world. My name
is Tom. 2 4 MNIST 784 (28 x 28) 28 x 28= ??? size Input representation ... ... ...... Image Sentence ... ... ... ... 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? 13年9月28日土曜日
83.
言い換えると 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? NLPでNNを使いたい 単語の特徴をうまく捉えた表現の学習 13年9月28日土曜日
84.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 13年9月28日土曜日
85.
任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation Keywords 13年9月28日土曜日
86.
任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation Keywords 13年9月28日土曜日
87.
任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation Keywords 13年9月28日土曜日
88.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 13年9月28日土曜日
89.
Word representation 自然言語処理における 単語の表現方法 ベクトル (Vector Space
Model, VSM) 13年9月28日土曜日
90.
単語の意味をベクトルで表現 単語 → ベクトル dog いろいろな方法 -
One-hot - Distributional - Distributed ... 本題 Word representation 13年9月28日土曜日
91.
One-hot representation 各単語に個別IDを割り当て表現 dog 辞書V 0 1 2 236 237 3043 : the :
a : of : dog : sky : cat ......... ......... cat 0 |V| 1 00... ...000... 0 1 00... 000... 0 スパースすぎて訓練厳しい 汎化能力なくて未知語扱えず 13年9月28日土曜日
92.
Distributional representation 単語の意味は,周りの文脈によって決まる Standardな方法 13年9月28日土曜日
93.
Distributed representation dense, low-dimensional,
real-valued dog k k |V| ... Neural Language Model により学習 = Word embedding 構文的,意味的な情報 を埋め込む 13年9月28日土曜日
94.
Distributed Word representation Distributed
Phrase representation Distributed Sentence representation Distributed Document representation recursive勢の一強? さて... 13年9月28日土曜日
95.
Neural Language Model Distributed
Word Representation の学習 13年9月28日土曜日
96.
Neural Language Model 言語モデルとは P(“私の耳が昨日からじんじん痛む”) P(“私を耳が高くに拡散して草地”)
はぁ? うむ 与えられた文字列の 生成確率を出力するモデル 13年9月28日土曜日
97.
Neural Language Model N-gram言語モデル 単語列の出現確率を
N-gram ずつに分解して近似 次元の呪いを回避 13年9月28日土曜日
98.
Neural Language Model N-gram言語モデルの課題 1.
実質的には長い文脈は活用できない せいぜいN=1,2 2. “似ている単語”を扱えない P(house|green) 13年9月28日土曜日
99.
Neural Language Model とは Neural
Networkベースの言語モデル - 言語モデルの学習 - Word Embeddingsの学習 同時に学習する 13年9月28日土曜日
100.
Neural Language Model 単語そのもの その単語のembedding |辞書|次元の確率分布 どの単語が次に 出てくるかを予測 A
Neural Probabilistic Language Model (bengio+, 2003) 13年9月28日土曜日
101.
Neural Language Model n語の文脈が与えられた時 次にどの単語がどのく らいの確率でくるか 13年9月28日土曜日
102.
Neural Language Model 似ている単語に似たembeddingを与えられれば, NN的には似た出力を出すはず 語の類似度を考慮した言語モデルができる 13年9月28日土曜日
103.
Ranking language model[Collobert
& Weston,2008] 仮名 単語列に対しスコアを出すNN 正しい単語列 最後の単語をランダムに入れ替え > となるように学習 他の主なアプローチ 13年9月28日土曜日
104.
Recurrent Neural Network
[Mikolov+, 2010] t番⽬目の単語の⼊入⼒力力時に 同時にt-‐‑‒1番⽬目の内部状態を⽂文脈として⼊入⼒力力 1単語ずつ⼊入⼒力力 出⼒力力は同じく 語彙上の確率率率分布 word2vecの人 他の主なアプローチ 13年9月28日土曜日
105.
Recurrent Neural Network
[Mikolov+, 2010] t番⽬目の単語の⼊入⼒力力時に 同時にt-‐‑‒1番⽬目の内部状態を⽂文脈として⼊入⼒力力 1単語ずつ⼊入⼒力力 出⼒力力は同じく 語彙上の確率率率分布 word2vecの人 他の主なアプローチ 13年9月28日土曜日
106.
word2vec https://code.google.com/p/word2vec/ 研究 進展 人生
→ 苦悩 人生 恋愛 研究 → 進展 他に... 13年9月28日土曜日
107.
word2vec https://code.google.com/p/word2vec/ 単語間の関係のoffsetを捉えている仮定 king - man
+ woman ≒ queen 単語の意味についてのしっかりした分析 13年9月28日土曜日
108.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 13年9月28日土曜日
109.
先ほどは,単語表現を学習するためのモデル (Bengio’s, C&W’s, Mikolov’s) 以降は,NNで言語処理のタスクに 取り組むためのモデル (結果的に単語ベクトルは学習されるが
おそらくタスク依存なものになっている) 13年9月28日土曜日
110.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 13年9月28日土曜日
111.
Collobert & Weston[2008] convolutional-‐‑‒way はじめに 2008年の論文 文レベルの話のとこだけ 他に Multi-task
learning Language model の話題がある 13年9月28日土曜日
112.
Collobert & Weston[2008] convolutional-‐‑‒way はじめに ここは 2層Neural
Network 入力 隠れ層 出力層 13年9月28日土曜日
113.
convolutional-‐‑‒way はじめに Neural Networkに 入力するために どうやって 固定次元に変換するか 任意の長さの文 Collobert &
Weston[2008] 13年9月28日土曜日
114.
convolutional-‐‑‒way Collobert & Weston[2008] 13年9月28日土曜日
115.
convolutional-‐‑‒way 単語をd次元ベクトルに (word embedding +
α) Collobert & Weston[2008] 13年9月28日土曜日
116.
convolutional-‐‑‒way 3単語をConvolutionして localな特徴を得る Collobert & Weston[2008] 13年9月28日土曜日
117.
convolutional-‐‑‒way 共通の重み( ) Collobert &
Weston[2008] 13年9月28日土曜日
118.
max( ) convolutional-‐‑‒way 各次元の最大値を取得 max( ) max(
) 次元固定のキモ Collobert & Weston[2008] 13年9月28日土曜日
119.
convolutional-‐‑‒way 固定次元の入力 任意サイズの入力 Neural Networkで学習 Collobert &
Weston[2008] 13年9月28日土曜日
120.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 13年9月28日土曜日
121.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 13年9月28日土曜日
122.
Richard Socher一派 recursive-‐‑‒way Recursive Neural
Network Recursive Autoencoder Parsing Natural Scenes and Natural Language with Recursive Neural Networks (ICML2011) Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
123.
Richard Socher一派 recursive-‐‑‒way Recursive Neural
Network Recursive Autoencoder Parsing Natural Scenes and Natural Language with Recursive Neural Networks (ICML2011) Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
124.
recursive-‐‑‒way Recursive Autoencoder 感情分布の推定 Semi-Supervised Recursive
Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
125.
recursive-‐‑‒way Recursive Autoencoder Distributed Representation Autoencoder 感情分布の推定 Semi-Supervised Recursive
Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
126.
recursive-‐‑‒way Recursive Autoencoder 2つを1つに圧縮用の共通の Autoencoderを再帰的に適用 Stacked Autoencoderは 階層毎に別のものを訓練 Semi-Supervised
Recursive Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
127.
recursive-‐‑‒way Recursive Autoencoder 感情分布の推定 Semi-Supervised Recursive
Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
128.
Richard Socher一派 recursive-‐‑‒way Recursive Neural
Network Recursive Autoencoder Parsing Natural Scenes and Natural Language with Recursive Neural Networks (ICML2011) Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions (EMNLP2011) 13年9月28日土曜日
129.
recursive-‐‑‒way Recursive Neural Network Parsing Parsing
Natural Scenes and Natural Language with Recursive Neural Networks (ICML2011) 13年9月28日土曜日
130.
recursive-‐‑‒way Recursive Neural Network Neural
Network Parsing Parsing Natural Scenes and Natural Language with Recursive Neural Networks (ICML2011) 13年9月28日土曜日
131.
recursive-‐‑‒way Recursive Neural Network Parsing
Natural Scenes and Natural Language with Recursive Neural Networks (ICML2011) :モデル出力の最大スコア :正解構文木のスコア Parsing 構文木中の全非終端ノードのsの総和 13年9月28日土曜日
132.
Socher+’s ほかの 13年9月28日土曜日
133.
Compositional semantics Semantic Compositionality
through Recursive Matrix-Vector Spaces (EMNLP2012) 各単語にベクトルと行列を割り当て 意味 隣接語への影響 not, very ... 13年9月28日土曜日
134.
Paraphrase 入力: 二つの文 出力: 二つの文が同じ意味か否か Dynamic
Pooling and Unfolding Recursive Autoencoders for Paraphrase Detection (NIPS2011) - Unfolding RAEで文の意味 - Dynamic poolingで任意の長さの2文を比べる 13年9月28日土曜日
135.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 単語の表現⽅方法としての 密で低次元な連続値ベクトル(word embedding) 13年9月28日土曜日
136.
Keywords Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation word embeddingsを学習するためのアプローチ ex) Bengio’s, recurrent, ranking 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? 13年9月28日土曜日
137.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation 文ごとに長さが異なるのを扱うアプローチ 13年9月28日土曜日
138.
Keywords 任意の⻑⾧長さの⽂文を⼊入⼒力力とするには?? 単語(句句や⽂文も)をどうやって表現する?? Distributed word representation -‐‑‒
convolutional-‐‑‒way -‐‑‒ recursive-‐‑‒way Neural language model phrase, sentence-‐‑‒level representation Recursiveな方は途中のphraseやsentenceに おける単語ベクトルも保存 13年9月28日土曜日
139.
2つのモチベーション - NLPでニューラルネットを - 言語の意味的な特徴を NN→多層×→pretraining→breakthrough
!! 焦って早口過ぎてたら 教えて下さい 具体例の説明が重くなりすぎたかも... 13年9月28日土曜日
140.
2つのモチベーション - NLPでニューラルネットを - 言語の意味的な特徴を NN→多層×→pretraining→breakthrough
!! 焦って早口過ぎてたら 教えて下さい 具体例の説明が重くなりすぎたかも... 13年9月28日土曜日
141.
Distributed (Word|Phrase|Sentence|Document) Representation Recursive Autoencoder一強 他の枠組みは? どうする? よりよい単語の表現 Neural Language
Model 意味?? Compositional Semanticsという タスク自体は,deep learning 以外でも最近盛ん 13年9月28日土曜日
142.
既存タスクへの応用 単語類似度,分類,構造学習... 要約,翻訳,推薦, ...? - 学習された単語のembeddingを追加素性に使う 他の方法は? 13年9月28日土曜日
143.
おわり 13年9月28日土曜日
Download