人工知能はどんな梦を见るか？
4 likes3,273 views
2013年10月のCMU用のDeep Learning解説資料
![具体的には？
Deep Learning = Learning Hierarchical Representations
Y LeCun
MA Ranzato
It's deep if it has more than one stage of non-linear feature
transformation
Low-Level
Feature
ニューロン
Deep Learning = Learning Hierarchical Representations
Mid-Level
High-Level
Trainable
Feature
Feature
Classifier
It's deep if it has more than one stage of non-linear feature
transformation
神経細胞をまねた
Low-Level
Mid-Level
Feature
High-Level
Feature
Feature
計算素子
Trainable
Classifier
学習
データの
(from LeCun and
Ranzato, ICML
Tutorial ’13)
共通性を探す
Feature visualization of convolutional net trained on ImageNet from [Zeiler & Fergus 2013]
Saturday, October 19, 13
YL
MA Ra](https://image.slidesharecdn.com/talk2-131018212849-phpapp01/85/-82-320.jpg)
![g = Learning Hierarchical Representations
Y LeCun
MA Ranzato
具体的には？
s more than one stage of non-linear feature
ow-Level
Feature
Deep Learning = Learning Hierarchical Representations
Mid-Level
High-Level
Trainable
Y LeCun
MA Ranzato
Feature
Feature
Classifier
It's deep if it has more than one stage of non-linear feature
transformation
Low-Level
Feature
ニューロン
Deep Learning = Learning Hierarchical Representations
Mid-Level
High-Level
Trainable
Feature
Feature
Classifier
It's deep if it has more than one stage of non-linear feature
transformation
神経細胞をまねた
Low-Level
Mid-Level
Feature
High-Level
Feature
Feature
計算素子
Trainable
Classifier
学習
onvolutional net trained on ImageNet from [Zeiler & Fergus 2013]
(from LeCun and
Ranzato, ICML
Tutorial ’13)
データの
共通性を探す
Feature visualization of convolutional net trained on ImageNet from [Zeiler & Fergus 2013]
Saturday, October 19, 13
YL
MA Ra](https://image.slidesharecdn.com/talk2-131018212849-phpapp01/85/-83-320.jpg)
人工知能はどんな梦を见るか？
- 1. 人工知能はどんな 夢を見るか？ Creator’s MeetUp, Oct 19th, 2013 東京大学 工学系研究科 木脇太一 Saturday, October 19, 13
- 3. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり やってること Saturday, October 19, 13
- 4. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり エンジニア やってること Saturday, October 19, 13
- 5. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり エンジニア ＋ やってること Saturday, October 19, 13
- 6. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり エンジニア ＋ 研究者 やってること Saturday, October 19, 13
- 7. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり エンジニア ＋ 研究者 やってること Saturday, October 19, 13
- 8. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり エンジニア ＋ 研究者 2 やってること Saturday, October 19, 13
- 9. 自己紹介 名前 木脇 太一（きわき たいち） ひととなり エンジニア ＋ 研究者 2 やってること これから話すよ！ Saturday, October 19, 13
- 10. 今日話すこと Saturday, October 19, 13
- 11. 今日話すこと Deep Learningって何？ Saturday, October 19, 13
- 25. コップを取る Saturday, October 19, 13
- 26. コップを取る Saturday, October 19, 13
- 27. コップを取る Saturday, October 19, 13
- 28. コップを取る Saturday, October 19, 13
- 29. コップを取る Saturday, October 19, 13
- 30. コップを取る Saturday, October 19, 13
- 31. コップを取る Saturday, October 19, 13
- 32. コップを取る Saturday, October 19, 13
- 33. コップを取る Saturday, October 19, 13
- 37. もっと他にも 別のアングル 別のものを取ってほ しい 更に難しいタスク 別のコップ Saturday, October 19, 13 買い物, etc.
- 38. もっと他にも 別のアングル 別のものを取ってほ しい 更に難しいタスク 別のコップ 買い物, etc. エンジニア100万人いても どらえもん１人作れない orz Saturday, October 19, 13
- 44. 階層性 Saturday, October 19, 13
- 48. 階層性 (Camel by Pablo Picasso) http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 49. 階層性 (Camel by Pablo Picasso) http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 50. 階層性 (Camel by Pablo Picasso) http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 51. 階層性 (Camel by Pablo Picasso) 層 階 http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 52. 階層性 (Camel by Pablo Picasso) 層 階 http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 53. 階層性 (Camel by Pablo Picasso) 層 階 http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 54. 階層性 (Camel by Pablo Picasso) 層 階 http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 55. 階層性 (Camel by Pablo Picasso) 層 階 http://www.mathworks.com/matlabcentral/ fx_?les/23253/4/gabor.png wikipedia Saturday, October 19, 13
- 56. 教師なし Saturday, October 19, 13
- 57. 教師なし Saturday, October 19, 13
- 58. 教師なし Saturday, October 19, 13
- 59. 教師なし Saturday, October 19, 13
- 60. 教師なし Saturday, October 19, 13
- 64. 汎用性 Saturday, October 19, 13
- 65. 汎用性 音声 Saturday, October 19, 13
- 66. 汎用性 音声 Saturday, October 19, 13 視角
- 68. 汎用性 目からの情報を音声を 処理する部分につなぎ かえると？ 音声 Saturday, October 19, 13 視角 “目が見える様になる”
- 71. 具体的には？ Saturday, October 19, 13
- 72. 具体的には？ Saturday, October 19, 13
- 73. 具体的には？ Saturday, October 19, 13
- 80. 具体的には？ ニューロン Deep Learning = Learning Hierarchical Representations YL MA Ra 神経細胞をまねた It's deep if it has more than one stage of non-linear feature transformation Low-Level Mid-Level Feature High-Level Feature Feature 計算素子 Trainable Classifier 学習 データの 共通性を探す Saturday, October 19, 13
- 81. 具体的には？ ニューロン Deep Learning = Learning Hierarchical Representations YL MA Ra 神経細胞をまねた It's deep if it has more than one stage of non-linear feature transformation Low-Level Mid-Level Feature High-Level Feature Feature 計算素子 Trainable Classifier 学習 データの (from LeCun and Ranzato, ICML Tutorial ’13) Saturday, October 19, 13 共通性を探す
- 82. 具体的には？ Deep Learning = Learning Hierarchical Representations Y LeCun MA Ranzato It's deep if it has more than one stage of non-linear feature transformation Low-Level Feature ニューロン Deep Learning = Learning Hierarchical Representations Mid-Level High-Level Trainable Feature Feature Classifier It's deep if it has more than one stage of non-linear feature transformation 神経細胞をまねた Low-Level Mid-Level Feature High-Level Feature Feature 計算素子 Trainable Classifier 学習 データの (from LeCun and Ranzato, ICML Tutorial ’13) 共通性を探す Feature visualization of convolutional net trained on ImageNet from [Zeiler & Fergus 2013] Saturday, October 19, 13 YL MA Ra
- 83. g = Learning Hierarchical Representations Y LeCun MA Ranzato 具体的には？ s more than one stage of non-linear feature ow-Level Feature Deep Learning = Learning Hierarchical Representations Mid-Level High-Level Trainable Y LeCun MA Ranzato Feature Feature Classifier It's deep if it has more than one stage of non-linear feature transformation Low-Level Feature ニューロン Deep Learning = Learning Hierarchical Representations Mid-Level High-Level Trainable Feature Feature Classifier It's deep if it has more than one stage of non-linear feature transformation 神経細胞をまねた Low-Level Mid-Level Feature High-Level Feature Feature 計算素子 Trainable Classifier 学習 onvolutional net trained on ImageNet from [Zeiler & Fergus 2013] (from LeCun and Ranzato, ICML Tutorial ’13) データの 共通性を探す Feature visualization of convolutional net trained on ImageNet from [Zeiler & Fergus 2013] Saturday, October 19, 13 YL MA Ra
- 84. 例：画像認識 IMAGENET 画像データベース Our model ● ● 2012年のコンテストで Max-pooling layers follow first, second, and fifth convolutional layers Deep Learningを使った The number of neurons in each layer is given チームが快勝 by 253440, 186624, 64896, 64896, 43264, 4096, 4096, 1000 (Krizhevsky et al., 2012) Saturday, October 19, 13
- 85. 例：音声認識 英語ー中国語 Jelly Beanの音声認識 同時通訳 (MS) (by Google) http://news.xinhuanet.com/english/photo/2012-10/26/c_131931521.htm Saturday, October 19, 13
- 87. 参考 Bengio, 2012 ICML Tutorial LeCun and Ranzato, 2013 ICML Tutorial A Fast Learning Algorithm for Deep Belief Nets, Hinton et al., 2006 人工知能学会誌 連載解説 Deep Learning The Future of Robotics and AI, Ng, Youtube Saturday, October 19, 13
- 88. Keywords Unsupervised Feature Learning Semi-supervised Learning Deep Belief Nets Restricted Boltzmann Machines Deep Auto Encoders Deep Neural Networks Saturday, October 19, 13