[Oracle Code Night] 機械学習入門 : ディープラーニングの基礎から転移学習まで

Transcript

1. 2021 2 26 ~
2. 2 Copyright © 2021 Oracle and/or its affiliates. • • • (NN) • • • • NN • •
3. 3 Copyright © 2021 Oracle and/or its affiliates. • • • • Etc. ?%
4. 4 Copyright © 2021 Oracle and/or its affiliates. • • • • K-means • SVM • etc. •
5. 5 Copyright © 2021 Oracle and/or its affiliates. : Wikipedia https://ja.wikipedia.org/wiki/%E7%A5%9E%E7%B5%8C%E7%B4%B0%E8%83%9E • ニューロン：脳の神経細胞 • シナプス：ニューロンと他のニューロンとの接合部分 • シナプス結合：ニューロンから別のニューロンへ信号が伝達される状態 • 活性化：ニューロンが受け取った刺激の大きさが特定の大きさを超える と、次のニューロンへと信号を伝達する。この状態を活性化と呼ぶ。 • 人工ニューロン(ノード)：ニューロンに相当。重みとバイアスを持つ。 • エッジ：シナプス結合に相当。ノード間を接続する。 • 重み、バイアス：シナプス結合の強度。ノード間接続時の信号の大きさの 計算に利用される。 • 活性化関数：ノード間の信号の大きさを計算するための関数。次ノードへ 伝える値を決める関数。
6. 6 Copyright © 2021 Oracle and/or its affiliates. ( ) 𝑦 = 𝑓 ෍ 𝑖=0 𝑛 𝑤𝑖 𝑥𝑖 + 𝑏 𝑥0 w0 + 𝑥1 w1 +… +𝑥𝑛 𝑤𝑛 + 𝑏 = b input x w n 𝑥0 𝑥1 𝑥2 w0 w1 𝑤2 y 𝑥𝑛 input output 𝑏 0 ⇒ 1 0 ⇒ 0 • ⇒ 1 ⇒ 0 • ⇒ 1 ⇒ 0 • ⇒ 1 ⇒ 0 input output • ⇒ 1 • ⇒ 0 → 1 → 0 → 1 → 2 → 1 → 3 → 1 → 3 → 7 1*2 + 0*1 + 1*3 +3 = 8 7 ⇒ 1 7 ⇒ 0 )
7. 7 Copyright © 2021 Oracle and/or its affiliates. • ( ) • • ( ) • ( ) • 𝑥0 𝑥1 𝑥2 w0 w1 𝑤2 y 𝑥𝑛 input output 𝑏 ) 65 80 70 A 45 60 98 B 65 55 80 C 65 70 63 1 3 2 40 38 43
8. 8 Copyright © 2021 Oracle and/or its affiliates. ReLU Tanh Sigmoid Step • 0 0 0 1 • NN (※ ) • 0 1 • 1 0 • -1 1 • 1 -1 • 0 0 0 • NN
9. 9 Copyright © 2021 Oracle and/or its affiliates. ReLU Tanh Sigmoid Step • 0 0 0 1 • NN (※ ) • 0 1 • 1 0 • -1 1 • 1 -1 • 0 0 0 • NN ( ) ( )
10. 10 Copyright © 2021 Oracle and/or its affiliates. Softmax • • • ( ) (※ Sigmoid ) • 0 1 • 1 • (※ )
11. 11 Copyright © 2021 Oracle and/or its affiliates. Softmax ReLU ReLU 0.8 Softmax ReLU ReLU 0.2 1 0 Cat or Tiger ? ( ) Cat Tiger 𝑥0 𝑥1 𝑥2 𝑏 𝑏 𝑏 𝑏 𝑏 𝑏 w w w w w w w w w w w w w w [Cat, Tiger] = [1, 0] 100 pixel 60 pixel w ( ) ( ) ( ) w ( )
12. 12 Copyright © 2021 Oracle and/or its affiliates. = ( ) = ( ) (w) Softmax ReLU ReLU 0.8 Softmax ReLU ReLU 0.2 1 0 Cat or Tiger ? ( ) Cat Tiger 𝑥0 𝑥1 𝑥2 𝑏 𝑏 𝑏 𝑏 𝑏 𝑏 w w w w w w w w w w w w w w [Cat, Tiger] = [1, 0] 100 pixel 60 pixel ( )
13. 13 Copyright © 2021 Oracle and/or its affiliates. ( ) w ( )
14. 14 Copyright © 2021 Oracle and/or its affiliates. ( ) ≒(( ) – ( ))2 ) • 100 • A 90 • B 120 ) • ⇒ [Cat, Tiger] = [1, 0] • A ⇒ [Cat, Tiger] = [0.8, 0.2] • B ⇒ [Cat, Tiger] = [0.2, 0.8] A (1 x log0.8 + 0 x log0.2) (log0.8) (0.09) 0.09 ⇒ ≒ log(( ) x ( )) B (1 x log0.2 + 0 x log0.8) (log0.2) (0.69) 0.69 ⇒ A (100 - 90)2 100 ⇒ B (100 - 120)2 400 ⇒
15. 15 Copyright © 2021 Oracle and/or its affiliates. (Stochastic Gradient Descent, SGD) (w) (w) (w) ( ) ) (w)
16. 16 Copyright © 2021 Oracle and/or its affiliates. Stochastic Gradient Descent(SGD) w momentumSGD SGD AdaGrad SGD RMSprop AdaGrad AdaDelta SGD Adaptive Moment Estimation(Adam) momentumSGD RMSprop (SDG)
17. 17 Copyright © 2021 Oracle and/or its affiliates.
18. 18 Copyright © 2021 Oracle and/or its affiliates. CNN ( ) • 28x28 / • 6 • 0 9 0 9 • 28x28 / • 1 •
19. 19 Copyright © 2021 Oracle and/or its affiliates. CNN 784 2 16 10 768 1 16 0 1 2 8 9 8 784 786 10 0 9 10 768 x 1 input input input input
20. 20 Copyright © 2021 Oracle and/or its affiliates. • • • • • • • • W W W
21. 21 Copyright © 2021 Oracle and/or its affiliates. w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w • ( ) • w ( ) • w
22. 22 Copyright © 2021 Oracle and/or its affiliates. w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w
23. 23 Copyright © 2021 Oracle and/or its affiliates.
24. 24 Copyright © 2021 Oracle and/or its affiliates.
25. 25 Copyright © 2021 Oracle and/or its affiliates. : http://image-net.org : https://cifar.ca/ • ImageNet • 160 x 160 pixel • 140 • • CIFAR-10 • 32 x 32 pixel • 5 • • MobileNetV2 • 1000 • 10 W W W ※MobileNetV2by Google) NN ?%