Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Deep genenergyprobdoc

1,931 views

Published on

This is explanation about Energy-Based GaN by Bengio

Published in: Data & Analytics
  • Be the first to comment

Deep genenergyprobdoc

  1. 1. Deep Directed Generative Models with Energy-Based Probability Estimation By Yoshua Bengio mabonki0725 ()1 June 16, 2017
  2. 2. (1) Alpha SL 64 RL Q- (2) AI (3) (4) VAE GAN Energy-Base ( ) IRL Figure:
  3. 3. Eθ(x) -NFAHJ x Eθ(x) x Pθ(x) = 1 Z(θ) exp (−Eθ(x)) Figure: x=0 =1 x= =0 3 / 11
  4. 4. EΘ(x) ˜Pθi (x) = 1 1 + exp −WT i x + bi PΘ(x) = 1 ZΘ i ˜Pθi (x) = 1 ZΘ eEΘ(x) EΘ(x) = i log 1 + e−(W T i x+bi) ! Eθ(x) = 1 σ2 xT x − bT x − i log 1 + eW T i x+bi 4 / 11
  5. 5. L(Θ, D′ ) = − 1 N N i=1 log Pθ(x(i) ) (5) PositivePhase( ) NegativePhase( ∂L(Θ, D′) ∂Θ = − 1 N N i=1 ∂ log Pθ(x(i)) ∂Θ = − 1 N N i=1 ∂ log exp(Eθ(x(i) )) Zθ ∂Θ (6) = − 1 N N i=1 ∂ log exp(Eθ(x(i))) ∂Θ + ∂ log 1 ZΘ ∂Θ (7) = 1 N N i=1 ∂Eθ(x(i)) ∂Θ − Ex∼Pθ(x) ∂Eθ(x) ∂Θ (8) ≈ Ex+∼PD(x) ∂Eθ(x+) ∂Θ P ositiveP hase − Ex−∼Pθ(x) ∂Eθ(x) ∂Θ NegativeP hase (9) 5 / 11
  6. 6. L(Θ, D′) x+ x− Pφ(y = 1|x) = σ(−Eφ(x)) = 1 1 + exp(−Eφ(x)) (10) P(y = 0) = p(y = 1) = 1 2 Positive Phase Negative Phase E(x,y)∼P (x,y) − ∂ log Pφ(y|x) ∂φ = E(x,y)∼P (x,y) − ∂ log Pφ(y = 1|x)yPφ(y = 0|x)(1−y) ∂φ (11) = − 1 2 Ex+∼PD(x) ∂ log Pφ(y = 1|x+) ∂φ + Ex−∼PΘ(x) ∂ log Pφ(y = 0|x−) ∂φ (12) = 1 2 Ex+∼PD(x) Pφ(y = 0|x+ ) ∂Eφ(x+) ∂φ − Ex−∼PΘ(x) Pφ(y = 1|x− ) ∂Eφ(|x−) ∂φ ≈ 1 4 Ex+∼PD(x) ∂Eφ(x+) ∂φ − Ex−∼PΘ(x) ∂Eφ(x−) ∂φ (13) 6 / 11
  7. 7. Deep Generative Model with Enegrgy-Based 2 Positive Phase Negative Phase DeepLearning Figure: 7 / 11
  8. 8. Deep Energy Model Positive Phase x+ Negative Phase x− Energy Energy Eθ(x) = 1 σ2 xT x − bT x − i log 1 + eW T i fφ(x)+bi (14) fφ(x) (7) ∂L(Θ,D′ ) ∂Θ Deep Energy Model ∂L(Θ, D′) ∂Θ ≈ Ex+∼PD(x) ∂Eθ(x+) ∂Θ P ositiveP hase − Ex−∼Pθ(x) ∂Eθ(x) ∂Θ NegativeP hase (15) 8 / 11
  9. 9. Deep Generative Model (Negative Phase ) DKL(Pφ(x)||Pθ(x)) = Ex−∼Pφ(x) − log Pθ(x− ) − H(Pφ(x)) (16) Negativ Phase ∂ ∂φ Ex−∼Pφ(x) − log Pθ(x− ) = ∂ ∂φ Ez−∼P i(z) − log Pθ(Gφ(z)) (17) = Ez∼P (z) ∂Eθ(Gφ(z)) ∂φ (18) ≈ 1 N N i=1 ∂Eθ(Gφ(zi)) ∂φ zi ∼ P(z) (19) Gφ(zi) Deep Network zi ∼ P(z) (1 ∼ −1) Negativ Phase H(Pφ(x)) ≈ αi H(N(µαi , σi)) = αi 1 2 log 2 exp(πσ2 αi ) (20) 9 / 11
  10. 10. Figure: Figure: 10 / 11
  11. 11. VAE GAN 1 (VAE+GAN VAE GAN / Controlable Text Generation by Salakhutdinov 2 Generator Discrimetor CNN Adversarial Neural Machine Translation Energy-Base GAN Energy 11 / 11

×