Likelihood

Sequence Generation

Conditional Sequence Generation

Maximizing Expected Reward

Policy Gradient

Conditional GAN

Abtractive Summarization

GAN + Autoencoder

Feature Extraction:

InfoGAN

VAE-GAN

BiGAN

Triple GAN

Feature Disentangle  v. 解开

J-S divergence proplem

Wasserstein GAN:

Earth Mover's Distance

Lipschitz Function

intractable adj. 棘手的 <==> difficult

f-divergence

exponential  adj. 指数

Theory behind GAN:

Divergence

KL Divergence

sample v. 抽样

J-S divergence

Unsupervised Conditional Generation

• Direct Transfomation
• Projection to Common Space:

CycleGAN:

Cycle consistency

GAN: Generative Adversarial Network

since sliced bread

Disciminator

Step 1: Fix G, update D

Step 1: Fix D, update G

Can Generator learn by itself?

Auto-encoder

Decoder = Generator

Can Discriminator generate?

Why CNN for image

Filter: 3×3

stride 步长

Feature Map

几个 Filter 几个 image

Max Pooling

Deep dream: Exaggerate

Deep style: 

Fat + Short vs. Thin +Tall

Deep ==> Modularization

Why Deep? Training Data 不够

GMM

Univerality Therorem

Analogy

End-to-end Learning

ReLU:

• Leaky ReLU
• Parametric ReLU

Maxout: 

ReLU is a special case of Maxout.

Learnable activation function

RMSProp:

Momentum:

RMSProp + Momentum ==> Adam

Regularization:

Dropout

Backpropagation

to compute gradients efficiently

Chain Rule:

dz/dx = dz/dy × dy/dx

• Forward pass
• Backward pass

Fully Connected Feedforward Network

Output Layer = Multi-class Classifier

Example

Step 1: Function Set

Step 2: Goodness of a Function

Cross Entropy

Step 3: Find the best Function(Gradient Descent)

no squarre error

Discriminative 有时优于 Generative（几率模型：Naive Bayes）

Multi-class Classification

Softmax ==> 0<y<1

Limitation of Logistic Regression

Classificaiton as Regression

Generative Model:

P(x) = 

Gaussian Distribution

Find Maximum Likelihood (mean*, covariance*)

All dimensions are independent ==> Naive Bayes Classifier

σ(z)=1/ (1+exp(-z))

On-line vs Off-line:

Momentum

Adagrad

RMSProp

Adam

Real Application

Adagrad

root mean square

g（gradient）: 偏微分

best step: |First derivative| / Second derivative

Stochastic Gradient Descent

Feature Scaling

Taylor Series

error 来源：bias 和 variance

mean: μ

variance: σ^2

s^2 是 σ^2的估测值

E[f*] = f^-: f* 的期望值

简单的模型 Variance 较小，简单的模型受数据波动影响小

复杂模型的 Bias 更小

Regularization ==> 使曲线变平滑6

Cross Validation

x_i: features

input: x^n

output: y^^n

function: f_n

Loss function L（function 的 function）: 

• Input: a function
• Output: how bad it is
• L(f) = L(w, b)

Step3: Best Function

f* = arg min L(f)

w*, b* = arg min L(w, b)

Gradient Descent:

• initial value w^0
• dL/dw|w=w^0
• 若 negative，增加 w
• 若 positive，减小 w
• η（learning rate）: 参数更新的幅度 -η(dL/dw|w=w^0)
• Local optimal: 局部最优
• global optimal: 全局最优
• 两个参数 w, b: 分别对 w, b 求偏微分
• ▽L: gradient 梯度

convex 凸面的 adj.

引入更复杂的函数：

x_cp^2

Overfitting

Back to Step 1: Redesign

• x_s = species of x
• 不同物种，不同 w, b
• δ(x_s = )
• = 1, if x_s = Pidgey
• = 0, otherwise

Back to Step 2: Rularization（调整）

不考虑 b 

select λ