Machine Learning & Deep Learning

AI vs. Human

AI 기술의 새로운 등장 Artificial Intelligence Machine Learning Deep Learning

강의 계획 강의 발표 평가 Machine Learning & Deep Learning 기초 Deep Learning Book Machine learning basic concepts Linear regression Logistic regression (classification) Multivariable (Vector) linear/logistic regression Neural networks 발표 Deep Learning Book 평가 중간고사: 45% 기말고사: 45% 출석: 10%

강의 교재 (발표) http://www.deeplearningbook.org/ 저자: Ian Goodfellow, Yoshua Bengio and Aaron Courville

Machine Learning Basics

Basic concepts What is ML? What is learning? What is regression? supervised learning unsupervised learning What is regression? What is classification?

Supervised/Unsupervised learning learning with labeled examples - training set

Supervised/Unsupervised learning

Supervised Learning

Supervised Learning Training data set

Supervised Learning AlphaGo

Types of Supervised Learning

Predicting final exam score : Regression

Predicting Pass/non-pass : binary classification

Predicting grades (A, B, …) : multi-labels classification

Linear Regression

Predicting exam score: regression

Regression data

Linear Hypothesis

Linear Hypothesis

Which hypothesis is better?

Cost function How fit the line to our (training) data

Cost function How fit the line to our (training) data

Cost function

ML Goal To minimize the cost function 𝑎𝑟𝑔𝑚𝑖𝑛 {𝑊,𝑏} 𝑐𝑜𝑠𝑡(𝑊,𝑏)

Hypothesis and Cost

Simplified hypothesis

What cost(W) looks like?

What cost(W) looks like?

What cost(W) looks like?

How to minimize the cost

Gradient descent algorithm

Gradient descent algorithm How it works?

Cost function: formal definition

Cost function: formal definition

Cost function: convex function

Cost function: convex function

Multi-variable linear regression

Predicting exam score: regression using two inputs (x1, x2)

Hypothesis

Cost function

Matrix notation

Matrix notation Hypothesis without b

Logistic regression

Classification

Classification

Classification

Linear Regression을 이용한 classification 0.5

Linear Regression을 이용한 classification 0.5

Linear Regression을 이용한 classification Linear regression hypothesis H(x) = Wx + b 1보다 크고 0보다 작을 수도 있음 0 ~ 1 사이의 값으로 scaling 필요

Logistic regression Sigmoid function (or Logistic function) 이용

Logistic hypothesis

Cost function

New cost function for logistic regression

New cost function for logistic regression y=1 일 때, H(x)=1로 예측 => cost = 0 H(x)=0로 예측 => cost =  y=0 일 때, H(x)=0로 예측 => cost = 0 H(x)=1로 예측 => cost = 

New cost function for logistic regression 𝐶 𝐻 𝑥 , 𝑦 =−𝑦 log 𝐻 𝑥 − 1−𝑦 log⁡(1−𝐻 𝑥 )

Gradient descent algorithm

Multinomial classification

Logistic regression HL(X) = WX z = H(X) g(z) : 0 ~ 1 사이의 값을 가지도록 HR(X) = g(HL(X)) z X 𝑌 : 0 ~ 1 W

Multinomial classification

Multinomial classification X W z 𝒀 : 0 ~ 1 X W z 𝒀 : 0 ~ 1 X W z 𝒀 : 0 ~ 1

Multinomial classification X W z 𝒀 𝑨 B X W z 𝒀𝑩 C X W z 𝒀 𝑪

Multinomial classification X W z 𝒀 𝑨 𝑤𝐴1 𝑤𝐴2 𝑤𝐴3 𝑤𝐵1 𝑤𝐵2 𝑤𝐵3 𝑤𝐶1 𝑤𝐶1 𝑤𝐶1 𝑥1 𝑥2 𝑥3 B X W z 𝒀𝑩 C X W z 𝒀 𝑪

Softmax function 𝒀 𝑨 𝒀 𝑩 X Y = 𝒀 𝑪 W 0.7 0.2 0.1 = 2.0 1.0 0.1 S(Y) = 2.0 1.0 0.1 0.7 0.2 0.1 Y = probability

Cost function 0.7 0.2 0.1 1 0 0 𝐷 𝑆, 𝐿 =− 𝑖 𝐿 𝑖 ∗log⁡( 𝑆 𝑖 ) S(Y) Cross-entropy S(Y) L = Y (실제값) 0.7 0.2 0.1 1 0 0 𝐷 𝑆, 𝐿 =− 𝑖 𝐿 𝑖 ∗log⁡( 𝑆 𝑖 )

Cross-entropy cost function 𝐷 𝑆, 𝐿 = 𝑖 {𝐿 𝑖 ∗−log⁡( 𝑆 𝑖 )} 예를 들어, 실제값 Y = L = [0 1]t

Logistic cost vs. Cross entropy cost 𝐶 𝐻 𝑥 , 𝑦 =−𝑦 log 𝐻 𝑥 − 1−𝑦 log⁡(1−𝐻 𝑥 ) 𝐷 𝑆, 𝐿 =− 𝑖 𝐿 𝑖 ∗log⁡( 𝑆 𝑖 )

Final cost (loss) function 𝐿𝑜𝑠𝑠 𝑊 = 1 𝑁 𝑖 𝐷(𝑆 𝑖 , 𝐿 𝑖 ) Training data

Gradient descent 𝑤𝑖←𝑤𝑖−𝛼 𝜕𝐿𝑜𝑠𝑠 𝑊 𝜕𝑤𝑖

Learning rate

Large learning rate: overshooting

Small learning rate: takes too long

Optimal learning rates ? Observe the cost function Check it goes down in a reasonable rate

Data preprocessing

Data preprocessing for gradient descent w2 w1

Data preprocessing for gradient descent w2 w1

Data preprocessing for gradient descent

Standardization

Overfitting

Overfitting The ML model is very good only with the training data set Memorization Not good at test set or in real use

Overfitting

Solutions for overfitting More training data Reduce the number of features Regularization

Regularization Try not to have too big number in the weights 특정 wi 가 커지는 경우

Regularization 𝐿𝑜𝑠𝑠 𝑊 = 1 𝑁 𝑖 𝐷(𝑆(𝑊𝑋 𝑖 +𝑏), 𝐿 𝑖 )+ 𝑤 𝑖 2 𝐿𝑜𝑠𝑠 𝑊 = 1 𝑁 𝑖 𝐷(𝑆(𝑊𝑋 𝑖 +𝑏), 𝐿 𝑖 )+ 𝑤 𝑖 2 : regularization strength 범위: 0 ~ 1