What is Machine Learning

Machine Learning Algorithms

 

1.  Supervised learning (지도 학습)

Learns from being given "right answers"

used most in real-world applications , most rapid advacements and innovation

learn x(input) to y(output mapping)

give learning algortihm examples to learn

-particular type of Supervised learning Algorithm : Regression(회귀)

Regression : predict a number of infinitely many possible outputs

 - Training set : Data used to train the model

 - X = 'input' variable, feature

 - Y = 'output' variable, 'target' variable

 - m = number of training examples

- (X,Y) = single training example

- (X(i), Y(i)) = i th training example  ** 제곱이나 루트가 아니니 헷갈려하지 말것 Index일뿐

- 'y-hat' = estimate or prediction for y in ML ** if Y, This is a target : actual true value in the training set

- Model = function f (X -> f -> Y^)

- f_wb = f(x) = wx + b ( w,b : parameters, Coefficients, weights) (w = slope(기울기) , b = intercept(절편))

- y-hat은 선형회귀선 위의 점. 'X' Marks means y^(i)

 

How to represent model f ?

** Python - .shape를 입력하면 (행렬의 차원수, 하나의 원소에 포함된 element갯수)를 반환 ,

   따라서 .shape[0] = 행렬의 차원수, shape[1]= 하나의 원소에 포함된 element 갯수

 

Cost Function J_wb (비용함수, 손실함수) also called squared error cost function **w에 대한 함수 , f는 x에 대한 함수!!**

 - commonly used for linear regression 

Cost function

 y-hat = fx^(i) + b , that is

Goal is minimize the J(w,b) (해당 값이 낮을 수록 Good Model)

 

Gradient Descent (경사하강법)

Algorithm that can use to try to minimize any function , not just cost function for linear regression

여러개의 parameter가 존재할때 cost function을 minimize하기 위해 사용

 - Start with some w, b

 - Keep changing w, b to reduce J(w,b)

 - Until settle at or near a minimum ( minimum은 한개 보다 많을 수 있음)

 

How to get Gradient descent Algorithm

W(Updated Parameter) = w - (a *d/dw J(w,b)) ,  B(Updated Parameter) = b - (a*d/db J(w,b))

 - Repeat until convergence

 - a : learning rate 0<a<1 . if alpha large, very aggressive gradient descent procedure, small alpha; baby steps

   if learning rate too small, 매우 작은 변화를. Gradient descent may be slow

  elif learning rate too large, 매우 큰 변화. Gradient descent may : Overshoot(초과), never reach min. Fail to converge(모여들다) ,  diverge(갈라지다)

 - if W already minimum, New update W = w , d/dw(J_wb) = 0 (slope = 0)

  or get nearer a minimum, Derivative(미분) becomes smaller , Update steps become smaller , Can reach minimum     without decreasing learning rate a

 - d/dw J(w,b) : direction, size want to take

 - Simultaneously update w and b 

순서 주의!

미분한것을 풀어냄

To be more precise : Use 'Batch' gradient descent (Batch = 집단)

 - Every steps of gradient descent. 

 

 

 

-Second major type of Supervised learning Algorithm : Classification (분류)

Classification : predict categories (non- numeric) (ex. dog, cat) , small number of possible outputs

 

 

 

 

-2. Unsupervised learning (비지도 학습)

Given data that isn't associated with any output labels y

Find something interesting in ualabeled data

Data only comes wtih inputs X, but not output labels Y.

Algorithm has to find structure in the data.

 

-Clustering Algorithm 

Takes Data without labels and tries to automatically group them into clusters

Group simliar data points together (ex. Google news : related news) 

 

-Anomaly detection

Find unusual data points

 

-Dimensionally reduction 

Compress data using fewer numbers

 

 

3. Recommender systems

4. Reinforcement learning