EECS 836: Machine Learning

Zijun Yao
Assistant Professor, EECS Department
The University of Kansas

1
Agenda

• Linear Regression model

• Model definition

• Loss function

• Optimizing parameters

2
 Supervised learning setup
• Given a collection of records (training set) Training
• Each record is characterized by a pair (x, y) set

• x: feature, attribute, independent variable

Learning
• y: target, label, dependent variable algorithm
• Goal
• Learn a model (function ) so that f (x) can Function
correctly predict 𝑦ො for the corresponding value of y
• Tasks x 𝑦ො
• Regression: to predict a continuous value 𝑦ො Feature Predicted
• Classification: to predict a category class 𝑦ො value of y

3
* x is bolded because it represents a set of features; y is not because it is just a value.
House price prediction - regression

Size of House
# of Bedrooms f Price of House
…….

4
 Linear regression

Independent variables (features) x Dependent variables (targets) y

• Given
• Data
• Corresponding labels

• Goal: find a continuous function that models the continuous points

5
 3 ML steps for linear regression

Step 1: Step 2: Step 3: pick

define a set goodness the best
of function* of function function

Define a model Measure the error Optimizing parameters

*A set of function means same model but with different values of parameter. 6
 Step 1: Model definition

Predicted value of y

Target 1st Feature 2nd Feature d-th Feature

A linear relationship between feature and target Data 7

 Step 1: Model definition

Bias: a fixed offset Weights: significance of each feature

Parameters

Predicted value of y

Target 1st Feature 2nd Feature d-th Feature

A linear relationship between feature and target Data 8

 Step 1: Define set of functions
Size of
house

𝑓 𝑥 = price 𝑦
𝑥𝑠𝑖𝑧𝑒 , 𝑥𝑏𝑎𝑡ℎ
# of bath

9
 Step 1: Define set of functions
Size of
house

𝑓 𝑥 = price 𝑦
𝑥𝑠𝑖𝑧𝑒 , 𝑥𝑏𝑎𝑡ℎ
# of bath

Linear Regression model

10
 Step 1: Define set of functions
Size of
house

𝑓 𝑥 = price 𝑦
𝑥𝑠𝑖𝑧𝑒 , 𝑥𝑏𝑎𝑡ℎ
# of bath

Linear Regression model

w and b are parameters (can be any value)

A set of f1:
f1 , f 2 
function f2:
With different values of parameters …… infinite 11
Step 1: A variant form of linear model

1st Feature 2nd Feature d-th Feature

12
Step 1: A variant form of linear model

Equivalence
by

1st Feature 2nd Feature d-th Feature

0st Feature 1st Feature 2nd Feature d-th Feature

13
Step 1: A variant form of linear model

0st Feature 1st Feature 2nd Feature d-th Feature

Prediction Parameters

Linear Regression model

Function of 𝑥 Features

14
Agenda

• Linear Regression model

• Model definition

• Loss function

• Optimizing parameters

15
 Step 2: Goodness of function
How good is a function? - measure the difference between predicted and true y

Model

A set of
function

f1 , f 2 

Training
Data

16
 Step 2: Goodness of function
How good is a function? - measure the difference between predicted and true y

Model function function

input: Output (scalar): label value:

A set of (𝟏)
𝒙𝒔𝒊𝒛𝒆 ∶ 4,043 sqft
function (𝟏)
𝒙𝒂𝒈𝒆 : 26 years 𝑦ො (1) = f(𝐱 (1) ) 𝑦 (1) 784,000

f1 , f 2 

Training (𝟐)
𝒙𝒔𝒊𝒛𝒆 ∶ 4,976 sqft
Data (𝟐) 𝑦ො (2) = f(𝐱 (2) ) 𝑦 (2) 724,900
𝒙𝒂𝒈𝒆 : 8 years
Suppose we have two house features in
this data: size and age Superscript means the data index
 Step 2: Goodness of function
How good is a function? - measure the difference between predicted and true y

Model function function

input: Output (scalar): label value:
Measure the difference
A set of (𝟏)
𝒙𝒔𝒊𝒛𝒆 ∶ 4,043 sqft
function (𝟏)
𝒙𝒂𝒈𝒆 : 26 years 𝑦ො (1) = f(𝐱 (1) ) 𝑦 (1) 784,000

f1 , f 2 

Training (𝟐)
𝒙𝒔𝒊𝒛𝒆 ∶ 4,976 sqft
Data (𝟐) 𝑦ො (2) = f(𝐱 (2) ) 𝑦 (2) 724,900
𝒙𝒂𝒈𝒆 : 8 years
Suppose we have two house features in
this data: size and age Superscript means the data index
Step 2: Measure error
How good is a function? - Loss function 𝐿

A set of Model
Input: a function Output: the loss - how far is
function f1 , f 2  and data prediction from true value
𝑛 Sum of square error (SSE)
2
Goodness of 𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑓 𝑥 (𝑖)
function f 𝑖=1
Sum over examples Estimated y based
on input function
Training 1
𝑛
2
Averaged by n, you have mean squared error (MSE) loss ෍ 𝑦 (𝑖) − 𝑓 𝑥 (𝑖)
Data 𝑛
𝑖=1

Loss function also called cost function and objective function

 Step 2: Measure error
How good is a function? - Loss function 𝐿
𝑛
Input: a function and data 2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑓 𝑥 (𝑖)
output: the loss - how far is 𝑖=1
prediction from true value Sum over examples
True y
Estimated y based
on input function
𝐿 𝑤𝑠𝑖𝑧𝑒 , 𝑤𝑎𝑔𝑒 , 𝑏
𝑛
2
(𝑖) (𝑖)
=෍ 𝑦 (𝑖) − 𝑏 + 𝑤𝑠𝑖𝑧𝑒 ∙ 𝑥𝑠𝑖𝑧𝑒 + 𝑤𝑎𝑔𝑒 ∙ 𝑥𝑎𝑔𝑒
𝑖=1

Loss function also called cost function and objective function

Step 2: Loss function

𝑛
2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑓 𝑥 (𝑖)
𝑖=1

Estimated y based on
input function

A simple case where only one feature is used to predict y 21

Step 2: Loss function

𝑛
2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑓 𝑥 (𝑖)
𝑖=1

A simple case where only one feature is used to predict y When there are 2 features to predict y 22
Step 2: Intuition of loss function
• Let’s use a simple case with only one feature
𝑛
2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑤∙𝑥 𝑖

𝑖=1
𝑓 𝑥 𝐿 𝑓

3 15
𝑤=1
𝑦 2 L 𝑓 10

1 5

𝑏=0
0 1 2 3 0 0.5 1 1.5
𝑥 𝑤
𝐿 𝑓 = (1−1)2 +(2−2)2 +(3−3)2 = 0
23
Step 2: Intuition of loss function
• Let’s use a simple case with only one feature
𝑛
2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑤∙𝑥 𝑖

𝑖=1
𝑓 𝑥 𝐿 𝑓

3 15

𝑤 = 0.5 10
𝑦 2 L 𝑓

1 5

𝑏=0
0 1 2 3 0 0.5 1 1.5
𝑥 𝑤
𝐿 𝑓 = (1−0.5)2 +(2−1)2 +(3−1.5)2 = 3.5
24
Step 2: Intuition of loss function
• Let’s use a simple case with only one feature
𝑛
2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑤∙𝑥 𝑖

𝑖=1
𝑓 𝑥 𝐿 𝑓

3 15

𝑦 2 L 𝑓 10

1 5
𝑤=0
𝑏=0
0 1 2 3 0 0.5 1 1.5
𝑥 𝑤
𝐿 𝑓 = (1−0)2 +(2−0)2 +(3−0)2 = 14
25
Step 2: Intuition of loss function
• Let’s use a simple case with only one feature
𝑛
2
𝐿 𝑓 =෍ 𝑦 (𝑖) −𝑤∙𝑥 𝑖

𝑖=1
𝑓 𝑥 𝐿 𝑓
Loss function L is
3 15 a concave

𝑦 2 L 𝑓 10

1 5
𝑤=0
𝑏=0
0 1 2 3 0 0.5 1 1.5
𝑥 𝑤
𝐿 𝑓 = (1−0)2 +(2−0)2 +(3−0)2 = 14
26
 Step 2: Intuition of loss function

One-feature case Two-feature case

15

L(𝑓)
𝐿 𝑓 10

0 0.5 1 1.5 𝑤1
𝑤2
𝑤

Loss function tracks the performance of model as parameters change

27
Agenda

• Linear Regression model

• Model definition

• Loss function

• Optimizing parameters

28
Step 3: Find the best function

𝐿 𝑤, 𝑏
𝑛
A set of Model 𝑖
2
=෍ 𝑦𝑖 − 𝑏 + 𝑤𝑠𝑖𝑧𝑒 ∙ 𝑥𝑠𝑖𝑧𝑒 + 𝑖
𝑤𝑎𝑔𝑒 ∙ 𝑥𝑎𝑔𝑒
function f1 , f 2  𝑖=1

Loss
Goodness of
function f

Training
Data

29
Step 3: Find the best function

𝐿 𝑤, 𝑏
𝑛
A set of Model 𝑖
2
=෍ 𝑦𝑖 − 𝑏 + 𝑤𝑠𝑖𝑧𝑒 ∙ 𝑥𝑠𝑖𝑧𝑒 + 𝑖
𝑤𝑎𝑔𝑒 ∙ 𝑥𝑎𝑔𝑒
function f1 , f 2  𝑖=1

Loss The “best” – the function

Goodness of gives minimum loss
function f
Optimizing Search 𝑤, 𝑏 to find
Parameters minimum 𝐿
Training
Data

30
Step 3: Find the best function

𝐿 𝑤, 𝑏
𝑛
A set of Model 𝑖
2
=෍ 𝑦𝑖 − 𝑏 + 𝑤𝑠𝑖𝑧𝑒 ∙ 𝑥𝑠𝑖𝑧𝑒 + 𝑖
𝑤𝑎𝑔𝑒 ∙ 𝑥𝑎𝑔𝑒
function f1 , f 2  𝑖=1

Loss The “best” – the function

Goodness of gives minimum loss
function f
𝑓 ∗ = 𝑎𝑟𝑔 min L 𝑓 Optimizing Search 𝑤, 𝑏 to find
𝑓
Parameters minimum 𝐿
𝑤 ∗ , 𝑏 ∗ = 𝑎𝑟𝑔 min L 𝑤, 𝑏
𝑤,𝑏
Training 𝑛
2
Data (𝑖) (𝑖)
= 𝑎𝑟𝑔 min ෍ 𝑦 (𝑖) − 𝑏 + 𝑤𝑠𝑖𝑧𝑒 ∙ 𝑥𝑠𝑖𝑧𝑒 + 𝑤𝑎𝑔𝑒 ∙ 𝑥𝑎𝑔𝑒
𝑤,𝑏
𝑖=1
31
 𝐿(𝑤)
Derivatives Tangent line (𝑤=1)

𝐿(𝑤)
• The derivative of loss function 𝐿 is

The sensitivity to change of loss function

with respect to a change in a parameter 𝑤

32
 𝐿(𝑤)
Derivatives Tangent line (𝑤=1)

𝐿(𝑤)
• The derivative of loss function 𝐿 is

• Partial Derivatives: Let be loss function of parameters

Function derivative with respect to one of

parameters 𝑤𝑖 , with the others held constant

33
 𝐿(𝑤)
Derivatives Tangent line (𝑤=1)

𝐿(𝑤)
• The derivative of loss function 𝐿 is

• Partial Derivatives: Let be loss function of parameters

• Gradients
Gradients is a vector consists of
partial derivative of each parameter

34
 𝐿(𝑤)
Derivatives Tangent line (𝑤=1)

𝐿(𝑤)
• The derivative of loss function 𝐿 is

• Partial Derivatives: Let be loss function of parameters

• Gradients How to reduce Loss function?

Subtract gradient from each
parameter w
Gradient is a direction that increase the value of 𝐿(𝒘)
35
Step 3: Gradient descent
∗
• Consider loss function 𝐿(𝑤) with one parameter w: 𝑤 = 𝑎𝑟𝑔 min
𝑤
𝐿 𝑤

➢ (Randomly) Pick an initial value w0 at time 0

𝑑𝐿
➢ Compute | 0
Loss 𝑑𝑤 𝑤=𝑤
𝐿 𝑤 Negative Increase w

Positive Decrease w

w0 w
36
Step 3: Gradient descent
∗
• Consider loss function 𝐿(𝑤) with one parameter w: 𝑤 = 𝑎𝑟𝑔 min
𝑤
𝐿 𝑤

➢ (Randomly) Pick an initial value w0 at time 0

𝑑𝐿
➢ Compute | 𝑡 𝑑𝐿
Loss 𝑑𝑤 𝑤=𝑤 𝑤 𝑡+1 ← 𝑤𝑡 −𝛼 |𝑤=𝑤 𝑡
𝐿 𝑤 𝑑𝑤

𝛼 is called
w0 𝑑𝐿 w1 “learning rate” w
−𝛼 |𝑤=𝑤 0 37
𝑑𝑤 Usually small, like 0.05
Step 3: Gradient descent
∗
• Consider loss function 𝐿(𝑤) with one parameter w: 𝑤 = 𝑎𝑟𝑔 min
𝑤
𝐿 𝑤

➢ (Randomly) Pick an initial value w0

𝑑𝐿
➢ Compute | 0 1 0
𝑑𝐿
Loss 𝑑𝑤 𝑤=𝑤 𝑤 ←𝑤 −𝛼 |𝑤=𝑤 0
𝐿 𝑤 𝑑𝑤
𝑑𝐿
➢ Compute |𝑤=𝑤 1 2 1
𝑑𝐿
𝑑𝑤 𝑤 ←𝑤 −𝛼 |𝑤=𝑤 1
𝑑𝑤
repeat iterations until convergence
Local global
minima minima
w0 w1 w2 wT w
38
 𝜕𝐿
Step 3: Gradient descent 𝛻𝐿 = 𝜕𝑤
𝜕𝐿
• How about two parameters? 𝜕𝑏 gradient
𝑤 ∗ , 𝑏 ∗ = 𝑎𝑟𝑔 min 𝐿 𝑤, 𝑏
𝑤,𝑏

➢ (Randomly) Pick an initial value for each parameter w0, b0

𝜕𝐿 𝜕𝐿
➢ Compute | 0 0, | 0 0
𝜕𝑤 𝑤=𝑤 ,𝑏=𝑏 𝜕𝑏 𝑤=𝑤 ,𝑏=𝑏

1 0
𝜕𝐿 𝜕𝐿
𝑤 ←𝑤 −𝛼 |𝑤=𝑤 0 ,𝑏=𝑏0 𝑏1 ← 𝑏0 − 𝛼 |𝑤=𝑤 0 ,𝑏=𝑏0
𝜕𝑤 𝜕𝑏
𝜕𝐿 𝜕𝐿
➢ Compute | 1 1, | 1 1
𝜕𝑤 𝑤=𝑤 ,𝑏=𝑏 𝜕𝑏 𝑤=𝑤 ,𝑏=𝑏

𝜕𝐿 𝜕𝐿
𝑤2 ← 𝑤1 −𝛼 |𝑤=𝑤 1 ,𝑏=𝑏1 2 1
𝑏 ← 𝑏 − 𝛼 |𝑤=𝑤 1 ,𝑏=𝑏1
𝜕𝑤 𝜕𝑏
repeat iterations until convergence
39
Step 3: Gradient descent
• Gradient of Linear Regression

40
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 41
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 42
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 43
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 44
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 45
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 46
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 47
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 48
 Step 3: Gradient descent
𝑓 𝑥 𝐿 𝑤, 𝑏 Find minimum!
High 𝐿

𝑤
Low 𝐿
𝑏

How does the loss get minimized by gradient descent

Slide by Andrew Ng 49
Step 3: Gradient descent
• Small gradient can slow down or halt the optimization
Loss
Very slow at
the plateau
Stuck at
saddle point
Stuck at local minima

𝜕𝐿 ∕ 𝜕𝑤 𝜕𝐿 ∕ 𝜕𝑤 𝜕𝐿 ∕ 𝜕𝑤
≈0 =0 =0

The value of the parameter w 50

 Step 3: Gradient descent – learning rate 𝜶

Monitor the loss at each iteration

- Search 𝛼 by order of magnitude
at first (like, 10−1 … 10−5 )
- Tune 𝛼 locally to achieve efficient
convergence

[Link]
Linear algebra review
• Vector in ℝ𝑑 is an ordered set of 𝑑 real values
• Matrix in ℝ𝑛×𝑚 is a 𝑛 by 𝑚 object with 𝑛 rows and 𝑚
columns
• Transpose

• Matrix production

52
4×2 2×3 4×3
Vectorization of linear regression
• Benefits of vectorization
• More compact equations
• Faster code (using optimized matrix libraries)
• Linear regression model:

• Let

• In vectorized form, linear regression model:

53
Vectorization of linear regression
• Consider the model for 𝑛 instance

• Let

ℝ(𝑑+1)×1 ℝ𝑛×(𝑑+1)
• In vectorized form, linear regression model:
54
Vectorization of linear regression
• For the loss function

One time calculation, without iterating through all data samples.

55
 Improving learning
• Feature scaling (or normalization)
• Ensure all features have similar scales
• Gradient descent would converge faster

56
[Link]
Feature standardization
• Rescale features to have zero mean and unit variance

• Let 𝜇𝑗 be the mean of feature 𝑗

• Let 𝑠𝑗 be the standard deviation of feature 𝑗

• Replace each value with

for 𝑗 = 1 … 𝑑 (not 𝑥0 )

• Must apply the same transformation for both training and testing instances

• Outliers can cause problems

57
 Regularization
• A method to control the complexity of model, avoid overfitting
• Why - address overfitting issues by keeping 𝑤 small
• How - Penalize for large value of 𝑤𝑗
• Can incorporate into the loss function
• Works well when we have a lot of features

𝑛 𝑑
2
Also called 𝐿2 -norm
𝐿 𝑓 =෍ 𝑦 𝑖 −𝑓 𝑥 𝑖 + 𝜆 ෍ 𝑤𝑗2
𝑖=1 𝑗=1

model fit to data regularization

o 𝜆 is the predefined hyperparameter to control the degree of regularization
o No regularization on 𝑤0 (bias 𝑏)
58
Summary
• Problem: estimate a real value
• Model: 𝑦ො = 𝑏 + 𝑤1 𝑥1 + 𝑤2 𝑥2 + ⋯ + 𝑤𝑑 𝑥𝑑
• Loss function: sum of square (SSE)
𝑛 𝑛
1
𝐿 𝐰, 𝑏 = ෍ 𝑙 𝑖 𝐰, 𝑏 = ෍ (𝑦 𝑖
− 𝑦ො (𝑖) )2
𝑛
𝑖=1 𝑖=1

• Optimize parameters by Gradient Descent method

• Choose a starting point
• Repeat
• Compute gradient
• Update parameters
59
Demo
• Use ML library
• [Link]

60