9/5/23

Learning from Data

Model “car”
Modeling & Computation Predict
Future Data
Train

Past Data
(perhaps with labels)

1 2

Machine Learning in Practice

Modeling Computation
Modeling

3 4

1
 9/5/23

Linear Models Linear Models

• Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house). • Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house).
• Prediction: 𝑓 𝐱 = 𝐱 % 𝐰 (e.g., housing price).

𝑥" # of bedrooms 𝑥" # of bedrooms 𝑤"

𝑥# # of bathrooms 𝑥# # of bathrooms 𝑤#
𝐱 = 𝑥$ square feet
𝐱 = 𝑥$ square feet 𝐰 = 𝑤$

⋮ ⋮ ⋮ ⋮ ⋮
𝑥! age of house 𝑥! age of house 𝑤!

5 6

Linear Models Linear Models

• Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house). • Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house).
• Prediction: 𝑓 𝐱 = 𝐱 % 𝐰 (e.g., housing price). • Prediction: 𝑓 𝐱 = 𝐱 % 𝐰 (e.g., housing price).

Question: How to find 𝐰?

• 𝑓 𝐱 = 𝑤"𝑥" + 𝑤#𝑥# + ⋯ + 𝑤! 𝑥!
• 𝑤", 𝑤#, ⋯ , 𝑤! : weights Price = $0.5M

Prediction:
Features of a House
𝑓 𝐱 = 𝐱% 𝐰
𝐱 ∈ ℝ!

7 8

2
 9/5/23

Linear Models Linear Models

• Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house). • Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house).
• Prediction: 𝑓 𝐱 = 𝐱 % 𝐰 (e.g., housing price). • Prediction: 𝑓 𝐱 = 𝐱 % 𝐰 (e.g., housing price).

Question: How to find 𝐰? Question: How to find 𝐰?

• Training features: 𝐱", ⋯ , 𝐱 & ∈ ℝ! . • Training features: 𝐱", ⋯ , 𝐱 & ∈ ℝ! .

• Training targets: 𝑦", ⋯ , 𝑦& ∈ ℝ. • Training targets: 𝑦", ⋯ , 𝑦& ∈ ℝ.
" #

⋯ • Loss function: 𝐿 𝐰 = ∑& %

'(" 𝐱 ' 𝐰 − 𝑦' .
&

totally 𝑛 houses

9 10

Linear Models Linear Models Are Not Expressive

• Feature vector: 𝐱 ∈ ℝ! (e.g., features of a house). Example: Given a person’s photo, predict her/his age.
• Prediction: 𝑓 𝐱 = 𝐱 % 𝐰 (e.g., housing price).
Age = 36
Question: How to find 𝐰?

• Training features: 𝐱", ⋯ , 𝐱 & ∈ ℝ! . Photo (features)

• Training targets: 𝑦", ⋯ , 𝑦& ∈ ℝ.
" #
• Loss function: 𝐿 𝐰 = ∑& %
'(" 𝐱 ' 𝐰 − 𝑦' .
&
• Least squares regression: 𝐰 ⋆ = min 𝐿 𝐰 .
𝐰

11 12

3
 9/5/23

Linear Models Are Not Expressive Traditional Approaches

Example: Given a person’s photo, predict her/his age.
SIFT Feature
Age = 36

Photo (features) HOG Feature

Question: Can we use linear regression?

• Linear regression assumes the target is a weighted average of every

pixel in the photo. LBP Feature

• Linear regression works poorly for age prediction.

13 14

Traditional Approaches Convolutional Neural Networks (CNNs)

SIFT Feature

HOG Feature Linear Model

“dog”

LBP Feature

15 16

4
 9/5/23

Applications of CNNs Applications of CNNs: Medical Diagnosis

• CNNs are suitable for image data. Example: Skin cancer diagnosis
• CNNs convert images to effective
representations. (Feature extraction.) • Input: an image.
• Outputs:
• Is it skin cancer?
• Applications: • Benign or malignant?
• Image/video recognition.
• Face recognition. • The same accuracy as human experts.
• Image generation.
• …
Reference
[1] Esteva et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature, 2017.

17 18

Applications of CNNs: Medical Diagnosis Applications of CNNs: Self Driving Cars

• CNNs play an import role in self driving cars.

• Understand images taken by the cameras.
• Recognize signs, cars, pedestrians, and obstacles.
• CNN is not everything; self-driving car is a sophisticated system.

19 20

5
 9/5/23

Recurrent Neural Networks (RNNs) Applications of RNNs: Machine Translation

• RNNs naturally fit sequence data, e.g.,

• time series data (e.g., stock price, weather),
• text data,
• speech data…

21 22

Applications of RNNs: Machine Translation Applications of RNNs: Speech Recognition

23 24

6
 9/5/23

Deep Reinforcement Learning (DFL) Applications of DRL: Games

• DFL has applications in robotics, video game, and finance.

25 26

Applications of DRL: Robotics Applications of DRL: Robotics

Control Theory v.s. DRL Control Theory v.s. DRL

Boston Dynamics' Atlas

27 28

7
 9/5/23

What we have learned so far…

• ML tasks: regression, classification, …
• ML models: linear models, CNNs, RNNs, …

Example: least squares regression model Computations

! " &
• Loss function: 𝐿 𝐰 = ∑
" #$!
𝐱#% 𝐰 − 𝑦# .
• Optimization: 𝐰⋆ = min 𝐿 𝐰 .
𝐰

• How to solve the model?

29 30

Computational Methods Gradient Descent

Example: least squares regression model
• ML tasks: regression, classification, …
! &
• ML models: linear models, CNNs, RNNs, … • Loss function: 𝐿 𝐰 = ∑"#$! 𝐱#% 𝐰 − 𝑦# .
"
• Optimization: 𝐰⋆ = min 𝐿 𝐰 .
𝐰
Example: least squares regression model
! &
• Loss function: 𝐿 𝐰 = " ∑"#$! 𝐱#% 𝐰 − 𝑦# . +,
Gradient:
• Optimization: 𝐰⋆ = min 𝐿 𝐰 . +𝐰
𝐰 • 𝐰 is a 𝑑-dimensional vector.
• 𝐿 𝐰 is a scalar.
• Computations: solve the model using numerical algorithms, e.g., )*
gradient descent (GD) or stochastic descent (SGD). • Thus )𝐰
is a 𝑑-dimensional vector.

31 32

8
 9/5/23

Gradient Descent Gradient Descent

Example: least squares regression model Example: least squares regression model
! " & ! &
• Loss function: 𝐿 𝐰 = ∑ 𝐱#% 𝐰 − 𝑦# . • Loss function: 𝐿 𝐰 = ∑"#$! 𝐱#% 𝐰 − 𝑦# .
" #$! "
• Optimization: 𝐰⋆ = min 𝐿 𝐰 . • Optimization: 𝐰⋆ = min 𝐿 𝐰 .
𝐰 𝐰

+, Gradient descent algorithm Gradient descent algorithm

Gradient:
+𝐰
• 𝐰 is a 𝑑-dimensional vector. • Randomly initialize 𝐰+. • Randomly initialize 𝐰+.
• 𝐿 𝐰 is a scalar. • For 𝑡 = 0 to 𝑇: • For 𝑡 = 0 to 𝑇:
"# "#
)* • Gradient at 𝐰! : 𝐠 ! = │𝐰 ! ; • Gradient at 𝐰! : 𝐠 ! = │𝐰 ! ;
"𝐰 "𝐰
• Thus )𝐰
is a 𝑑-dimensional vector. • 𝐰!%& = 𝐰! − 𝛼 𝐠 ! . • 𝐰!%& = 𝐰! − 𝛼 𝐠 ! .

33 34

Gradient Descent Computational Challenges

Example: least squares regression model
!
• Loss function: 𝐿 𝐰 = ∑"#$! 𝐱#% 𝐰 − 𝑦# .
& • Big data: too many training samples.
" • ImageNet: 14 million 256×256 images.
• Optimization: 𝐰⋆ = min 𝐿 𝐰 .
𝐰 • Big model: too many model parameters.
• ResNet-50 (a very popular CNN architecture) has 25 million parameters.
Variants of Gradient Descent

• Stochastic gradient descent (SGD).

• SGD with momentum.
• RMSProp.
• ADAM…

35 36

9
 9/5/23

Computational Challenges Computational Challenges

• Big data: too many training samples. • Big data: too many training samples.
• ImageNet: 14 million 256×256 images. • ImageNet: 14 million 256×256 images.
• Big model: too many model parameters. • Big model: too many model parameters.
• ResNet-50 (a very popular CNN architecture) has 25 million parameters. • ResNet-50 (a very popular CNN architecture) has 25 million parameters.

• Big data + big model bring computational challenges. • Big data + big model bring computational challenges.
• Training ResNet-50 on ImageNet using a single GPU takes around 14 • Training ResNet-50 on ImageNet using a single GPU takes around 14
days. days.

è Efficient algorithms and software systems are necessary.

37 38

Machine Learning in Practice Machine Learning in Practice

Modeling Computation Modeling Computation

• The model that fits the data and problem.

• Decide the network structures, activation
functions, loss functions, etc.
• Improve the prediction accuracy.
• Experience in ML models.
• Understanding of the problem and data.

39 40

10
 9/5/23

Machine Learning in Practice Machine Learning in Practice

Modeling Computation Modeling Computation

• Design or apply efficient algorithms. • Design or apply efficient algorithms.

• Implement the algorithm using systems like • Implement the algorithm using systems like
TensorFlow, Hadoop, etc. TensorFlow, Hadoop, etc.
• Optimize your code. • Optimize your code.
• Experience in algorithms. • Experience in algorithms.
• Experience in systems. • Experience in systems.

41 42

11