Dr.

Korris Chung
Dept of Computing

COMP4432
Roadmap
o Basic Concepts
o Data vs Feature vs Model
o Theoretical/Mathematical aspect
o Machine Learning - Why? What? and Where?
o Related Disciplines
o ML Models
o Supervised Learning
o Unsupervised Learning
o Issues & Resources
o Take-home messages!
2
 Basic Concepts

(to be learned) in this course

3
Basic Concepts

4
Really Important Views/Concepts
o Data vs Feature

Could be a major data analytics work --- Feature engineering (FE)!

If your model doesn’t work, it may not be due to the model itself (e.g.
simple model, bad training). It could be due to the ineffective FE!
5
Really Important Views/Concepts
ML concerns with the model
itself, assuming appropriate
o Data vs Feature vs Model feature input and desired
output.

Deep learning model could step in to carry out feature

learning (not FE) and “machine learning”!!! 6
Really Important Views/Concepts

If you can comprehend these concepts and manipulate

them sophistically, you could be very successfully in any
ML scientist/developer job and/or Data
Scientist/Analytics job!

Probably, you can drop this course!

7
Let’s go a bit deeper!

8
Feature space and decision boundary

9
Supervised Learning (Classification) Space

10
 Unsupervised Learning (Clustering) Space

The data itself does not have

What label information!
dimension is it?
What is its
physical
meaning?
11
Example: Face clusters

12
Example:
Face Classification

13
After all, how are data samples
represented?
Iris dataset: sepal length and width, petal length

14
Face geometry?

15
Social Network Geometry?

How are social network data represented? 16

How should stock data be represented?

It requires careful, domain

knowledge driven feature
engineering work, e.g.
technical analysis.

17
After all, can data representation be learned?

Yes, through Deep Learning

(Representation Learning)
18
Yet more advanced applications

So, what are the data, feature and output in this application?
19
Yet more advanced applications…
It’s just so powerful…
Image Completion
 ML Concept -
Theoretical/Mathematical
Aspect

(to be learned) in this course

22
Stochastic Gradient Descent (SGD)

23
Regularization

Dataset is the same but

fitting is “improved”, i.e.,
generalize better to unseen
data!

24
Bias-Variance Trade-off

25
 Why?
What?
Where?

26
Why “Learn”?
§ Machine learning is programming computers to
optimize a performance criterion using example data or
past experience.
§ There is no need to “learn” to calculate payroll (which
is deterministic)
§ Learning is used when:
§ Human expertise does not exist (navigating on Mars),
§ Humans are unable to explain their expertise (speech
recognition, painting style)
§ Solution changes in time (routing on a computer network,
stock market)
§ Solution needs to be adapted to particular cases (user
biometrics, Chinese chatbot)

27
What We Talk About When We
Talk About “Learning”
§ Learning general models from a data of particular
examples
§ Data is cheap and abundant (data warehouses, data
marts); knowledge is expensive and scarce.
§ Build a model that is a good and useful
approximation to the data.

28
What is Machine Learning?
Definition Stuff:

29
What is Machine Learning?
§ Machine Learning
§ Study of algorithms that improve their performance at some task with
experience
§ Optimize a performance criterion using example data or past
experience.
§ Role of Statistics: Inference from a sample
§ Role of Computer science: Efficient algorithms to
§ Solve the optimization problem, e.g. via SGD
§ Representing and evaluating the model for inference

30
Why Study Machine Learning?
- Developing Better Computing Systems
§ Develop systems that are too difficult/expensive to
construct manually because they require specific detailed
skills or knowledge tuned to a specific task (feature
engineering bottleneck).
§ Develop systems that can automatically adapt and
customize themselves to individual users (self-learning).
§ Personalized news or email filter
§ Personalized tutoring
§ Discover new knowledge from large databases (data
mining).
§ Market basket analysis (e.g. diapers and beer)
§ Medical text mining

31
Where is Machine Learning?

32
Related Disciplines
§ Artificial Intelligence
§ Data Mining
§ Probability and Statistics
§ Information theory
§ Numerical optimization
§ Computational complexity theory
§ Control theory (adaptive)
§ Psychology (developmental, cognitive)
§ Neurobiology
§ Linguistics
§ Philosophy

33
ML Hierarchy

34
Growth of Machine Learning
§ Machine learning is preferred approach to
§ Speech recognition, Natural language processing
§ Computer vision
§ Medical outcome analysis
§ Robot control
§ Computational biology
§ This trend is accelerating
§ Improved machine learning algorithms
§ Improved data capture, networking, faster computers
§ Software too complex to write by hand
§ New sensors / IO devices
§ Internet of Things (IoT) and 5G Network
§ Demand for self-customization to user, environment
§ It turns out to be difficult to extract knowledge from human
expertsàfailure of expert systems in the 1980’s.
35
Models
§ Association Analysis (in Data Mining)
§ Supervised Learning
§ Classification
§ Regression/Prediction
§ Unsupervised Learning
§ Reinforcement Learning (in AI)
§ Also, semi-supervised learning (in ML)

36
Learning Associations (DM stuff)
Famous quote: 60% of the customers
buy diaper also buy beer
§ Market Basket Analysis (MBA):
P (Y | X ) probability that somebody who buys X also buys Y where
X and Y are products/services.
Example: P(Diaper|Beer) = 100%; P(Beer|Diaper) = 75%

Market-Basket transactions
TID Items
1 Bread, Milk
2 Bread, Diaper, Beer, Eggs
3 Milk, Diaper, Beer, Coke
4 Bread, Milk, Diaper, Beer
5 Bread, Milk, Diaper, Coke
37
Classification
§ Example: Credit scoring
§ Differentiating
between low-risk and
high-risk customers
from their income and
savings

ML/DM Model Features are

Discriminant: “income and
savings”!
IF income > θ1 AND savings > θ2 Why not
THEN low-risk ELSE high-risk “age”?
38
Classification: Applications
• Aka Pattern Recognition (PR)
• Face recognition: Pose, lighting, occlusion (glasses,
beard), make-up, hair style
• Character recognition: Different handwriting styles.
• Speech recognition: Temporal dependency.
• Use of a dictionary or the syntax of the language.
• Sensor fusion: Combine multiple modalities; e.g. visual (lip
image) and acoustic for speech
• Medical diagnosis: From symptoms to illnesses
• Web Advertizing: Predict if a user clicks on an ad on the
Internet.

39
Face Recognition
Training examples of a person

Test images

AT&T Laboratories, Cambridge UK

http://www.uk.research.att.com/facedatabase.html

40
Prediction: Regression

§ Example: Price of a used

car
§ x : car attributes y = wx+w0
y : price
y = g (x | 𝜃 )
g ( ) model,
𝜃 learnable parameters
(w, w0)

41
Supervised Learning: Classification
§ Example: Cancer diagnosis
Patient ID # of Tumors Avg Area Avg Density Diagnosis
1 5 20 118 Malignant
2 3 15 130 Benign Training Set
3 7 10 52 Benign
4 2 30 100 Malignant

§ Use this training set to learn how to classify patients where

diagnosis is not known:
Patient ID # of Tumors Avg Area Avg Density Diagnosis
101 4 16 95 ?
102 9 22 125 ? Test Set
103 1 14 80 ?

Input Data Classification

§ The input data is often easily obtained, whereas the
classification or label is not.
Notationally,
Supervised Learning: Classification

Goal: Learn a model from labeled data; Use training set +

some learning method to produce a predictive model; Use
this predictive model to classify new data.
43
Pictorially,
Supervised Learning: Classification
What’s happening in the feature space?

• Methods: Support Vector Machines (SVM), neural networks (NN),

decision trees, K-nearest neighbors, naive Bayes, etc.
44
More sophisticated classification

45
Unsupervised Learning
§ Learning “what normally happens”
§ No output or label
§ Clustering: Grouping similar instances
§ Other applications: Summarization, Association
Analysis
§ Example applications
§ Customer segmentation in CRM
§ Image compression: Color quantization
§ Bioinformatics: Learning motifs

46
Notationally,
Unsupervised Learning: Clustering

47
Pictorially,
Unsupervised Learning: Clustering

What’s happening in the feature space?

48
Pictorially,
Unsupervised Learning: Clustering
What’s happening in the feature space?

• Methods: K-means, gaussian mixtures, hierarchical

clustering, spectral clustering, etc. 49
Reinforcement Learning
§ Topics:
§ Policies: what actions should an agent take in a particular
situation
§ Utility estimation: how good is a state (àused by policy)
§ No supervised output but delayed reward
§ Credit assignment problem (what was responsible for
the outcome?)
§ Applications:
§ Game playing
§ Robot in a maze
§ AWS DeepRacer
§ Multiple agents, partial observability, ...

50
Issues in Machine Learning
§ What algorithms can approximate functions well and
when?
§ How does the number of training examples influence
accuracy?
§ Problem representation / feature extraction
§ Integrating learning with systems
§ What are the theoretical limits of learnability?
§ Continuous (life-long) learning
§ Transfer learning
§ Few-shot learning
§ Interpretable ML (or explainable AI)
§ Many others…
51
Measuring Performance
• Generalization accuracy
• Solution correctness
• Solution quality (length, efficiency)
• Speed of performance (scalability)

52
Scaling issues in ML
• Number of
• Inputs
• Outputs (e.g. Extreme Classification (with lots of labels))
• Batch vs real-time
• Training vs testing

53
Resources: Datasets
§ UCI Repository:
http://www.ics.uci.edu/~mlearn/MLRepository.html
§ UCI KDD Archive:
http://kdd.ics.uci.edu/summary.data.application.html
§ Kaggle: https://www.kaggle.com/datasets
§ Tianchi (天池): https://tianchi.aliyun.com/dataset/
and many others…

54
Resources: Journals
§ Journal of Machine Learning Research
www.jmlr.org
§ Machine Learning
§ IEEE Transactions on Neural Networks
§ IEEE Transactions on Pattern Analysis and Machine
Intelligence
§ Annals of Statistics
§ Journal of the American Statistical Association
§ ...
55
Resources: Conferences
§ International Conference on Machine Learning (ICML)
§ European Conference on Machine Learning (ECML)
§ Neural Information Processing Systems (NIPS)
§ Computational Learning
§ International Joint Conference on Artificial Intelligence (IJCAI)
§ ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)
§ IEEE Int. Conf. on Data Mining (ICDM)
§ Etc.

56
Take-home Messages
About ML and Data Analytics (DA)
§ DA concerns more about data+feature
§ ML concerns more about model

About this subject

§ Introductory course that covers a range of fundamental machine learning
techniques and some popular deep learning models.
§ It requires lots of abstract thinking.
§ A bit more theoretical/mathematical oriented than Data Mining or Big Data
Analytics like subjects.
§ ML development platform to be self-learnt and practiced in competition.
§ It’s going to be fun and hard work.

57
Acknowledgement
§ Slides of
§ E. Alpaydin, Introduction to Machine Learning. 2nd Ed.
MIT Press, 2010.
§ C.F. Eick, U of Houston.
§ Photos from Internet

58