CSE 445: Machine Learning

Introduction
 Resources
 Slides provided in course should be enough – but there is a plethora of
fantastic resources available, so use them!
 Recommended Books:
 Hands-On Machine Learning with Scikit-Learn, Keras, and Tensorflow by Aurelien
Geron (will be followed extensively in the course with code examples from
https://github.com/ageron/handson-ml )
 Pattern Recognition and Machine Learning by Christopher Bishop (excellent
resource for mathematical foundations)
 Elements of Statistical Learning by Jerome Friedman et al (good reference)
 Additional Material:
 Andrew Ng’s course on Machine Learning available on Coursera
 CS 189, Berkeley
 CS 229, Stanford
 Helpful Prerequisites
 MAT361- Probability & Statistics
 Probability distribution, Random Variable, Conditional Probability, Variance (some
of the important concepts to recall to name a few)
 MAT125 – Linear Algebra
 Matrix Multiplication, Eigenvalues, Eigenvectors

 Basic programming background in Python (an OK understanding of python

syntax is all that’s necessary – Geron’s textbook has excellent code examples)

 None of them are compulsory – easier to grasp the material if completed

 Assessment (tentative)
 In-class pop quizzes on Socrative (15%)
 Midterm (20%)
 Final (30%)
 Project (30%)
 Class Participation (5%)
 Course Project
 Groups of up to 4 members (4 is a hard maximum)
 Video Demo submission at the end of the semester, and in-person/online
presentation at the end of the semester
 4-6 page Report due at semester end, IEEE format – must include link to Github
repo

 Potential Topics (few examples):

 Covid-19
 Computer Vision
 Natural Language Processing
 Reinforcement Learning
 Speech & Music Recognition
 Biomedical Imaging and Biosignals
 What is Artificial Intelligence?

The science of making machines that:

Think like people Think rationally

Act like people Act rationally

“Machines that act rationally” – a fairly broad definition!

 What is Machine Learning?
Tom Mitchell (1998): a computer program is
said to learn from experience E with respect
to some class of tasks T and performance
measure P, if its performance at tasks in T, as
measured by P, improves with experience E.

Example:
Task: Playing Checkers
Experience (data): games played by the
program (with itself)
Performance measure: winning rate Image from Tom Mitchell’s homepage
 Definition of Machine Learning
Arthur Samuel (1959): Machine Learning is the
field of study that gives the computer the ability
to learn without being explicitly programmed.

Photos from Wikipedia

 Traditional Programming

• Traditional Programming: writing a set of RULES to find

ANSWERS from DATA
 The ML Approach
Machine Learning: Use DATA and ANSWERS to learn the underlying set of RULES

Great for:

• Problems that require a lot of fine-

tuning or long list of rules
• Changing environments – ML
systems can ADAPT

• Getting insights from large amounts

of data

• Complex problems that yields no good

solution with traditional approach
 Deep Learning
 Subset of ML - loosely mimics
structure/function of human brain
 Unlike traditional ML, does not require
manual feature extraction
 Keeps getting better with more data
(typically)
 Computer Vision (CNN, GAN)
 Natural Language Processing (RNN,
LSTM)
 Automatic Speech Recognition (RNN)
 Summary – AI vs ML vs DL
 Subsets of each other
 1950 – 1990: AI in the form of Expert systems (airplane
autopilot) and Games (checkers, chess)
 1990- : Statistical Approaches with ML, busts AI winter
 2010 - : Deep Learning revolutionizes CV, NLP among
other applications
 Narrow AI
 Systems can do a few defined things (such as playing
chess, or driving a car) as well, or better than humans
 Can’t do EVERYTHING a human being can do – yet
 AI is not “taking over the world” anytime soon
 Tell your uncles to relax and stop using Whatsapp
 What kind of ML system is it?
 Useful to classify ML systems based on the following criteria:
1. Does it require human supervision? 3. Does the system build a predictive model?
 Model-based Learning
 Supervised Learning
 Instance-based Learning
 Semisupervised Learning
 Unsupervised Learning
 Reinforcement Learning
• These are not exclusive – can be
combined
2. Can it learn incrementally on the fly?
 Online Learning
• e.g. Spam filter may learn on the fly
 Batch Learning with a deep neural network – online,
model-based, supervised learning
system
 Supervised Learning
 Training data fed to algorithm
includes the desired
answers/solutions (labels)
 Example algorithms:
 Linear Regression
 Logistic Regression
 SVM
 Decision Tree
 Neural Network
 Unsupervised Learning
 Training data is unlabeled
 System learns without direct human
supervision
 Widely used in:
 Clustering
 Anomaly detection
 Association mining
 Data preprocessing
 Example algorithms:
 K-means
 PCA
 SVD
 ICA
 Semisupervised Learning
 Partially labeled data
 Unsupervised learning used
to cluster similar data
together
 Human input taken to label
the clusters
 e.g. Google Photos will
cluster similar faces, and ask
the user if they are the
same person
 Reinforcement Learning
 The learning system (agent) can:
 Observe the environment
 Select and perform an action based on
environment
 Get rewards/penalties as a result
 Based on the reward agent will changes
its state
 Agent aim: Maximize reward
 Learns what the best policy should be
 Policy defines what actions should be
chosen in a certain situation
 Very effective in controlled environments
(such as a game of chess)
 With the progress in deep learning,
increasingly used in more complex tasks
(such as driving the mars rover)
 Batch Learning vs Online Learning
 Batch Learning
 Not capable of learning after
deployment
 Must be retrained from scratch –
computationally expensive!
 Online Learning
 Can continue to learn after
deployment
 Can take advantage of parallel
computing – no down time
 Preferred choice in production
 Instance-Based vs Model-Based Learning
 Two approaches to generalization
 Instance-based Learning
 Memorize known data
 Use similarity measure to generalize
new instances
 e.g. new instance is a triangle because
it’s similar to the other triangles
 Model-based Learning
 Build model from training data
 Predict based on model output
Example ML Task: Does money make people happy?

• Life Satisfaction data from OECD

• GDP per capita data from IMF

What relationship
can we infer between
life satisfaction and
GDP per capita from
the graph?
 Model Selection
 Based on data, we can select a linear model of life satisfaction with just one
feature/attribute: GDP per capita
𝑙𝑖𝑓𝑒_𝑠𝑎𝑡𝑖𝑠𝑓𝑎𝑐𝑡𝑖𝑜𝑛 = 𝜃0 + 𝜃1 ∗ 𝐺𝐷𝑃_𝑝𝑒𝑟_𝑐𝑎𝑝𝑖𝑡𝑎
 The model has two parameters: the y-intercept 𝜃0 and the slope 𝜃1
 How to figure out the parameter values?
 Performance Measure
 Define a utility function (how good is the fitted line?), or a cost function (how bad is
the fitted line?)
 Linear Regression
 Training: try to minimize distance between linear model’s prediction on the line, vs
the actual training examples, until the estimated parameter values converge
 In this example, Linear Regression gives 𝜽𝟎 = 𝟒. 𝟖𝟓 and 𝜽𝟏 = 𝟒. 𝟗𝟏 ∗ 𝟏𝟎−𝟓
 Problems with Machine Learning
 3 V’s of Big Data
 Volume, Variety, Velocity
 Problem #1: Training data!
 Insufficient quantity
 Nonrepresentative data
 Poor-quality data
 Problem #2: How “fit” is it?
 Overfitting data
 Underfitting data
 Problem #3: Which features should be used?
 Deep Learning automates feature selection
 Overfitting
 Most common problem in ML – do not overgeneralize!
 The polynomial model is better than the linear model on training
 How about testing?
 How to avoid overfitting
 Tip #1: REGULARIZATION – USE IT
 Constrain model to keep it simple – reduce risk of overfitting
 If you can stand on one leg, you’ll be able to stay balanced with two legs
 Hyperparameters – control level of regularization
 Tip #2: Get more training data, and reduce noise in it
 Model Evaluation
 How good is your model?
 Test it on new data – data not seen by the model ever before!
 Keep 80% for training, set 20% for testing
 NEVER go below 10% test data – better model is better than better
“accuracy”
 How to regularize?
 Keep a portion of training data held out for validation
 Alternatively, use cross-validation (many validation sets instead of one)
 Pick the hyperparameters that work best on validation for your model
on the test dataset
 Ratios
 A great model
 trained with 60% training data, 20% validation data, and 20% testing data
 An okay model
 trained with 70% training data, 15% validation data, and 15% testing data
 A barely-acceptable model
 trained with 80% training data, 10% validation data, and 10% testing data
 Models with worse ratios – hacks
 Unless there’s millions of instances in the dataset
 “No Free Lunch” theorem
 Only way to know for sure which model works best is to evaluate them
 Make reasonable assumptions about your data to select model