MSc: Machine Learning

Machine Learning

• Course name: Machine Learning
• Course number: R-01

Prerequisites

• CSE117 — Data Structures and Algorithms: python, numpy, basic object-oriented concepts, memory management
• CSE201 — Mathematical Analysis I
• CSE203 — Mathematical Analysis II
• CSE202 — Analytical Geometry and Linear Algebra I
• CSE204 — Analytic Geometry And Linear Algebra II

Course characteristics

Key concepts of the class

• Machine learning paradigms
• Machine Learning approaches, and algorithms

What is the purpose of this course?

There is a growing business need of individuals skilled in artificial intelligence, data analytics, and machine learning. Therefore, the purpose of this course is to provide students with an intensive treatment of a cross-section of the key elements of machine learning, with an emphasis on implementing them in modern programming environments, and using them to solve real-world data science problems.

Course Objectives Based on Bloom’s Taxonomy

What should a student remember at the end of the course?

By the end of the course, the students should be able to recognize and define

• Different learning paradigms
• A wide variety of learning approaches and algorithms
• Various learning settings
• Performance metrics
• Popular machine learning software tools

What should a student be able to understand at the end of the course?

By the end of the course, the students should be able to describe and explain (with examples)

• Difference between different learning paradigms
• Difference between classification and regression
• Concept of learning theory (bias/variance tradeoffs and large margins etc.)
• Kernel methods
• Regularization
• Ensemble Learning
• Neural or Deep Learning

What should a student be able to apply at the end of the course?

By the end of the course, the students should be able to apply

• Classification approaches to solve supervised learning problems
• Clustering approaches to solve unsupervised learning problems
• Ensemble learning to improve a model’s performance
• Regularization to improve a model’s generalization
• Deep learning algorithms to solve real-world problems

Course evaluation

Course grade breakdown

		Proposed points
Labs/seminar classes	20	0
Interim performance assessment	30	40
Exams	50	60

If necessary, please indicate freely your course’s features in terms of students’ performance assessment: None

Grades range

Course grading range

		Proposed range
A. Excellent	90-100
B. Good	75-89
C. Satisfactory	60-74
D. Poor	0-59

If necessary, please indicate freely your course’s grading features: The semester starts with the default range as proposed in the Table [tab:MLCourseGradingRange], but it may change slightly (usually reduced) depending on how the semester progresses.

Resources and reference material

• T. Hastie, R. Tibshirani, D. Witten and G. James. An Introduction to Statistical Learning. Springer 2013.
• T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. Springer 2011.
• Tom M Mitchel. Machine Learning, McGraw Hill
• Christopher M. Bishop. Pattern Recognition and Machine Learning, Springer

Course Sections

The main sections of the course and approximate hour distribution between them is as follows:

Course Sections

Section	Section Title	Teaching Hours
1	Supervised Learning	24
2	Decision Trees and Ensemble Learning	8
3	Unsupervised Learning	8
4	Deep Learning	12

Section 1

Section title:

Supervised Learning

Topics covered in this section:

• Introduction to Machine Learning
• Derivatives and Cost Function
• Data Pre-processing
• Linear Regression
• Multiple Linear Regression
• Gradient Descent
• Polynomial Regression
• Splines
• Bias-varaince Tradeoff
• Difference between classification and regression
• Logistic Regression
• Naive Bayes
• Bayesian Network
• KNN
• Confusion Metrics
• Performance Metrics
• Regularization
• Hyperplane Based Classification
• Perceptron Learning Algorithm
• Max-Margin Classification
• Support Vector Machines
• Slack Variables
• Lagrangian Support Vector Machines
• Kernel Trick

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. Is it true that in simple linear regression ${\textstyle R^{2}}$ and the squared correlation between X and Y are identical?
2. What are the two assumptions that the Linear regression model makes about the Error Terms?
3. Fit a regression model to a given data problem, and support your choice of the model.
4. In a list of given tasks, choose which are regression and which are classification tasks.
5. In a given graphical model of binary random variables, how many parameters are needed to define the Conditional Probability Distributions for this Bayes Net?
6. Write the mathematical form of the minimization objective of Rosenblatt’s perceptron learning algorithm for a two-dimensional case.
7. What is perceptron learning algorithm?
8. Write the mathematical form of its minimization objective for a two-dimensional case.
9. What is a max-margin classifier?
10. Explain the role of slack variable in SVM.

Typical questions for seminar classes (labs) within this section

1. How to implement various regression models to solve different regression problems?
2. Describe the difference between different types of regression models, their pros and cons, etc.
3. Implement various classification models to solve different classification problems.
4. Describe the difference between Logistic regression and naive bayes.
5. Implement perceptron learning algorithm, SVMs, and its variants to solve different classification problems.
6. Solve a given optimization problem using the Lagrange multiplier method.

Test questions for final assessment in this section

1. What does it mean for the standard least squares coefficient estimates of linear regression to be scale equivariant?
2. Given a fitted regression model to a dataset, interpret its coefficients.
3. Explain which regression model would be a better fit to model the relationship between response and predictor in a given data.
4. If the number of training examples goes to infinity, how will it affect the bias and variance of a classification model?
5. Given a two dimensional classification problem, determine if by using Logistic regression and regularization, a linear boundary can be estimated or not.
6. Explain which classification model would be a better fit to for a given classification problem.
7. Consider the Leave-one-out-CV error of standard two-class SVM. Argue that under a given value of slack variable, a given mathematical statement is either correct or incorrect.
8. How does the choice of slack variable affect the bias-variance tradeoff in SVM?
9. Explain which Kernel would be a better fit to be used in SVM for a given data.

Section 2

Section title:

Decision Trees and Ensemble Methods

Topics covered in this section:

• Decision Trees
• Bagging
• Boosting
• Random Forest
• Adaboost

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. What are pros and cons of decision trees over other classification models?
2. Explain how tree-pruning works.
3. What is the purpose of ensemble learning?
4. What is a bootstrap, and what is its role in Ensemble learning?
5. Explain the role of slack variable in SVM.

Typical questions for seminar classes (labs) within this section

1. Implement different variants of decision trees to solve different classification problems.
2. Solve a given classification problem problem using an ensemble classifier.
3. Implement Adaboost for a given problem.

Test questions for final assessment in this section

1. When a decision tree is grown to full depth, how does it affect tree’s bias and variance, and its response to noisy data?
2. Argue if an ensemble model would be a better choice for a given classification problem or not.
3. Given a particular iteration of boosting and other important information, calculate the weights of the Adaboost classifier.

Section 3

Section title:

Unsupervised Learning

Topics covered in this section:

• K-means Clustering
• K-means++
• Hierarchical Clustering
• DBSCAN
• Mean-shift

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. Which implicit or explicit objective function does K-means implement?
2. Explain the difference between k-means and k-means++.
3. Whaat is single-linkage and what are its pros and cons?
4. Explain how DBSCAN works.

Typical questions for seminar classes (labs) within this section

1. Implement different clustering algorithms to solve to solve different clustering problems.
2. Implement Mean-shift for video tracking

Test questions for final assessment in this section

1. K-Means does not explicitly use a fitness function. What are the characteristics of the solutions that K-Means finds? Which fitness function does it implicitly minimize?
2. Suppose we clustered a set of N data points using two different specified clustering algorithms. In both cases we obtained 5 clusters and in both cases the centers of the clusters are exactly the same. Can 3 points that are assigned to different clusters in one method be assigned to the same cluster in the other method?
3. What are the characterics of noise points in DBSCAN?

Section 4

Section title:

Deep Learning

Topics covered in this section:

• Artificial Neural Networks
• Back-propagation
• Convolutional Neural Networks
• Autoencoder
• Variatonal Autoencoder
• Generative Adversairal Networks

What forms of evaluation were used to test students’ performance in this section?

|a|c| & Yes/No
Development of individual parts of software product code & 1
Homework and group projects & 1
Midterm evaluation & 1
Testing (written or computer based) & 1
Reports & 0
Essays & 0
Oral polls & 0
Discussions & 1

Typical questions for ongoing performance evaluation within this section

1. What is a fully connected feed-forward ANN?
2. Explain different hyperparameters of CNNs.
3. Calculate KL-divergence between two probability distributions.
4. What is a generative model and how is it different from a discriminative model?

Typical questions for seminar classes (labs) within this section

1. Implement different types of ANNs to solve to solve different classification problems.
2. Calculate KL-divergence between two probability distributions.
3. Implement different generative models for different problems.

Test questions for final assessment in this section

1. Explain what is ReLU, what are its different variants, and what are their pros and cons?
2. Calculate the number of parameters to be learned during training in a CNN, given all important information.
3. Explain how a VAE can be used as a generative model.

Exams and retake planning

Exam

Exams will be paper-based and will be conducted in a form of problem solving, where the problems will be similar to those mentioned above and will based on the contents taught in lecture slides, lecture discussions (including white-board materials), lab materials, reading materials (including the text books), etc. Students will be given 1-3 hours to complete the exam.

Retake 1

First retake will be conducted in the same form as the final exam. The weight of the retake exam will be 5% larger than the passing threshold of the course.

Retake 2

Second retake will be conducted in the same form as the final exam. The weight of the retake exam will be 5% larger than the passing threshold of the course.