This course presents an in-depth study of statistical machine learning approaches. It aims to provide the student with a solid understanding of methods for learning and inference in structured probabilistic models, with a healthy balance of theory and practice. It will cover topics on the semantics of direct and undirected representations in probabilistic graphical models, exact and approximate inference, and learning of model parameters and structure.

In this course, we will study a class of statistical inference models known as Probabilistic Graphical Models (PGMs). PGMs are a great example of how Computer Science and Statistics can work together. PGMs use graph data structures to represent domains with large amounts of variables and specialised algorithms for efficient inference over these graphical models. Therefore, PGMs have pushed the limits of probability theory to the scale and rate necessary to provide automated reasoning in modern AI systems.

During this course, we will cover several graphical models, including Bayesian networks, Markov networks, Conditional Random Fields, Markov chains, Hidden Markov Models, Kalman Filters and Markov decision processes. We will have a clear understanding of how these models work as well as their main algorithms for inference and learning. We will also cover several algorithms used to learn parameters and make inferences such as Monte Carlo Markov Chains (MCMC), Gibbs Sampling, Viterbi and the Baum-Welch algorithms, among others.

Machine learning is at the intersection of Artificial Intelligence, Computer Science and Statistics. While the main goal of this course is to go beyond the basics of machine learning as provided by COMP9417 (focused on probabilistic modelling and inference), we will adopt a similar teaching rationale, where theory, algorithms and empirical analysis are all important components of the course. Therefore, the lectures, tutorials and assessments are designed to address these components jointly.

This course relates more directly to the following courses:

• COMP9417 - Machine Learning and Data Mining 
• COMP3411/COMP9414 - Artificial Intelligence and 
• COMP4418 - Knowledge Representation and Reasoning

Similar to COMP9417, this course focuses on Machine Learning techniques. However, COMP9418 focuses more on models that use probabilities as the primary language for knowledge representation and reasoning. Machine Learning targets data-driven models, while the probabilistic models studied in COMP9418 are inherently interpretable and can be designed from experience, data, or a mixture of both.

COMP9418 shares similarities with other courses covering knowledge representation and reasoning, such as CPMP3411/COMP9414 and COMP4418. However, we focus on probabilistic models, while these courses have a broader view covering other representations, such as propositional and first-order logic and fuzzy sets. COMP9418 covers several probabilistic models, such as Bayesian and Markov Networks, Hidden Markov Models, Kalman Filters, and Conditional Random Fields, and demonstrates they are closely related to a common framework of probabilistic graphical models.

Machine learning is at the intersection of Artificial Intelligence, Computer Science and Statistics. While the main goal of this course is to go beyond the basics of machine learning as provided by COMP9417 (focused on probabilistic modelling and inference), we will adopt a similar teaching rationale, where theory, algorithms and empirical analysis are all important components of the course. Therefore, the lectures, tutorials and assessments are designed to address these components jointly.

The course involves lectures and practical work.

• Lectures: Aim to summarise the concepts and present case studies.
• Tutorials: Aim to reinforce the topics covered in lectures and will cover theoretical and practical exercises. The practical part of the tutorials will be based on a bring-your-own-device approach, where students will be introduced to the technology required for the assignments and follow a series of programming and data analysis questions. There will be no formal assessment of the tutorials.
• Assignments: Aim the same as the tutorials at a higher degree of difficulty and will be assessed.
• Final exam: Aim to assess the understanding of the course content and application in different use cases.

Engagement Tools and Blended Learning

• All lectures (slides/recordings) will be on the Web.
• All tutorial and lab materials (questions before, solutions after) will be on the Web.
• All assignments will have specifications on the Web and online submission.
• The final exam will likely be online.
• Forum for answering questions using WebCMS3.

In this course, we will study a class of statistical inference models known as Probabilistic Graphical Models (PGMs). PGMs are a great example of how Computer Science and Statistics can work together. PGMs use graph data structures to represent domains with large amounts of variables and specialised algorithms for efficient inference over these graphical models. Therefore, PGMs have pushed the limits of probability theory to the scale and rate necessary to provide automated reasoning in modern AI systems.

During this course, we will cover several graphical models, including Bayesian networks, Markov networks, Conditional Random Fields, Markov chains, Hidden Markov Models, Kalman Filters and Markov decision processes. We will have a clear understanding of how these models work as well as their main algorithms for inference and learning. We will also cover several algorithms used to learn parameters and make inferences such as Monte Carlo Markov Chains (MCMC), Gibbs Sampling, Viterbi and the Baum-Welch algorithms, among others.

The following formula describes how the mark will be computed and how the hurdle will be enforced:

• quizzes = mark for quizzes (out of 10)
• ass1 = mark for assignment 1 (out of 15)
• ass2 = mark for assignment 2 (out of 15)
• exam = mark for exam (out of 60)
• okExam = exam >= 24/60
• mark = quizzes + ass1 + ass2 + exam
• grade = HD | DN | CR | PS if mark >= 50 && okExam
• grade = FL if mark < 50
• grade = UF if mark >= 50 && !okExam

Please note this is a tentative schedule. All dates are only indicative and subject to change.

We will post new content to WebCMS every Monday morning (before the lecture starts). We will email all students on Monday mornings, informing them of the new content and due dates for the week.

For every week in which an assessment item is due (quiz or assignment), the due date is Sunday at 6 p.m.

Prescribed Book

• [Book] Modeling and Reasoning with Bayesian Networks. Adnan Darwiche. Cambridge. 2009 [ eBook ] [ Print ].

Recommended Books

• [Book] Probabilistic Graphical Models: Principles and Techniques. Daphne Koller and Nir Friedman. MIT Press. 2009 [ Print ].
• [Book] Probabilistic Graphical Models: Principles and Applications. Luis Enrique Sucar. Springer. 2015.
• [Book] Bayesian Reasoning and Machine Learning. David Barber. Cambridge University Press. 2012.
• [Book] Machine Learning: A Probabilistic Perspective. Kevin P. Murphy. MIT Press. 2012.
• [Book] Pattern recognition and machine learning. Christopher M. Bishop. Springer, 2006.

This course is evaluated using the myExperience system.

In the previous offering of this course, students suggested some changes in the content sequence and the addition of new material covering continuous distributions. In conversation with the students, we also noted that the tutorial code needed to be faster to support their assessment implementations.

Based on their comments, we have placed the MAP lecture earlier in the course and reduced its content to allow space for a new lecture covering Gaussian Bayesian networks. We reimplemented the tutorial code, replacing an unordered dictionary with a NumPy array to increase code efficiency. We also improved code organisation using an object-oriented implementation.

We thank all the students who provided feedback on this course through MyExperience, email and conversations. These students include Martin Eftimoski, Gareth Dando, Lucky Zhan, Oliver Li and Darren Chong.