COMP329 Advanced ML & Deep Learning

Course syllabus · Semester 1 · 15 credits

COMP329 Syllabus

The definitive statement of what COMP329 covers, how it is taught, how it is assessed, and what is expected of you. Read it alongside the course overview page and the orientation unit before the first laboratory session.

6 modules 18 units 152 tutorial pages 15 credits 14 teaching weeks

1 Course description

COMP329 develops the machine learning foundations established in COMP328 into the architectures, training techniques and engineering practices used in contemporary deep learning systems.

The course begins with the mathematics of learning at scale: gradient-based optimisation, regularisation and principled model evaluation. It then builds upwards through neural network architectures — fully connected, convolutional, recurrent and transformer — before turning to generative modelling and representation learning.

Throughout, the emphasis is on implementation. You will build networks from first principles before using established frameworks, so that you understand what the libraries are doing on your behalf and can debug them when they fail.

The final third of the course addresses the realities of applied deep learning: data pipelines, transfer learning, compute constraints, interpretability, fairness and safe deployment. You finish with a capstone that takes a deep learning system from problem definition to a working, documented, evaluated artefact.

How to read this document

Sections 1–6 describe what the course is. Sections 7–11 describe how it runs and how you are assessed. Sections 12–15 cover resources, policies and administrative matters. Anything marked indicative is confirmed in the course Moodle site at the start of the semester.

2 Course details

Course at a glance

Course code and title
COMP329 — Advanced Machine Learning and Deep Learning
Level
Undergraduate · Final year / advanced
Credits
15 credits, one semester
Prerequisites
COMP328 Fundamentals of Machine Learning, or equivalent. COMP326 Data Science Foundations provides the assumed data-handling background. COMP327 is recommended but not required.
Assumed prior knowledge
Comfort with Python, vectors and matrices, partial derivatives and the chain rule, and basic probability. The course reviews the optimisation mathematics it relies on but does not reteach calculus or linear algebra.
Delivery
Two lectures and one GPU laboratory session per week, plus guided self-study and asynchronous tutorial pages.
Contact hours
Approximately 36 hours of scheduled teaching; 150 hours total notional learning
Expected weekly workload
12–14 hours, including model training time
Language and tools
Python 3, PyTorch, TensorFlow/Keras, scikit-learn, Hugging Face Transformers, Jupyter with CUDA acceleration
Assessment
5 assignments, 18 unit quizzes, 1 midterm, 1 final exam and a capstone project
Compute
Shared GPU workstations in the laboratory. All assignments are designed to be completable on free cloud notebook tiers — no personal GPU required.
Course coordinator
Computing & Information Systems teaching team (see Moodle for named staff and office hours)
Neural networks Transformers Generative models Responsible AI Capstone

3 Learning outcomes

On successful completion of COMP329 you will be able to design, implement and critically evaluate advanced machine learning and deep learning systems. Outcomes are numbered LO1–LO6 and are referenced in assessment briefs.

  1. LO1 — Learning theory and optimisation. Explain how gradient-based optimisation trains large models; apply regularisation and generalisation techniques deliberately; select evaluation methods appropriate to the problem and the data.
  2. LO2 — Deep architectures. Implement neural networks from first principles; design and train convolutional and recurrent architectures; diagnose and remedy training pathologies such as vanishing gradients.
  3. LO3 — Transformers and transfer learning. Explain attention and the transformer architecture; fine-tune pretrained models for new tasks; choose between training from scratch and transfer learning on the basis of evidence.
  4. LO4 — Generative and unsupervised methods. Apply representation learning to unlabelled data; compare autoencoders, GANs, VAEs and diffusion models; evaluate generative output both quantitatively and critically.
  5. LO5 — Applied deep learning. Build end-to-end pipelines for vision and language tasks; deploy and monitor models in a production setting; manage data, compute and reproducibility constraints.
  6. LO6 — Responsible advanced AI. Interpret model behaviour and communicate uncertainty; assess fairness, robustness and safety risks; document and justify design decisions professionally.

Relationship to COMP328

COMP328 covers machine learning fundamentals: regression, classification, decision trees, ensembles and the core evaluation framework. COMP329 assumes all of that and moves into deep architectures, representation learning, generative models and production deployment. Material from COMP328 is not retaught.

4 Curriculum position

COMP329 is the final course in the four-course AI and data sequence. It builds directly on the machine learning foundations established in COMP328 and completes the progression from data literacy to advanced AI engineering.

  1. COMP326 · Data Science Foundations
  2. COMP327 · Generative AI
  3. COMP328 · Machine Learning
  4. COMP329 · Advanced ML & Deep Learning
How the four courses divide the territory
Course Focus Assumed on entry to COMP329
COMP326 Data literacy, statistics, analytics, visualisation, the data lifecycle Yes — data handling, EDA, statistical reasoning
COMP327 AI literacy and responsible use of generative AI Recommended
COMP328 Machine learning fundamentals and evaluation Yes — prerequisite
COMP329 Advanced architectures, generative modelling and deployment This course

5 Course structure

The course comprises six modules and eighteen units, delivered as 152 tutorial pages: theory, worked derivations, PyTorch walkthroughs, GPU lab activities and a knowledge check per unit.

Course structure at a glance
Module Focus Units Tutorials
1 · Advanced ML Foundations Optimisation, regularisation, generalisation and rigorous evaluation 3 26
2 · Neural Networks & Deep Learning Backpropagation, training at depth and convolutional architectures 3 27
3 · Sequence Models & Attention Recurrent networks, attention, transformers and transfer learning 3 25
4 · Generative & Unsupervised Learning Representation learning, generative models and modern synthesis 3 25
5 · Applied Deep Learning Systems Vision and language applications, deployment and MLOps 3 25
6 · Responsible AI & Capstone Interpretability, fairness, safety and the capstone project 3 24
Total Six modules · eighteen units 18 152

6 Unit catalogue

Each unit follows the same rhythm: study the theory, implement it, train and evaluate it on GPU, then check your understanding with a quiz.

All eighteen units, with tutorial counts and associated assessment
Unit Title Tutorials Assessment
U01 Optimisation and Gradient Methods 9 A1
U02 Regularisation and Generalisation 9 A1
U03 Model Selection and Evaluation 8 A1
U04 Neural Network Fundamentals 9 A2
U05 Training Deep Networks 9 A2
U06 Convolutional Neural Networks 9 A2
U07 Recurrent Networks 8 A3
U08 Attention and Transformers 9 A3
U09 Pretraining and Transfer Learning 8 A3
U10 Unsupervised Representation Learning 8 A4
U11 Generative Models 9 A4
U12 Diffusion and Modern Synthesis 8 A4
U13 Computer Vision Applications 8 A5
U14 Natural Language Applications 8 A5
U15 Deployment and MLOps 9 A5
U16 Interpretability and Explainability 8 Capstone
U17 Robustness, Fairness and Safety 8 Capstone
U18 Capstone Preparation and Integration 8 Capstone
Eighteen units 152

Module pages

Full unit descriptions, tutorial sequences and learning activities are on the module landing pages: Module 1, Module 2, Module 3, Module 4, Module 5, Module 6.

7 Teaching and learning

Every unit follows the same four-part structure, so you always know what is expected and how to prepare for assessment.

  1. Study the theory

    Derivations and architecture diagrams explain why a method works, not just how to call it.

  2. Implement it

    Build the model in NumPy first, then reproduce it in PyTorch so the framework holds no mysteries.

  3. Train and evaluate

    Run GPU lab experiments, tune hyperparameters and interpret the results critically.

  4. Check your understanding

    Finish with self-check questions, a unit quiz and a short reflective summary.

Laboratory sessions

One timetabled GPU laboratory per week. Sessions are hands-on: you run code, break it, fix it and record what you observe. Attendance is strongly recommended and is recorded, but is not itself assessed.

Compute discipline

Model training consumes wall-clock time you cannot compress. Start experiments early, use small subsets while developing, and only scale up once the pipeline is proven correct. Working within a realistic compute budget is an assessed professional skill.

Reproducibility

Every notebook you submit must be runnable end to end from a clean kernel, with random seeds set and dependencies listed. Reproducibility is part of the marking criteria for every assignment.

Independent study

Scheduled contact time is roughly a third of the notional learning hours. The remainder is independent work:

  • Reading and re-working the tutorial pages before lectures.
  • Completing lab notebooks and extension exercises.
  • Training runs — often best left running while you do something else.
  • Writing up: evaluation, ablation and error analysis.
  • Capstone work from week 8 onwards.

8 Weekly schedule

The schedule below is indicative. Assessment deadlines are fixed; the mapping of units to weeks may shift slightly in response to cohort progress.

Fourteen teaching weeks, plus revision and examination periods
Week Module · Unit Focus Assessment
1 M1 · U01 Optimisation and gradient methods A1 released
2 M1 · U02 Regularisation and generalisation
3 M1 · U03 Model selection and evaluation A1 due
4 M2 · U04 Neural network fundamentals A2 released
5 M2 · U05 Training deep networks
6 M2 · U06 Convolutional neural networks A2 due
7 M3 · U07 Recurrent networks A3 released
8 M3 · U08 Attention and transformers Capstone proposal
9 M3 · U09 Pretraining and transfer learning A3 due
10 M4 · U10–U11 Representation learning; generative models Midterm (Modules 1–3)
11 M4 · U12 Diffusion and modern synthesis A4 released
12 M5 · U13–U14 Vision and language applications A4 due
13 M5 · U15 & M6 · U16 Deployment and MLOps; interpretability A5 released
14 M6 · U17–U18 Fairness and safety; capstone integration A5 due
Rev. Revision Consolidation and exam preparation Capstone submission
Exam Examination period Cumulative final examination Final exam

Deadlines fall at 23:59 on the Friday of the week shown unless stated otherwise in Moodle.

9 Assessment

Assessment is weighted towards implementation and the capstone. Five assignments build progressively towards the final project, so that by the time you begin the capstone you have already practised every component skill.

35%

Five assignments

Optimisation, network implementation, transformers, generative models and deployment.

30%

Capstone project

A complete deep learning system, from problem definition to evaluated deployment.

15%

Final exam

Cumulative, with an emphasis on architecture reasoning and critical evaluation.

10%

Midterm

Modules 1–3: optimisation, deep networks and sequence models.

10%

Unit quizzes

Eighteen short online quizzes, one per unit, with unlimited attempts.

Assessment components, outcomes and timing
Component Weight Outcomes Timing
Assignment 1 — Optimisation and Regularisation 6% LO1 Week 3
Assignment 2 — Neural Network Implementation 7% LO2 Week 6
Assignment 3 — Transformers and Transfer Learning 7% LO3 Week 9
Assignment 4 — Generative Modelling 7% LO4 Week 12
Assignment 5 — Deployment Project 8% LO5 Week 14
Unit quizzes (18 × 0.56%) 10% LO1–LO6 Ongoing
Midterm examination 10% LO1–LO3 Week 10
Capstone project 30% LO1–LO6 Weeks 8–15
Final examination 15% LO1–LO6 Examination period
Total 100%

Pass requirements

To pass the course you must achieve an overall mark of at least 40% and submit the capstone project. A capstone that is not submitted is recorded as a fail regardless of the mark on other components. Weightings and thresholds are indicative and confirmed in the course syllabus on Moodle.

Baselines are mandatory

Every assignment involving a trained model must compare against at least one simple baseline. A complex model that is not shown to outperform a trivial one is not a successful submission — this applies to the capstone in particular.

10 Assignment briefs

Full briefs, marking rubrics and starter notebooks are released on Moodle in the week indicated. Links below point to the public brief pages.

  • Assignment 1 Optimisation and Regularisation — implement and compare training strategies Brief
  • Assignment 2 Neural Network Implementation — build and train a network from first principles Brief
  • Assignment 3 Transformers and Transfer Learning — fine-tune a pretrained model Brief
  • Assignment 4 Generative Modelling — train and critically evaluate a generative model Brief
  • Assignment 5 Deployment Project — package, serve and monitor a trained model Brief

What every submission must contain

  • A runnable notebook, executed end to end from a clean kernel.
  • Fixed random seeds and a pinned dependency list.
  • At least one baseline for comparison.
  • A short written evaluation, including where the model fails.
  • An honest statement of compute used and time taken.

How submissions are marked

  • Correctness — does the implementation do what it claims?
  • Evaluation rigour — are the comparisons fair and complete?
  • Critical interpretation — do you understand your own results?
  • Reproducibility — can someone else run it?
  • Communication — is the report clear and well structured?

11 Capstone project

The capstone runs across the second half of the course. You take a real problem from definition through to a trained, evaluated and documented deep learning artefact, supported by twelve structured phases.

What you submit

  • Project proposal and problem framing
  • Dataset and compute plan
  • Model implementation and training notebook
  • Evaluation report including baselines and ablations
  • Deployed or packaged system
  • Final written report and oral presentation

Portfolio ready

A trained model, reproducible training code and an honest evaluation report together make a strong technical portfolio piece for graduate study or industry roles.

Baselines are mandatory

Every capstone must compare against at least one simple baseline. A complex model that is not shown to outperform a trivial one is not a successful project.

Dataset choice

You may choose your own dataset or select from a curated list of approved sources. Either way, you must document provenance, licensing and any preprocessing decisions as part of the submission.

  1. Problem IdentificationDefine a task where learning from data is genuinely justified.
  2. Stakeholder AnalysisIdentify who uses the system and what failure would cost them.
  3. Dataset SelectionSource, document and assess data suitability and licensing.
  4. Data PipelineBuild reproducible loading, splitting and augmentation.
  5. Baseline ImplementationEstablish a simple benchmark before anything complex.
  6. Architecture DesignChoose and justify an architecture for the task.
  7. Training and TuningTrain, monitor and tune within your compute budget.
  8. Evaluation and AblationMeasure performance rigorously and test your design choices.
  9. InterpretationExplain what the model learned and where it fails.
  10. Responsible AI ReviewAssess fairness, robustness and misuse risk.
  11. DeploymentPackage and serve the model with monitoring in place.
  12. Report and PresentationDocument and defend the full system.

12 Reading and resources

There is no single required textbook. The tutorial pages are self-contained, and the texts below are recommended for depth, alternative explanations and reference.

Core texts

Prince, S. — Understanding Deep Learning
The closest match to this course. Clear, modern and freely available online. Recommended as the primary companion text.
Goodfellow, I., Bengio, Y. & Courville, A. — Deep Learning
The canonical reference. Dense but authoritative; best used for the theory of optimisation, regularisation and generative modelling.
Zhang, A. et al. — Dive into Deep Learning
Interactive, code-first and free. Excellent for the implementation-heavy units in Modules 2 and 3.
Géron, A. — Hands-On Machine Learning
Practical end-to-end workflows in scikit-learn, Keras and TensorFlow. Useful for Module 5.

Key papers

  • Vaswani et al. (2017) — Attention Is All You Need
  • He et al. (2015) — Deep Residual Learning for Image Recognition
  • Kingma & Welling (2013) — Auto-Encoding Variational Bayes
  • Goodfellow et al. (2014) — Generative Adversarial Networks
  • Ho et al. (2020) — Denoising Diffusion Probabilistic Models
  • Devlin et al. (2018) — BERT; Radford et al. (2019) — GPT-2

Tools and technology

All tools are free and open source. Laboratory sessions run on shared GPU workstations, and every assignment is completable on free cloud notebook tiers.

Python 3 The language for the entire course.
PyTorch The primary framework for building and training networks.
TensorFlow / Keras An alternative framework, used for comparison and deployment.
scikit-learn Classical baselines, metrics and preprocessing utilities.
Hugging Face Pretrained transformers, datasets and fine-tuning workflows.
Jupyter & GPU Notebook environment with CUDA acceleration for training labs.

Setup

Follow the PyTorch setup guide before your first laboratory session. If you intend to work on your own machine, verify that your environment can run the week 1 notebook end to end.

13 Course policies

The policies below apply to COMP329. Where a policy conflicts with a university-wide regulation, the university regulation takes precedence.

Academic integrity

All submitted work must be your own. You may discuss ideas with classmates, but the code, analysis and writing you submit must be produced by you, and any external source — including code snippets, tutorials and pretrained models — must be cited.

Collusion, plagiarism and the submission of work generated by another person are academic misconduct and are handled under the university's academic integrity regulations.

Use of generative AI tools

Generative AI tools may be used as a learning aid: to explain a concept, suggest a debugging approach, or help you phrase a paragraph. They may not be used to produce submitted code or written analysis wholesale.

  • Disclose any use of generative AI in a short statement accompanying each submission.
  • You remain fully responsible for the correctness of everything you submit.
  • You must be able to explain any line of your submission in person.
  • Unacknowledged AI-generated content is treated as academic misconduct.

These expectations build directly on COMP327 Responsible and Effective Use of Generative AI.

Late submission and extensions

Standard late penalties apply to assignments unless an extension has been granted in advance. Extensions are considered for illness, caring responsibilities and other significant circumstances.

Request an extension through the university's extenuating circumstances process before the deadline where possible. Capstone milestones cannot be extended beyond the final submission date.

Attendance

Attendance at laboratory sessions is recorded. It does not contribute directly to your mark, but the sessions are where most of the practical skill in this course is developed, and students who miss them consistently perform measurably worse on the assignments.

Acceptable use of GPU and cloud compute

Shared GPU workstations are for coursework only. Do not run cryptocurrency mining, unrelated research, or long unattended jobs that block other students. Use job scheduling and checkpointing, and kill jobs you are not actively monitoring.

If you use external cloud credits, you are responsible for staying within your allocation. Do not commit API keys, tokens or personal data to any repository.

Data ethics and licensing

Every dataset you use must be lawfully obtained and appropriately licensed. You must document provenance and any preprocessing that could affect individuals represented in the data.

Do not use personal, sensitive or identifiable data without explicit approval. Where a dataset contains human subjects, consider whether your intended use is ethical as well as legal, and discuss it with the teaching team if you are unsure.

Accessibility and inclusive teaching

The course materials aim to meet accessibility standards: semantic markup, keyboard-navigable interfaces, sufficient colour contrast, and alternatives to colour-only encoding in visualisations. If you encounter a barrier, report it — it will be fixed.

Reasonable adjustments are available through the university's disability service. Please contact the teaching team early so that adjustments can be put in place before the first assessment.

Wellbeing

This is a demanding course with long training runs and open-ended projects. If you are struggling, contact the teaching team early — it is almost always easier to help before a deadline than after it.

University wellbeing and counselling services are available and are independent of your academic record.

14 Getting help

Laboratory sessions

  • Best place for hands-on debugging help.
  • Bring a minimal reproduction of the problem.
  • Demonstrators can read your code and your error trace.

Office hours

  • For conceptual questions and project direction.
  • Times published on Moodle each week.
  • No appointment needed — just turn up.

Discussion forum

  • For questions that other students will also have.
  • Search before posting — most issues are already answered.
  • Never post full solutions to assessed work.

How to ask a good technical question

State what you expected, what actually happened, the exact error message, the smallest piece of code that reproduces it, and what you have already tried. This is the same discipline required in your evaluation reports, and it is assessed.

15 Version control and document status

Syllabus revision history
Version Date Change
1.0 Semester 1 Initial syllabus published alongside the COMP329 course site.
Minor clarifications are made in place; substantive changes are announced on Moodle.

This page is the web version of the syllabus. In the event of a discrepancy, the version published on Moodle is authoritative.

👤