Module 6 · Unit 20 · Final Unit

Course Integration

This is where the pieces become a whole. Unit 20 draws together twenty units across six modules, prepares you for the final examination, and supports the completion of your capstone project — the single artefact that demonstrates everything you have learned.

Start the unit → Learning outcomes

Unit at a glance

9 TutorialsReview, workshop and exam prep
6 Modules revisitedFoundations through responsible ML
6 Hours of studyReview, practice and project work
1 Capstone + final examBoth culminate in this unit

01 Overview

Twenty units ago you began with a question: what does it mean for a machine to learn? You have since worked through data preparation and the Python ecosystem, supervised learning and model evaluation, decision trees and ensembles and support vector machines, clustering and dimensionality reduction and anomaly detection, neural networks and deep learning, and — in this final module — deployment, ethics, governance and fairness.

That is a lot of material. The purpose of Unit 20 is to make it cohere. A course that leaves you with twenty disconnected topics has failed, however well each individual unit was taught. What matters is whether you can move fluidly between the layers — recognising that a choice about a fairness metric constrains your evaluation strategy, that a serving pattern determines what you can monitor, that a data collection decision made in week one becomes an ethical liability in production.

This unit has three jobs. First, to review the conceptual spine of the course so that the connections are explicit rather than implied. Second, to prepare you for the final examination — not by drilling facts, but by working through the kinds of integrative problems the exam is designed to test. Third, to support your capstone, which is where the course asks you to demonstrate the full lifecycle from problem selection to ethics analysis on a project of your own choosing.

Why this unit matters

Professional machine learning work does not arrive pre-sorted into units. A real project asks you simultaneously to choose a model, justify it, evaluate it fairly, deploy it responsibly and account for its effects. Unit 20 is the rehearsal for that — the point at which the course stops teaching you topics and starts asking you to practise a discipline.

The four threads that run through the course

Concepts — how do models learn from data? Learning paradigms, loss functions, optimisation, generalisation and the bias–variance trade-off.
Skills — how do you build and evaluate a model? Data preparation, algorithm selection, training, cross-validation and honest performance measurement.
Practice — how do you run a model in the world? Production pipelines, serving patterns, monitoring, drift detection and the MLOps lifecycle.
Responsibility — how do you govern what you have built? Bias, privacy, transparency, fairness metrics, regulation and organisational accountability.

Every unit in COMP328 belongs to one or more of these threads. If you can place a topic on this map and explain how it connects to its neighbours, you have understood the course.

02 Learning outcomes

By the end of Unit 20 — and therefore by the end of COMP328 — you should be able to do the following. These outcomes describe the integrated capabilities the course is designed to produce.

Knowledge — what you will understand

Explain how the major topics of the course relate to one another, and place any given technique within the overall machine learning lifecycle.
Describe the bias–variance trade-off and connect it to model selection, regularisation, ensemble methods and evaluation strategy.
Explain how the choice of learning paradigm — supervised, unsupervised, deep — determines what questions a system can answer.
Articulate how deployment, fairness and governance constraints feed back into modelling and data decisions.
Summarise the ethical and regulatory landscape within which modern machine learning systems operate.

Skills — what you will be able to do

Given an unfamiliar problem, select an appropriate approach and justify it in terms of data, task and constraints.
Design an end-to-end machine learning solution covering data, modelling, evaluation, deployment and monitoring.
Evaluate a model's performance honestly, including across subgroups and under distribution shift.
Produce a complete capstone deliverable: problem framing, data work, model, evaluation, deployment plan, ethics analysis and governance plan.
Answer integrative examination questions that require combining concepts from multiple modules.

Professional competencies — how you will think

Approach an ambiguous problem by decomposing it into data, modelling, evaluation and operational questions.
Recognise when a question is technical, when it is a value judgement, and when it is both.
Communicate a complete machine learning project — including its limitations and risks — to a mixed audience.
Continue learning independently: read papers, follow tooling, and evaluate new methods against the foundations established in this course.

03 The course map

The table below is the single most useful revision artefact in this unit. It maps every module and unit to its core ideas. If you can talk through this table from memory — explaining, not reciting — you are ready for the exam.

COMP328 — complete course map
Module	Units	Core ideas
1 · Foundations	1–3	What machine learning is; the ML lifecycle; data, features and labels; train/validation/test splits; the Python ecosystem (NumPy, Pandas, Matplotlib, scikit-learn).
2 · Supervised Learning	4–6	Regression and classification; least squares; the sigmoid and decision boundaries; accuracy, precision, recall, F1, ROC; cross-validation; the bias–variance trade-off.
3 · Classical Algorithms	7–9	Decision trees, entropy and information gain; ensembles, bagging and feature importance; support vector machines, margins and kernels.
4 · Unsupervised Learning	10–13	K-means and hierarchical clustering; the curse of dimensionality and PCA; association rules and Apriori; anomaly detection and its applications.
5 · Neural Networks	14–16	Artificial neurons and backpropagation; activation functions; regularisation and dropout; convolution, pooling and CNN architectures.
6 · Responsible & Applied ML	17–20	Deployment pipelines and serving patterns; monitoring and drift; bias, privacy, transparency and accountability; fairness metrics and mitigation; regulation and governance.

The lifecycle as a single picture

Read across the course map and a single process emerges. Every machine learning project moves through the same broad stages, and every technique you have learned belongs somewhere on this path.

Frame the problem
What decision is being made? Who is affected? Is machine learning the right tool at all?
Collect and prepare data
Sources, quality, features, labels, splits. Where most bias enters and most time is spent.
Select and train a model
Algorithm choice, loss function, optimisation, regularisation. The part everyone thinks of as "machine learning".
Evaluate honestly
Held-out performance, cross-validation, subgroup analysis, calibration. Where good projects are separated from lucky ones.
Deploy and monitor
Serving pattern, pipeline, versioning, drift detection. Where a model stops being an experiment.
Govern and account
Fairness evaluation, documentation, audit, oversight. Where the system becomes answerable to the people it affects.

Study tip

The most common mistake in revision is to study units in isolation. Instead, pick a single technique — say, logistic regression — and trace it through all six stages above. What data does it need? How is it trained? How is it evaluated? How is it served? What fairness questions does it raise? Doing this for three or four techniques will teach you more than re-reading twenty units.

04 Reading roadmap

Unit 20 is delivered as nine tutorials. Tutorials 1–7 review the six modules in sequence. Tutorial 8 is a working session on the capstone. Tutorial 9 prepares you for the final examination. If you are short on time, prioritise T-01, T-08 and T-09 — but the module reviews are where the connections become visible.

T-01

The Course in Retrospect

The four threads — concepts, skills, practice, responsibility — and how the twenty units map onto them.

T-02

Foundations Review

Module 1 revisited: the ML lifecycle, data fundamentals and the Python ecosystem, with the connections to later modules made explicit.

T-03

Supervised Learning Review

Module 2 revisited: regression, classification and evaluation, unified by the bias–variance trade-off and the question of generalisation.

T-04

Classical Algorithms Review

Module 3 revisited: trees, forests and SVMs as three different answers to the same question — how to draw a boundary that generalises.

T-05

Unsupervised Learning Review

Module 4 revisited: clustering, dimensionality reduction, association and anomaly detection as tools for unlabelled data.

T-06

Neural Networks Review

Module 5 revisited: from a single neuron to convolutional architectures, and what deep learning changed — and did not change.

T-07

Responsible ML Review

Module 6 revisited: deployment, ethics, governance and fairness as four faces of one obligation — running systems that deserve trust.

T-08

Capstone Workshop

A structured working session on the final project: consolidating feedback, completing the deliverables and preparing the presentation.

T-09

Exam Preparation

Integrative problem types, worked examples, and a revision strategy that builds understanding rather than recall.

05 Assessment

Unit 20 carries no new topic assessment of its own — the course is assessed through the five assignments, the twenty unit quizzes, the final examination and the capstone project. This unit is where the final two culminate.

Quiz questions

Integrative questions spanning all six modules. This quiz is deliberately the most demanding in the course.

40%

Capstone project

The full lifecycle: problem framing, data work, modelling, evaluation, deployment plan, ethics analysis and governance plan, plus presentation.

30%

Final examination

Integrative written examination covering the whole course. Designed to test reasoning and synthesis, not recall.

How the course outcomes are assessed

Capstone deliverable structure
Section	What it contains	Drawn from
1 · Problem framing	Decision, stakeholders, success criteria, why ML is appropriate	Units 1, 18
2 · Data	Sources, provenance, cleaning, features, splits, limitations	Units 2, 3
3 · Modelling	Approach selection, baselines, training, hyperparameter choices	Units 4–9, 14–16
4 · Evaluation	Metrics, cross-validation, subgroup performance, error analysis	Units 6, 10–13
5 · Deployment plan	Serving pattern, pipeline, versioning, monitoring, drift response	Unit 17
6 · Ethics analysis	Stakeholders, harms, bias audit, privacy review, mitigations	Unit 18
7 · Governance plan	Fairness definition and metric, audit results, regulatory classification, oversight	Unit 19
8 · Presentation	Communicating the project to a mixed technical and non-technical audience	All units

Before you submit

The single most common weakness in capstone submissions is a missing or thin evaluation section. A project that builds an impressive model but cannot say honestly how well it performs — including on the subgroups it might harm — will not score well, however sophisticated the technique. Reserve real time for Section 4.

06 Summary

Unit 20 has done three things: reviewed the course, prepared you for the examination, and supported the capstone. The ideas worth carrying out of COMP328 are these:

Machine learning is a lifecycle, not an algorithm. Problem framing, data work, evaluation, deployment and governance are all part of the discipline. Technique is one component among several.
Generalisation is the central problem. The bias–variance trade-off, regularisation, cross-validation, ensembles and deep learning architectures are all attempts to answer the same question: will this work on data the model has never seen?
Evaluation is where rigour lives. A model is only as good as the honesty of its evaluation. Held-out sets, subgroup analysis and calibration are not formalities.
Deployment changes the problem. Once a model has users, the questions shift from accuracy to reliability, latency, drift and failure modes.
Fairness involves choices, not solutions. The impossibility results show that competing fairness definitions cannot all be satisfied. Someone must choose, and that choice must be accountable.
Responsibility is not optional. Bias, privacy, transparency and accountability are not add-ons to a machine learning system — they are properties of how it was designed and governed.

You have completed COMP328 when you can…

…take an ambiguous real-world problem, decide whether machine learning is appropriate, design and evaluate a solution, plan its deployment and monitoring, analyse who it affects and how, and communicate the whole thing honestly — including what you do not know.

Where to go from here

This course has given you the foundations. What comes next depends on your direction. If you are continuing in machine learning, the natural next steps are deeper study of optimisation and probabilistic modelling, modern deep learning architectures, and the specialised literature in whichever domain you care about. If you are moving into practice, the most valuable next skill is likely to be software engineering discipline — testing, version control and system design — because that is what separates a model that works from a system that works.

Whichever path you take, the habits this course has tried to build are the ones that matter: ask who is affected, measure honestly, document what you did, and be willing to say what your system cannot do.

Two deliverables remain: the capstone and the exam.

Begin with T-01 — The Course in Retrospect — to see how the twenty units fit together, then move to T-08 for the capstone workshop. Nine tutorials, roughly six hours, and the end of COMP328.

Start T-01 → Capstone project