AI & Machine Learning Study Guide
Five units from the ML workflow through data preparation and evaluation to unsupervised methods and responsible AI. Each unit explains what to know, how to apply it, what typically goes wrong, and one study move that makes the material stick.
Five units
Foundations and workflow, data preparation and features, supervised learning, model evaluation, then unsupervised methods and responsible AI.
Python-based
Assumes the Python course or equivalent comfort with functions, collections, and pandas.
No single exam
Assessed by projects, analysis write-ups, and practice sets.
Where it leads
University machine learning and data science coursework, and portfolio projects.
Units
Unit 1 is the largest by a wide margin. The rest move faster, and the last one is the bridge into calculus.
Unit 1 · Foundational
What machine learning is, when it is the wrong tool, and the order the work happens in. Framing decisions made here determine whether anything downstream can be evaluated.
Unit 2 · Where the effort is
The unit that decides whether a project succeeds. A model can only learn what its data contains, including the data errors and the historical bias.
Unit 3 · Heavy
The algorithms themselves, treated as tools with known behavior rather than as the point of the course. What matters is knowing what each assumes and how each fails.
Unit 4 · Most assessed
Whether a result can be believed. This is the unit that separates a student who can run a model from one who can report an honest finding.
Unit 5 · Assessed seriously
Finding structure without labels, and the obligations that come with deploying any of this. The responsibility material is assessed with the same seriousness as the technical content.
Mathematical Practices
Four habits that separate a result you can believe from a number you cannot: frame first, inspect the data, beat a baseline, distrust one number.
Say what is being predicted, from what, and what success would mean numerically. A project that skips this produces a model nobody can evaluate.
Shape, types, missing values, and class balance, before any algorithm. Most project failures are visible in the first ten minutes of inspection.
Always know what a trivial rule scores. A model that cannot beat "predict the most common class" has not learned anything, whatever its accuracy looks like.
A single accuracy figure hides class imbalance and uneven errors. Ask for the breakdown before believing any headline.
Assessment
Two lists: one for selected-response work, one for anything you have to write out.
Study plan
Four moves, in the order that pays off.
Frame, prepare, train, evaluate, deploy, monitor. Being able to say where you are in that sequence prevents most confusion.
Leakage is the failure that produces spectacular results and useless models. Train yourself to ask whether each feature would exist at prediction time.
Split first, fit transformations on training data only. Get this wrong and every score you report is inflated.
Precision, recall, and the cost of each error type are the vocabulary of every serious evaluation discussion.
Real problems are imbalanced. Practise on a dataset where accuracy is misleading until the trap is obvious.
For every project, write what the model cannot do and who it might fail. This is assessed and it is also the honest thing to produce.
Curriculum scope
Machine learning is taught here as a workflow rather than a catalogue of algorithms. The recurring question is not which model to use but whether the result can be believed, which is why evaluation and data preparation carry more weight than model internals.
Essential foundations
Included in this sequence
Compare with the school syllabus
Glossary
Supervised learning
Training on labelled examples so a model can predict the label for new data.
Leakage
A feature that would not be available at prediction time, or that encodes the answer.
Baseline
The score a trivial rule achieves, against which any model must be compared.
Overfitting
Fitting noise in the training data, shown by a gap between training and test performance.
Underfitting
A model too simple to capture the pattern, shown by poor performance everywhere.
Precision
The proportion of predicted positives that are actually positive.
Recall
The proportion of actual positives that the model found.
Class imbalance
One class dominating the dataset, which makes accuracy misleading.
Proxy variable
A feature correlated with a protected attribute, carrying its information indirectly.
Drift
Change over time in the data a deployed model sees, degrading its accuracy.
