Machine Learning Fundamentals

In traditional programming you write rules and the computer applies them. In machine learning you supply examples of inputs and correct outputs, and the system derives the rules itself. The output is a model: a function that maps new inputs to predictions.

This is powerful precisely where writing rules is impractical — recognising objects in images, understanding language, predicting behaviour from many weakly-informative signals.

Understanding the fundamentals is what lets you judge whether a machine learning approach is appropriate at all. A meaningful share of proposed ML projects should be a SQL query, a rules engine, or a better-designed form.

The exams test this judgement directly. Scenarios where a simple deterministic approach meets the requirement, or where a pre-trained service already solves the problem, are common — and choosing to train a custom model is the wrong answer.

• Supervised learning — learning from labelled examples. Classification predicts a category; regression predicts a number.
• Unsupervised learning — finding structure without labels, such as clustering customers into segments.
• Feature — an input variable. Feature engineering — constructing better inputs — usually improves results more than changing the algorithm.
• Label — the correct answer for a training example.
• Training, validation and test sets — you train on one, tune on another, and measure honestly on a third that was never used for either.
• Overfitting — the model memorises the training data, performing excellently there and poorly on anything new.
• Underfitting — the model is too simple to capture the real pattern and performs poorly everywhere.
• Bias and fairness — a model trained on historical data reproduces the patterns in that data, including patterns that reflect past discrimination.

Metric selection is the most frequently tested piece of theory, because choosing the wrong one produces a model that scores well and is useless.

• Training–serving skew — features are computed one way during training and another way in production, so the model sees inputs unlike anything it learned from.
• Data drift — the input distribution changes over time and the model's assumptions quietly stop holding.
• Concept drift — the relationship between inputs and outputs changes, as when customer behaviour shifts.
• Feedback loops — the model's own predictions influence future data, reinforcing its existing biases.
• Silent failure — a degraded model keeps returning confident predictions. Nothing crashes, which is why monitoring prediction quality matters as much as monitoring latency.

• Classification — spam detection, fraud scoring, support ticket routing.
• Regression — demand forecasting, price estimation, time-to-failure prediction.
• Clustering — customer segmentation and anomaly detection.
• Recommendation — surfacing relevant products or content.
• Vision and language — extracting text from documents, classifying images, analysing sentiment.