Foundations of Artificial Intelligence and Machine Learning
What this domain tests
This domain checks whether you can pick the correct learning paradigm for a problem, reason about generalisation instead of memorisation, and read a learning curve or score report without confusing the data splits or the metric.
Core rules
- Supervised learning maps labelled inputs to outputs (classification, regression); unsupervised learning finds structure (clustering, dimensionality reduction, anomaly detection); reinforcement learning optimises a policy from reward signals over time.
- Generalisation error decomposes into bias (model too simple, underfits) and variance (model too sensitive to the training sample, overfits), plus irreducible noise.
- Control capacity with regularisation: L1/L2 penalties, early stopping, dropout, pruning, and limits on tree depth or feature count.
- Use a three-way split: train fits parameters, validation selects hyperparameters and model family, test estimates final performance and is touched once.
- The loss optimised during training is not automatically the metric the business cares about; cross-entropy is standard for classification, MAE/MSE for regression.
Worked example
A classifier scores 0.96 on training data and 0.71 on validation. The large gap plus a much lower validation score means high variance (overfitting); underfitting would show both scores low and close together. First actions: strengthen regularisation, reduce complexity, or add and augment training data, then re-tune on validation. If both scores had been 0.71, the diagnosis flips to high bias and you would add capacity or better features instead.
Traps that cost marks
Repeatedly tuning against the test set, which leaks information and inflates the reported score. Declaring a model good on accuracy alone when classes are imbalanced. Treating correlation between two features as proof that one causes the other. Forgetting that a validation score is itself an estimate with variance on small samples.