Machine Learning

Why it matters

Spam filters, recommender systems, credit scores, image recognition and language models are all built on the ideas in this course. Even if you do not want to be a data scientist or AI engineer, every developer now needs to understand when a model can be trusted.

What to know

• Supervised (labelled data) vs unsupervised learning; regression vs classification.
• Linear regression, logistic regression, k-nearest neighbours and naive Bayes.
• Decision trees and ensembles: random forests, gradient boosting.
• Training, validation and test data; k-fold cross-validation.
• Overfitting, the bias-variance trade-off and regularization (Ridge, Lasso).
• Clustering (k-means) and dimensionality reduction (PCA).
• Neural networks: the perceptron, the multilayer perceptron (MLP), activation functions, backpropagation and gradient descent.
• Metrics: MAE, RMSE and R²; accuracy, precision, recall, F1 and the confusion matrix.

Practice and projects

• Compare several models on a small dataset with scikit-learn using the same cross-validation folds; always add a baseline that predicts the mean.
• Write gradient descent and a simple MLP once yourself with NumPy; that is when you understand what the library does.
• Join beginner competitions on Kaggle and read other people's notebooks.

Common mistakes

• Using test data during training or preprocessing (data leakage); steps like scaling must be learned from training data only.
• Splitting a sorted dataset without shuffling; test folds end up from a single group and the score is wrong.
• Reporting the score of one lucky split; give the mean and spread from cross-validation.
• Looking only at accuracy on imbalanced data.