CSUC301 Machine Learning

Each file opens in any browser with no server and no network connection. Every guide is self-contained, so a single unit can be sent to a friend and will still work.

Mathematical Foundation

The machinery every later unit runs on: a model is a set of numbers, a loss says how wrong they are, and calculus says which way to move them. Built entirely on five students, hours practised against marks, so the arithmetic stays checkable by hand.

1.1 Systems of linear equations · 1.2 Convex and non-convex functions · 1.3 Loss functions and minimisation · 1.4 Gradient and gradient descent, batch and SGD · 1.5 Entropy and information gain · 1.6 Bias–variance trade-off · 1.7 Cost trade-offs and regularization · 1.8 Activation functions

8topics
6 hrs · CO1
ready

Supervised Learning: Models

Six algorithms fitted to the same eight students, and the finding that all six reach 100% training accuracy while disagreeing about every point in between. Split from 2B because 13 teaching hours does not fit in one file.

2.1 Architecture, model selection, hyperparameters · 2.2 Multiple regression, ridge and lasso · 2.3.1 Logistic regression · 2.3.2 k-NN · 2.3.3 Support vector machines · 2.3.4 Decision trees · 2.3.5 Bagging, forests, boosting · 2.3.6 Bayesian classifiers · 2.3.7 Choosing between them

9topics
13 hrs · CO2, CO5
ready

Supervised Learning: Evaluation & Responsible AI

The audit. Ten test students whose accuracy stays fixed at 0.7000 while the classifier changes from mediocre to good, thresholds derived from costs rather than convention, and a four-line proof that three fairness criteria cannot hold at once.

2.4.1 Confusion matrix · 2.4.2 Precision, recall, F-scores · 2.4.3 ROC, AUC, precision–recall · 2.4.4 Choosing a threshold · 2.4.5 Regression metrics · 2.4.6 Calibration and honest reporting · 2.5 Semi- and self-supervised learning · 2.6 Fairness, interpretability, privacy

8topics
CO2, CO5
ready

Unsupervised Learning

Twelve students with the labels taken away. Four clustering methods that all agree on three groups, four validation indices that cannot tell you whether those groups are right, and a PCA that optimally retains 56.7% of the variance while collapsing two of the three clusters onto the same point.

3.1.1 K-Means · 3.1.2 Hierarchical clustering · 3.1.3 DBSCAN · 3.1.4 Gaussian mixtures and EM · 3.2 Internal validation: elbow, silhouette, Davies–Bouldin, Calinski–Harabasz · 3.3 PCA with derivation, reconstruction error, t-SNE · 3.4 External metrics: ARI, NMI, homogeneity, completeness, V-measure · 3.5 Hyperparameter tuning without labels

8topics
7 hrs · CO2, CO3
ready

Neural Networks

Four points that no linear model can ever fit, and a nine-parameter network that solves them exactly with whole-number weights. Backpropagation worked by hand to a loss of exactly ln 2, one step at η = 1 that kills a hidden unit, and the finding that the architecture proven sufficient for XOR finds it from only 16.5% of random starts.

4.1 Perceptron, forward propagation, losses · 4.2 Convolution and pooling · 4.3 Activation trade-offs · 4.4 Backpropagation and parameter counts · 4.5 SGD, momentum, Adam · 4.6 Dropout, batch norm, early stopping · 4.7 Initialisation and tuning · 4.8 Learning curves, calibration, fairness

8topics
6 hrs · CO2, CO3
ready

Deep Learning Foundation

Three small shapes read two ways. A convolutional network sees each all at once and separates them easily; a recurrent one reads them row by row and loses a first-row difference by a factor of 39,000 in five steps. A gated cell holds it exactly. Closes with three AUCs of 1.0000 sitting beside an accuracy of 0.8333.

5.1 Deep learning versus machine learning · 5.2 CNN architecture: 1×1 bottlenecks, residual connections, receptive fields · 5.3 Transfer learning and where to draw the freeze line · 5.4 RNN, vanishing gradients over time, LSTM and GRU · 5.5 Top-K accuracy, one-vs-rest AUC, convergence curves

5topics
8 hrs · CO2, CO3
ready

Reinforcement Learning Fundamentals

Six squares, four actions, and no instructions. Every optimal value is a whole power of γ, and the hand-worked Q-learning trace shows reward crawling backwards one step per episode — nothing at the start square moves until episode four. Ends with a measured result that cuts both ways: greedy exploration wins outright here, and one small extra reward reverses it completely.

6.1 Agent, environment, reward and the discount · 6.2 Markov decision processes and the Bellman equations · 6.3 Q-learning derived, traced by hand, and the cost of exploring · 6.4 Cumulative reward, regret, episode length and honest reporting

4topics
5 hrs · CO4
ready

How to use these

The guides are written so a beginner and a strong student can read the same page and both get something. Each concept appears three times at increasing depth, so you choose your own path through it.

Machine Learning study guide

Units