Advanced Machine Learning Course

Lesson 1 • Mathematical Prerequisites for ML

Covers linear algebra, calculus, probability, and statistics essential for understanding ML algorithms. Provides the quantitative language used throughout every subsequent chapter.

Lesson 2 • The Machine Learning Problem Framework

Defines inputs, outputs, hypothesis spaces, and loss functions as a unified problem formulation. Connects abstract maths to concrete prediction tasks.

Lesson 3 • Reinforcement Learning Fundamentals

Presents the agent-environment loop, reward signals, and policy concepts. Establishes vocabulary for the advanced RL chapter later in the course.

Lesson 4 • Unsupervised and Self-Supervised Learning

Introduces clustering, density estimation, and representation learning without labels. Prepares learners for deep generative models and self-supervised pretraining.

Lesson 1 • Cross-Validation and Model Selection

Implements k-fold, stratified, and time-series cross-validation to produce unbiased performance estimates. Prevents data leakage errors that invalidate model comparisons.

Lesson 2 • Hyperparameter Tuning Strategies

Compares grid search, random search, and Bayesian optimisation for hyperparameter selection. Equips students to efficiently configure any model architecture.

Lesson 3 • Bias-Variance Trade-off and Regularisation

Analyses bias-variance decomposition and applies L1, L2, and dropout regularisation to control generalisation. Directly addresses overfitting encountered in every subsequent model chapter.

Lesson 1 • Nearest Neighbours and Kernel Methods

Analyses k-NN, kernel density estimation, and Gaussian processes as non-parametric approaches. Highlights when parametric models fail and instance-based methods excel.

Lesson 2 • Linear and Logistic Regression

Derives closed-form and gradient solutions for regression and classification with probabilistic interpretation. Establishes the linear baseline all complex models are compared against.

Lesson 3 • Ensemble Methods: Bagging and Boosting

Implements Random Forests, AdaBoost, Gradient Boosting, and XGBoost with theoretical justification. Demonstrates state-of-the-art performance on tabular benchmarks.

Lesson 4 • Decision Trees and Splitting Criteria

Constructs decision trees using information gain, Gini impurity, and variance reduction. Provides the building block for ensemble methods in the next section.

Lesson 1 • Recurrent and Sequence Models

Derives vanilla RNNs, LSTMs, and GRUs for sequential data with vanishing gradient analysis. Prepares learners for transformer-based sequence modelling.

Lesson 2 • Attention Mechanisms and Transformers

Derives scaled dot-product attention, multi-head attention, and the full transformer architecture. Bridges classical sequence models to modern large language model pretraining.

Lesson 3 • Feedforward Neural Networks

Builds multilayer perceptrons, derives backpropagation, and applies activation functions. Establishes the computational graph abstraction used by all deep architectures.

Lesson 4 • Training Deep Networks at Scale

Addresses gradient pathologies, mixed-precision training, and distributed data parallelism. Enables students to train large models on realistic hardware budgets.

Lesson 1 • Generative Adversarial Networks

Derives the GAN minimax objective, training dynamics, and mode collapse diagnosis. Introduces conditional and progressive GAN variants for controlled generation.

Lesson 2 • Clustering Algorithms and Evaluation

Implements K-means, DBSCAN, hierarchical clustering, and spectral clustering with rigorous evaluation metrics. Connects cluster quality to downstream task performance.

Lesson 3 • Dimensionality Reduction Techniques

Applies PCA, t-SNE, UMAP, and autoencoders to compress high-dimensional data. Distinguishes linear vs. nonlinear methods and their visualisation vs. compression use cases.

Lesson 1 • Deep Q-Networks and Extensions

Implements DQN with experience replay and target networks, then extends to Double DQN and Dueling DQN. Demonstrates deep RL on high-dimensional observation spaces.

Lesson 2 • Tabular RL: Dynamic Programming and TD

Solves MDPs with value iteration, policy iteration, Q-learning, and SARSA. Provides exact solutions that approximate methods in later sections are compared against.

Lesson 3 • Model-Based RL and Planning

Contrasts model-free and model-based RL, implements Dyna-Q, and introduces world models. Addresses sample efficiency limitations of pure model-free approaches.

Lesson 4 • Actor-Critic and PPO

Implements advantage actor-critic (A2C/A3C), PPO clipping, and GAE for stable policy updates. Covers the dominant algorithms used in modern RL applications.

Lesson 1 • ML Pipelines and Orchestration

Automates training, evaluation, and deployment workflows using pipeline orchestration tools. Enables reproducible, auditable ML experiments at team scale.

Lesson 2 • Scalable Training Infrastructure

Configures distributed training clusters, GPU resource scheduling, and cost-efficient spot instances. Addresses infrastructure bottlenecks that limit model scale.

Lesson 3 • Monitoring and Model Maintenance

Detects data drift, concept drift, and performance degradation in live systems. Implements alerting and retraining strategies to sustain model quality over time.

Lesson 4 • Model Serialisation and Serving

Packages models using standard formats, deploys REST and gRPC endpoints, and optimises inference latency. Connects model training artifacts to real-time prediction services.

Lesson 1 • Model Interpretability and Explainability

Applies SHAP, LIME, integrated gradients, and attention visualisation to explain model decisions. Connects explainability to stakeholder trust and regulatory compliance.

Lesson 2 • Emerging Frontiers in ML

Surveys foundation models, multimodal learning, and causal ML as active research directions. Positions students to evaluate and adopt emerging techniques critically.

Lesson 3 • Privacy-Preserving Machine Learning

Implements differential privacy, federated learning, and secure aggregation to protect training data. Addresses data minimisation requirements in privacy-sensitive deployments.

Lesson 4 • Fairness Metrics and Bias Mitigation

Defines demographic parity, equalised odds, and calibration fairness metrics, then applies pre/in/post-processing mitigations. Prepares learners for regulatory and ethical audits.