Teaching

Machine Learning

ELEC60019/ELEC70059, Autumn 2026/2027

Machine learning underpins technologies from medical diagnosis and autonomous systems to large language models. These systems rest on a common set of ideas: how a model learns from data, why it generalises beyond its training examples, and what goes wrong when it does not. This course teaches those ideas in depth, so that you can understand, assess and build on modern methods rather than only use them.

The course focuses on the mathematics behind machine learning. It starts with the components of learning, problem formulation, and simple predictors such as the perceptron and linear regression. It then asks when learning is feasible, covering empirical risk minimisation, generalisation bounds, the Hoeffding and VC inequalities, and the trade-off between performance and complexity. These concepts underlie current questions such as why large neural networks generalise and how much data a model needs. They also explain overfitting and regularisation, which reappear in deep learning as weight decay and dropout.

You will then study support vector machines and kernels, logistic regression, gradient descent, multilayer perceptrons and backpropagation, nearest neighbour classification, principal component analysis, and clustering with k-means, kernel k-means and Gaussian mixture models. Gradient descent and backpropagation are how deep networks and language models are trained today. Nearest neighbour search, dimensionality reduction and clustering are routinely used to retrieve, visualise and organise the embeddings these models produce. Python exercises in Colab give practice applying the theory to real problems.

Machine Learning for Safety-Critical Decision Making

ELEC70122, Spring 2026/2027

Machine learning is increasingly used to inform high-stakes decisions in healthcare, finance, criminal justice and defence. In these settings an accurate prediction is not enough: errors matter when they trigger actions, their consequences are rarely equal, and many failures only appear after deployment, when the data, population or environment has changed. Models that exploit spurious correlations, report overconfident probabilities or optimise the wrong objective can cause real harm.

Foundation models and large language models have made powerful general-purpose systems widely available, but pretraining alone does not ensure safe behaviour. Building models that can be trusted to act, or to support human decision makers, requires reasoning carefully about uncertainty, causality and evaluation.

In this course, you will learn how to build statistically sound ML models for risk-sensitive settings and how to judge when their outputs can be safely acted on. Topics include Bayesian modelling and posterior predictive decisions, calibration and conformal prediction, robustness to distribution shift, learning to defer to human experts, safe and offline reinforcement learning with constrained MDPs, off-policy evaluation and risk-sensitive value estimation, causal inference (causes of effects and effects of causes, treatment effect estimation), and alignment and auditing of foundation models through RLHF and inverse reinforcement learning. Methods are illustrated with real applications, including ICU decision making and HIV treatment optimisation.