This module focuses on the following topics (with the understanding that foundational ML concepts are covered in the AI Algorithms and Machine Learning module and only briefly reviewed here):

Concise review of ML concepts for advanced study:
Short recap of supervised/unsupervised learning, evaluation metrics and the basic ML pipeline (without repeating Level 5 detail).

Advanced deep learning architectures:
Multi-layer perceptions (MLPs) and notions of capacity and depth.
Convolutional Neural Networks (CNNs) for structured data such as images (conceptual and architectural focus, without overlapping with pipeline engineering content in Applied AI).
Sequence models (RNN, GRU, LSTM) and an introductory, conceptual overview of attention/Transformer architectures sufficient for understanding how such models are trained and behave.

Optimisation for Deep Learning:
Gradient-based optimisation algorithms (SGD, Momentum, Adam, RMSProp).
Choosing learning rates, schedules, warm-up, batch sizes.
Practical training issues such as vanishing/exploding gradients, non-convergence and common remedies.

Regularisation and Generalisation in Deep Models:
L1/L2 regularisation, dropout, data augmentation, early stopping, batch normalisation.
Bias-variance trade-off in the context of deep networks and model capacity.
Analysing overfitting in more complex scenarios.

Representation Learning and Embeddings:
The idea of representation/feature learning in deep networks.
Embeddings for discrete entities (Ex, words, items) at a conceptual level.
High-level understanding of transfer learning and fine-tuning (without domain-specific deployment detail, which is handled in Applied AI).

Advanced ML Topics (introductory, conceptual):
High-level overview of generative models (Ex, autoencoders, VAEs, GANs).
Conceptual introduction to reinforcement learning and the agent–environment framework (without deep algorithmic detail).

Experimental Design for Advanced ML/DL:
Designing experiments, ablation studies, and baselines vs advanced models.
Using validation and cross-validation in more complex settings.
Managing computational resources, ensuring reproducibility, and recording experiments (experiment tracking).

Robustness, Shift and Simple Fairness/Risk Considerations (technical level):
Concepts of out-of-distribution data, distribution shift and drift.
Behaviour of deep models under simple noise or perturbations.
Technical-level reference to bias and fairness issues (focused on model sensitivity to data distributions), without revisiting the philosophical/legal foundations covered in the Human-Centred AI module.

Reading, Interpreting and Communicating ML/DL Research:
Understanding the structure of an ML/DL research paper (problem, method, experiments, results).
Re-designing a simplified experiment based on an existing paper.
Presenting results and analysis in written and (optionally) short oral formats.

BCS / TechSkills / Employability:
Maths principles: Optimisation, regularisation, and concepts of model capacity/complexity.
Evaluation of systems: Designing metrics, robustness tests and analyses for complex models.
System modelling: Viewing ML/DL models as components within larger systems and analysing how design decisions affect overall behaviour.
Problem solving & management/planning: Planning experiments, managing computational resources and balancing trade-offs between quality, cost and time.
Legal, social, ethical and professional issues: Technical awareness of practical risks (Ex, overfitting in safety-critical domains) and their relationship to the need for rigorous evaluation; deeper conceptual/ethical issues are addressed in Human-Centred AI modules.

1. PRACTICAL - Portfolio (50%)

You will complete a set of guided but increasingly independent experiments within practical sessions to aid building the portfolio in which you:

Implement one deep neural network architecture (Ex, MLPs, CNNs for structured/image data, RNN/sequence models or basic attention-based architectures) for a specified modelling task.

Explore different 2 regularisation techniques (Ex, dropout, weight decay, early stopping, data augmentation) and optimisation strategies (Ex, SGD, Adam, learning-rate schedules), and record the resulting behaviour.

Conduct 3 ablation-style experiments to investigate the effect of key architectural choices or hyper-parameters.

Document your results and observations in a set of notebooks/code plus a short, structured reflective commentary. Focus should be of a brief summary focus that emphasizes key technical aspects.

2. REPORT - Individual Report (50%)

You will select a problem to explore and model (for example multi-class image classification, sequence prediction, time-series forecasting, or high-dimensional structured data modelling), either from a provided list of scenarios or by agreement with the module team, and produce a brief research-style report that includes:

A clear problem and dataset description (context, sources, features, challenges) and a justification of why advanced ML/DL models are appropriate.

The design of two or more competitive advanced architectures (for example, comparison of a classical ML model with a deep neural network, or comparison of two deep architectures such as a simple CNN vs a deeper/more modern variant, or an MLP vs a sequence-aware model).

A brief experimental protocol, including:
Train/validation/test splitting strategy.
Hyper-parameter tuning approach.
Ablation experiments to examine the contribution of key components.

An evaluation using suitable metrics, comparison of models, and analysis of their behaviour (Ex, sensitivity to noise, overfitting, performance on rare classes).

Discussion of limitations, risks (such as overfitting, data shift), generalisation, and implications for future development or research.

Presentation of findings in the style of a research report, with appropriate figures/tables and clear narrative.

All teaching sessions will use practical learning development activities, combined with theory-based lectures.

This will include:

Lectures, to introduce advanced concepts, architectures and design patterns in ML/DL.

Practical classes and workshops, lab-based sessions to implement and train advanced models using modern libraries (Ex, PyTorch, TensorFlow/Keras).

Seminars / journal clubs, to discuss research papers, case studies and experimental results.

Tutorials / project supervisions, to support you in designing experiments, interpreting results and preparing assessments.

Group discussions and problem-solving sessions, to analyse training failures, unexpected model behaviour and architectural/hyper-parameter choices.

1. Critically explain advanced machine learning and deep learning architectures, optimisation methods and regularisation techniques.

Knowledge & Understanding
Research Skills

2. Implement advanced ML/DL models using appropriate frameworks, training regimes and hyper-parameter strategies.

Application & Problem-Solving
Digital Literacy

3. Plan and conduct rigorous experimental studies for ML/DL systems, including hyper-parameter tuning, ablation experiments, and evaluation of these.

Critical Reasoning & Collaboration
Personal Development & Entrepreneurship

4. Critically evaluate the limitations, robustness issues and potential risks of advanced ML/DL models, communicating these clearly to technical and non-technical audiences.

Communication
Reflection

Students will require access to:

Networked computer labs with Python, Jupyter/VS Code and ML/DL libraries (Ex, PyTorch, TensorFlow/Keras, scikit-learn, NumPy, pandas).
GPU access (local or cloud-based) where possible, to support training deep learning models at a reasonable scale.
Version control platforms (Ex, GitHub/GitLab) for managing code and experiments.
Library databases and digital collections for research literature and reference materials.
The University VLE (Blackboard) for learning materials, announcements and assessment submissions.

Géron, A. (2022) Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow, 3rd edn. O’Reilly.

Bishop, C. M. (2024) Pattern Recognition and Machine Learning, updated edn. Springer.

Francois, C. (2025) Deep Learning with Python, Third Edition, Manning Publications.

Sridhar, S. and Narishman, D. (2025) Deep Learning, Pearson Publishing.

Selected papers from leading ML conferences/journals (Ex, NeurIPS, ICML, ICLR) available via the University library’s digital subscriptions.

In this module, you will deepen your understanding of machine learning by engaging with advanced models and deep neural networks. You will explore how modern architectures are designed and trained, how their behaviour can be analysed through rigorous experimentation, and how to interpret and communicate the strengths and limitations of these models in realistic settings. Through a combination of hands-on experimentation and a research-style project, you will develop the skills needed to work with state-of-the-art ML and deep learning techniques in both professional and academic environments.