Deep Learning

About this course

A modern deep-learning course built in PyTorch. Learners go from multilayer perceptrons and backpropagation through convolutional and recurrent networks to attention and the transformer, then learn what it takes to train well and efficiently in 2026: normalisation and residual design, AdamW and schedules, mixed precision, distributed data parallelism, scaling laws, state-space and mixture-of-experts alternatives, and reproducible experiments with honest ablations.

For: student, software engineer, data professional, ml practitioner.

Learning outcomes

On successful completion you will be able to:

1. 1Explain how deep networks learn, including backpropagation, initialisation, normalisation, optimisation and the factors that govern generalisation at scale.
2. 2Implement and train neural networks in PyTorch, from a multilayer perceptron to convolutional, recurrent and transformer models.
3. 3Diagnose training problems from loss curves, gradient statistics and activations, and fix them with appropriate techniques.
4. 4Apply transfer learning and pretrained models to vision and language tasks with limited labelled data.
5. 5Train efficiently on modern accelerators using mixed precision, input-pipeline optimisation, compilation and multi-GPU data parallelism.
6. 6Evaluate architectural choices (convolutional, recurrent, transformer, state-space, mixture-of-experts) for a problem, citing evidence on accuracy, compute and data requirements.
7. 7Conduct and report a deep-learning experiment reproducibly, including ablations, compute cost and an honest account of limitations.

Modules and lessons

1. Week 1Neural network fundamentals⌄

   • From perceptrons to multilayer networks
   • The PyTorch training looplab
2. Week 2Making training work⌄

   • Initialisation, normalisation and residual connections
   • Optimisation and regularisation in practicelab
3. Week 3Convolutional networks and vision⌄

   • Convolutional networks
   • Transfer learning and modern visionlab
4. Week 4Sequences and representations⌄

   • Embeddings and representation learning
   • Recurrent networks and their limits
5. Week 5Attention and transformers⌄

   • The attention mechanism
   • The transformer architecturelab
6. Week 6Training at scale⌄

   • Mixed precision and training performancelab
   • Distributed training basics
7. Week 7Scale and architecture choice⌄

   • Scaling laws and generalisation
   • Beyond the dense transformer: state-space models and mixture of experts
8. Week 8Robustness, efficiency and reporting⌄

   • Robustness, calibration and uncertainty
   • Efficient models and reproducible experimentslab

Assessment

• Quiz (15%)

  Eight weekly quizzes drawn from the lesson quiz banks; best seven count.

• Lab (35%)

  Six PyTorch labs with automated checks, from a reusable training loop to a profiled, mixed-precision transformer.

• Project (35%)

  Train or fine-tune a model on a real task and write it up as a short paper with baselines, ablations, compute cost and limitations.

• Viva (15%)

  Twenty-minute technical interview on the project and its training curves, held live on camera.

Pass mark 60%. Grades: Distinction, Merit, Pass.