The question behind the course

In applied deep learning the hard part is rarely “can I write a model?”. It is:

• Is the model learning the right thing?
• Why does it fail on some inputs?
• What does it actually see?
• How do I explain its decisions?

Deep learning is not just about accuracy — it is about understanding, diagnosing, visualising and improving.

01 — CNNs: learning visual features

Convolutional networks for image classification, with filter inspection and activation maps to see how early layers detect edges and textures and deeper layers build up shapes.

02 — U-Net: semantic segmentation

Instead of one label per image, the model predicts a mask for every pixel. This required dense prediction, skip connections to preserve spatial detail, and careful evaluation at object boundaries.

03 — Variational Autoencoders

Compressing images into a low-dimensional latent space and reconstructing them — a hands-on look at representation learning, generative modelling and the trade-off between compression and reconstruction quality.

04 — Sentiment analysis

Text classification from word patterns: tokenisation, embeddings and sequence models, with frequency analysis to interpret what drives each class.

Diagnosing training

Training curves reveal more than a final score — over- and under-fitting, learning-rate problems and data issues all show up here first.

What I learned

• Data quality, preprocessing and architecture choices matter more than model size
• Model inspection matters as much as model performance
• Visualisation exposes bias, overfitting and poor representations
• A model is useful when you can explain how it behaves and where it fails