Welcome to Module 4 — deep learning architectures
Before we begin
Module 3 gave you a fully connected network on MNIST. Real problems have structure:
Images have local patterns (edges, textures). Language has order and meaning beyond bag-of-words.
Module 4 introduces the architectures that dominated deep learning before transformers — and whose ideas still live inside modern models:
- CNNs for images
- RNNs for sequences
- LSTM / GRU for longer memory
- Embeddings for words (critical for GenAI)
Figure
Module 4 at a glance
What Module 4 covers
| Topic | What you will understand |
|---|---|
| CNN | Conv filters, pooling, why spatial structure matters |
| RNN | Hidden state over time, sequences |
| LSTM / GRU | Gates, long-range context |
| Embeddings | Words as vectors; pre-trained vs learned |
Before you start
Required: Module 3 project or comfort with PyTorch training loops.
Install before the project:
pip install torch pandas matplotlib scikit-learn
Lessons 1–5 are reading. Lesson 7 is the coding project.