I designed this series of interactive AI by Hand ✍️ diagrams to cover the foundational MLP architecture patterns. They are common building blocks, and they turn up again inside much larger architectures. Master these twelve and the frontier models stop looking like new inventions and start looking like arrangements of parts you already know.
MLP stands for Multi-Layer Perceptron, which I think is just jargon for a deep neural network. A perceptron is one layer: multiply by a weight matrix, then apply an activation. Multi-layer means we stack them. What changes from one pattern to the next is how layers are wired for different purposes.
1. Perceptron (free)
2. Wide (free)
3. Deep (free)
4. Binary Classifier (free)
5. Encoder (free)
6. Decoder (free)
7. Autoencoder (free)
8. Expand then Shrink (free)
9. Shrink then Expand (free)
10. GAN (free)
11. Two-Tower MLP (free)
12. Heads (free)
