Multilayer networks
Stack layers so each layer’s outputs become the next layer’s inputs (MLP).
A multilayer perceptron (MLP) builds hierarchical features: early layers detect simple patterns; later layers combine them. Depth (how many layers) and width (how many units) control capacity — more capacity can fit harder functions but overfits more easily.
Tiny MLP
Inputs $(x_1,x_2)$ → hidden layer of 3 neurons → one output neuron. Each hidden unit draws a soft half-plane in the input space; the output neuron mixes those three “votes” into a curved decision. XOR-like patterns that one neuron cannot separate become possible with a hidden layer.
Naming
“Deep learning” usually means many such layers (and often specialized architectures). The MLP is the conceptual starting point.
Key takeaways
- Layers compose functions.
- Hidden units learn intermediate features.
- Architecture is a design choice — deeper ≠ always better.