Neural Network
Layers of weighted sums and non-linear activations, trained by backpropagation.
- One hidden layer + non-linearity = universal function approximator.
- Backprop computes parameter gradients efficiently.
- Substrate for everything in the deep learning track.
When to use: When linear models cap out and you have enough data to fit non-linear interactions.
Example: 3-layer MLP on tabular fraud: input(50) → 64 (ReLU) → 32 (ReLU) → 1 (sigmoid). Trained with Adam for 30 epochs, beats logistic regression by 4 AUC points.