Learn patterns via Contrastive Divergence — Gibbs sampling between visible and hidden layers
An RBM has no intra-layer connections. Contrastive Divergence (CD-1): clamp visible, sample hidden (positive phase), then reconstruct visible from hidden and re-sample hidden (negative phase). Weight update: ΔW = η(⟨vh⟩_data - ⟨vh⟩_model).