σ(Wf·[h,x] + bf)
Decides what to forget
σ(Wi·[h,x] + bi)
Decides what to add
Ct = ft⊙Ct-1 + it⊙C̃t
Long-term memory
σ(Wo·[h,x] + bo)
Decides output
Purpose: Solves vanishing gradient problem in RNNs. Maintains long-term dependencies in sequences.
Gates: Forget (remove), Input (add), Output (expose) information from cell state
Cell State: Highway for information flow across time steps
Applications: Language modeling • Machine translation • Speech recognition • Time series • Music generation