Sequences have order. A recurrent neural network processes input one step at a time, carrying a hidden state forward that summarizes what came before. The same weights apply at every step, which lets the network handle variable-length input.

This recurrence gives the model memory, but it also makes training slow. Each step depends on the previous one, so computation cannot parallelize across time. Long sequences also cause gradients to vanish or explode.

Common RNN variants

• Simple recurrent networks
• Long short-term memory
• Gated recurrent units
• Bidirectional RNNs
• Deep stacked RNNs

Transformers replaced RNNs for most language tasks because attention parallelizes and captures long-range dependencies more reliably. RNNs still appear in streaming and embedded settings where their efficiency matters.