Architectures · intermediate · concept 91 of 311 · Updated Nov 4, 2026
Gated Recurrent Unit (GRU)
A GRU is a recurrent unit that uses two gates to decide what to keep and what to forget as it reads a sequence. It has fewer parameters than an LSTM and often performs about as well, so it trains faster and suits smaller datasets. Transformers replaced both for most language work, but GRUs remain useful on devices and in time series.
Key terms
Update gateReset gateVanishing gradientsSequence modelingLSTM comparison
Learn these first
Where you meet it in the real world
On-device speech, sensor and time series models, lightweight language models
Guides and articles
Gated recurrent units ↗
Aston Zhang and others
Understanding LSTM Networks ↗
Chris Olah