Recurrent Neural Network

A neural network that processes a sequence one step at a time, using a hidden state to remember the past.

A recurrent neural network is a class of network for processing sequences: at each time step, it combines the new input with the hidden state from the previous step (a vector that records history) to compute a new state, reusing the same parameters at every step. Plain RNNs suffer from vanishing gradients on long sequences and struggle to remember anything from far back, which led to gated variants: Hochreiter and Schmidhuber proposed Long Short-Term Memory (LSTM) in 1997, and Cho and colleagues proposed the simpler Gated Recurrent Unit (GRU) in 2014. RNNs must be computed step by step and are hard to parallelize during training, so Transformers have largely replaced them in language tasks; but because an RNN only needs to update a single state at each inference step, which is cheap, they're still commonly used in robotics for policies that need to remember history.

ExampleThe imitation-learning library robomimic offers a BC-RNN algorithm: an LSTM reads in observations from the past several steps and outputs the current action, and it's often used as a comparison baseline for newer methods like Diffusion Policy.

Also called

RNN

Related

Transformer · Long Short-Term Memory / Gated Recurrent Unit · Vanishing / Exploding Gradients · State Space Model · Recurrent State-Space Model · History Encoder

Sources

Dive into Deep Learning: Long Short-Term Memory (LSTM)
Dive into Deep Learning: Gated Recurrent Units (GRU)
robomimic Documentation: Overview