Browse Teach Curate

Stacked LSTMs

Layering LSTMs so each layer's hidden-state sequence is the next layer's input, adding depth in feature abstraction on top of depth in time. Two to four layers is the practical sweet spot; residual connections and gradient clipping keep deeper stacks trainable.

jeremy Video

in Machine Learning multilayer Recurrent Neural Networks (RNN) architecture for feature abstraction

Multi-layer Recurrent Neural Networks (RNNs) extend single-layer architectures to achieve hierarchical feature abstraction by stacking recurrent units where hidden states from lower temporal layers s…