LSTM

1997

Computer Science (theoretical)Machine Learning TheoryArchitecture Componentfoundational

Recurrent unit with gated memory cells (Hochreiter & Schmidhuber, "Long Short-Term Memory," Neural Computation 9(8), 1997) that solved the vanishing-gradient problem in training recurrent networks over long sequences, making deep sequence models practical for the first time.

Originators

Hochreiter, S.
Schmidhuber, J.

Landmark Paper

W2064675550 ↗

Not retracted (OpenAlex)

Checked 2026-09-19 — interim signal only, see docs/BASIC_ROADMAP.md Phase 10

Connections

is component of Sequence to Sequence Learning
basis: reasoned
Sutskever, Vinyals & Le's Seq2Seq (ref [35]) uses multilayered LSTM cells as the encoder and decoder; LSTM is the specific recurrent unit the framework is built from.
historically preceded Mamba
basis: reasoned
Appendix A of the Mamba paper: "Gating originally referred to the gating mechanisms of RNNs such as the LSTM (Hochreiter and Schmidhuber 1997)... Theorem 1 is an improvement of [LSSL]... generalizing to the ZOH discretization and input-dependent gates." Mamba's own Theorem 1 shows its selection mechanism recovers a gated-RNN recurrence as a special case, with LSTM named as the specific example of classical RNN gating.