← Back to Basic Science

Sequence to Sequence Learning

2014
Computer Science (theoretical)Machine Learning TheoryFrameworkfoundational

General encoder-decoder framework using multilayered LSTMs to map an input sequence to an output sequence of different length (Sutskever, Vinyals & Le, "Sequence to Sequence Learning with Neural Networks," NeurIPS 2014). Established neural sequence transduction as competitive with statistical machine translation and became the architecture attention was first added to.

Originators

  • Sutskever, I.
  • Vinyals, O.
  • Le, Q.V.

Landmark Paper

W2130942839 ↗
Not retracted (OpenAlex)

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

Connections

  • is prerequisite for Attention Mechanism
    basis: reasoned

    The fixed-length context-vector bottleneck in Sutskever et al.'s Seq2Seq (ref [35]) is the specific limitation Bahdanau et al.'s attention mechanism (ref [2]) was designed to remove, by letting the decoder attend over all encoder states instead of one compressed vector.