Sequence model architectures

• One-to-One: Single input -> Single output (e.g., basic RNN step)
• Many-to-One: Sequence input -> Single output (e.g., sentiment analysis, next-word prediction based on prior words). RNN can be used by taking the final hidden state/output
• One-to-Many: Single input -> Sequence output (Less common)
• Many-to-Many (Sequence-to-Sequence / Seq2Seq): Sequence input -> Sequence output
  • Variant 1: Output at each input step (e.g., video frame classification)
  • Variant 2 (Encoder-Decoder): Process entire input sequence first, then generate entire output sequence (e.g., machine translation)