Sign in

Randomized Positional Encodings Boost Length Generalization of Transformers.

Anian RuossGrégoire DelétangTim GeneweinJordi Grau-MoyaRóbert CsordásMehdi BennaniShane LeggJoel Veness
Published in: CoRR (2023)
Keyphrases
  • positional information
  • orders of magnitude
  • non binary
  • efficient learning
  • finite alphabet
  • real world
  • machine learning
  • information retrieval
  • case study
  • special case
  • sat encodings
  • total length