Conceptual
Login

Encoder and Vision Models That Still Ship Absolute Positional Embeddings

BERT-style retrieval encoders and standard ViTs still use learned absolute positions, so 'absolute is obsolete' is the wrong lesson to take away