Positional Embeddings in Transformer Models: Evolution from Text to Vision Domains

摘要

Positional encoding has become an essential element in transformer models, addressing their fundamental property of permutation invariance and allowing them to understand sequential relationships within data. This blog post examines positional encoding techniques, emphasizing their vital importance in traditional transformers and their use with 2D data in Vision Transformers (ViT). We explore two contemporary methods—ALiBi (Attention with Linear Biases) and RoPE (Rotary Position Embedding)—analyzing their unique approaches to tackling the challenge of sequence length extrapolation during inference, a significant issue for transformers. Additionally, we compare these methods' fundamental similarities and differences, assessing their impact on transformer performance across various fields. We also look into how interpolation strategies have been utilized to enhance the extrapolation capabilities of these methods; we conclude this blog with an empirical comparison of ALiBi and RoPE in Vision Transformers. To the best of our knowledge, this represents the first direct comparison of these positional encoding methods with those used in standard Vision Transformers.