Vision Transformers (ViT)

Relative Position Bias

Bias added to attention scores that depends on the relative positions of patches, improving the model's ability to understand spatial relationships without absolute position encoding.

← Terug