A Geometric Perspective on Robustness in Vision Transformers [R]

About this article

Hi everyone! I'm sharing a paper I've been working on that investigates how different positional encoding schemes (learned absolute, sinusoidal, and rotary) shape the internal representations of Vision Transformers, and how these representations relate to robustness under distributional shift. Paper PDF: https://github.com/mahmoud-mannes/neurips-geometry-paper/blob/main/paper/main.pdf Abstract: Positional embeddings (PEs) in Vision Transformers (ViTs) are known to impact performance and robus...