Collapsing redundant layers: a unified framework for faster ViT inference without accuracy loss

burkov · x · 2026-09-24

Vision Transformers dominate computer vision but their size and compute demands create deployment bottlenecks. Prior work shows ViTs contain depthwise redundancy: groups of adjacent layers performing nearly identical transformations.

This article establishes a unified framework for layer redundancy and evaluates whether contiguous redundant layers can be collapsed into a single standard learned layer, cutting parameters and inference compute without degrading accuracy — addressing the fact that existing compression methods either fail to speed up inference or cause noticeable accuracy losses.

Original post →

More from Research

Research channel →