PCAFormer

Vision Transformer efficiency project using PCA-based token compression inside attention layers.

PCAFormer explores how to reduce transformer attention cost without losing core visual understanding.

Core ideas:

  • Implemented PCA-guided token compression within transformer blocks to shrink attention compute.
  • Benchmarked compressed-token variants against baseline ViT settings to analyze accuracy and runtime tradeoffs.
  • Used ablation-style experiments to understand where compression is most effective in the architecture.

This project focuses on practical model efficiency for large-scale vision workloads.