Reproduce “From Activation to Intervention: Interpreting Sparse Visual Features through Reconstruction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jianyang Gu, Jacob Beattie, Mridul Khurana, Julia Chae, Xi Xiao, Cheng Zhang, Anuj Karpatne, Wei-Lun Chao and 2 more
The method makes vision-model features easier to interpret by changing a sparse feature in the model and reconstructing the resulting patch representation as an image. This visual change complements examples where the feature activates. On NABirds with DINOv3, it raises feature-explanation accuracy from 73.0% to 74.8%; this result is specific to that dataset, encoder, and benchmark comparison.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.