Reproduce “Pulling Back the Curtain on Deep Networks” 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

Maciej Satkiewicz, Roberto Corizzo, Marcin Pietroń

The authors introduce Semantic Pullbacks, a method that carries target signals backward through input-conditioned linear operations without differentiating through their input-dependent changes. PullbackAscent repeatedly uses this signal to modify images. On ResNet50, mean Infidelity falls to 5.384 from 29.480 for GradientAscent. Across 60 matched comparisons, 46 improve, seven deteriorate, and seven are inconclusive under the paper’s image-perturbation protocol.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.