Reproduce “Zero Grads: Learning Local Surrogate Losses for Non-Differentiable Graphics” 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
Michael Fischer, Tobias Ritschel
ZeroGrads optimizes settings in forward-only graphics and simulation pipelines by learning a local, differentiable loss surrogate from sampled outputs. It smooths nearby losses and focuses training on the current neighborhood; its gradient guides settings without differentiating through the pipeline. On a 35,152-weight MNIST MLP task, it converges under an equal function-evaluation budget where several baselines do not; this is not an elapsed-time comparison.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.