Reproduce “Convergence Properties of Stochastic Hypergradients” 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
Riccardo Grazzi, Massimiliano Pontil, Saverio Salzo
The authors introduce stochastic implicit differentiation for estimating hypergradients by using minibatch-based stochastic solvers for both the lower-level model fit and sensitivity linear system. Their analysis bounds mean-square error using solvers’ convergence errors and sampled hyperparameter-derivative variance. In 784-parameter regularization experiment, stochastic estimates’ error falls while the batch curve changes little; this measures approximation error, not speed or accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.