Reproduce “On the Iteration Complexity of Hypergradient Computation” 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, Luca Franceschi, Massimiliano Pontil, Saverio Salzo
The authors compare hypergradient estimates for bilevel optimization, where an outer objective depends on a costly lower-level solution. Iterative differentiation traces training updates; approximate implicit differentiation uses the fixed-point relationship. Under contraction assumptions, its conjugate-gradient variant often estimates gradients better and avoids trajectory storage. On 20 Newsgroups at t=50, it reports validation objective 0.946 and test accuracy 63.5%, versus 1.058 and 61.6% for iterative differentiation; this is not a general downstream or runtime guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.