AI Paper Atlas

Research map / Optimization / Optimizers

On the Convergence Theory for Hessian-Free Bilevel Algorithms

Daouda Sow, Kaiyi Ji, Yingbin Liang

NeurIPS 2022 · Optimization › Optimizers · approach: Hypergradient bilevel optimization · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
PZOBO instead of ProtoNetPZOBO · miniImageNet ResNet-12hours to 67% accuracy: 3.5 -> 2.0Improved
PZOBO instead of ProtoNetPZOBO · miniImageNet ResNet-12hours to 69% accuracy: 4.4 -> 2.8Improved
PZOBO instead of MAMLPZOBO · miniImageNet ResNet-12hours to 67% accuracy: 9.7 -> 2.0Improved
PZOBO instead of ANILPZOBO · miniImageNet ResNet-12hours to 67% accuracy: 6.3 -> 2.0Improved
PZOBO instead of MetaOptNetPZOBO · miniImageNet ResNet-12hours to 69% accuracy: 20+ -> 2.8Improved
PZOBO-S instead of other bilevel methodsPZOBO-S · MNIST deep HRaccuracy: 0.93 -> 0.98Improved
estimate only the response Jacobian (PZOBO) instead of full hypergradient (HOZOG)PZOBO · shallow HR two-layer netconvergenceImproved
vary number N of inner GD steps (ablation)PZOBO · shallow HRconvergenceNo clear effect
vary number N of inner SGD steps (ablation)PZOBO-S · MNIST deep HRconvergenceNo clear effect
PZOBO instead of HOZOGPZOBO · 20 Newsgroup hyperparameter optimizationconvergenceImproved
increase network size from CNN4 to ResNet-12PZOBO · miniImageNetadvantage over MAML and ANILImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy