AI Paper Atlas

Research map / Optimization / Optimizers

Adaptive Second Order Coresets for Data-efficient Machine Learning

Omead Pooladzandi, David Davini, Baharan Mirzasoleiman

ICML 2022 · Optimization › Optimizers · approach: Natural gradient and Fisher approximations · 5 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
select weighted coresets using Hessian-preconditioned gradients instead of gradient-only selectionResNet20 with SGD+Momentum · CIFAR10test accuracy: 43.6% ± 1.6 (75%) -> 55.4% ± 1.1 (74%)Improved
select weighted coresets using Hessian-preconditioned gradients instead of random samplingResNet20 with AdaHessian · CIFAR10test accuracy: 59.1% ± 2.8 (87%) -> 60.2% ± 0.5 (73%)Improved
use AdaCore-selected subsets rather than CRAIG subsetsResNet18 · Imbalanced CIFAR10, S=1%, R=20test accuracy: 48.6% ± 0.8 -> 57.3% ± 0.5Improved
use AdaCore-selected subsets rather than random subsetsResNet18 · Imbalanced CIFAR10, S=1%, R=20test accuracy: 54.7% ± 0.3 -> 57.3% ± 0.5Improved
use AdaCore-selected subsets rather than CRAIG subsetsResNet50 with SGD+Momentum · BDD100ktest accuracy: 72.3% -> 74.3%Improved
remove exponential averaging of gradient and Hessian (ablation)AdaCore with ResNet20 and AdaHessian · CIFAR10test accuracy: 60.2% ± 0.5 -> 58.4% ± 0.2Hurt
remove exponential averaging of gradient and Hessian (ablation)AdaCore with ResNet20 and SGD+Momentum · CIFAR10test accuracy: 55.4% ± 1.1 -> 51.5% ± 1.1Hurt
use exponential averaging of gradients only, without Hessian averaging (ablation)AdaCore with ResNet20 and AdaHessian · CIFAR10test accuracy: 60.2% ± 0.5 -> 59.8% ± 0.5Hurt
use exponential averaging of gradients only, without Hessian averaging (ablation)AdaCore with ResNet20 and SGD+Momentum · CIFAR10test accuracy: 55.4% ± 1.1 -> 53.2% ± 1.1Hurt
use exponential averaging of Hessian only, without gradient averaging (ablation)AdaCore with ResNet20 and AdaHessian · CIFAR10test accuracy: 60.2% ± 0.5 -> 60.2% ± 0.5No clear effect
use exponential averaging of Hessian only, without gradient averaging (ablation)AdaCore with ResNet20 and SGD+Momentum · CIFAR10test accuracy: 55.4% ± 1.1 -> 54.4% ± 1.1Hurt
increase the Hessian-estimation batch size to 512 (ablation)AdaCore with ResNet20 and AdaHessian · CIFAR10test accuracy: 60.2% ± 0.5 -> 57.2% ± 0.5Hurt

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