Research map / Optimization / Optimizers
Reducing the variance in online optimization by transporting past gradients
Sébastien Arnold, Pierre-Antoine Manzagol, Reza Babanezhad Harikandeh, Ioannis Mitliagkas, Nicolas Le Roux
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add implicit gradient transport with anytime tail averaging to Heavy Ball | Heavy Ball on MobileNetV2 · CIFAR10 test | accuracy: 91.55 -> 90.37 | Hurt |
| add implicit gradient transport with anytime tail averaging to Heavy Ball | Heavy Ball on LeNet5 · MNIST test | accuracy: 99.08 -> 99.19 | Improved |
| add implicit gradient transport with anytime tail averaging to Heavy Ball | Heavy Ball on logistic regression · Linear-MNIST test | accuracy: 92.52 -> 92.50 | No clear effect |
| add implicit gradient transport with anytime tail averaging to Heavy Ball | Heavy Ball on IMDb · IMDb test | accuracy: 86.90 -> 87.26 | No clear effect |
| add IGT gradients to Adam | Adam on MAML · Mini-ImageNet test | accuracy: 63.70% -> 65.16% | Improved |
| add IGT gradients with anytime tail averaging to Heavy Ball | Heavy Ball on MAML · Mini-ImageNet test | accuracy: 63.08% -> 64.57% | Improved |
| remove anytime tail averaging from Heavyball-ITA (ablation) | Heavyball-ITA on MobileNetV2 · CIFAR10 test | accuracy: 90.37 -> 12.05 | Hurt |
| remove anytime tail averaging from Heavyball-ITA (ablation) | Heavyball-ITA on CIFAR10 · CIFAR10 training | training error | Hurt |
| use increasing momentum without gradient transport (ablation) | Momentum with anytime tail averaging · CIFAR10 | training loss | Hurt |
| increase the minibatch size (ablation) | IGT · CIFAR10 training | steps to target training loss | Mixed |
| use anytime tail averaging with IGT (ablation) | Heavyball-IGT on ConvNet settings · MNIST and CIFAR10 optimization | training error | Improved |
Closest papers by meaning
- Implicit Gradient Regularization (ICLR 2021 · similarity 0.87)
- Variance-Reduced Gradient Estimation via Noise-Reuse in Online Evolution Strategies (NeurIPS 2023 · similarity 0.87)
- The Implicit Regularization of Stochastic Gradient Flow for Least Squares (ICML 2020 · similarity 0.87)
- Variance-Reduced Stochastic Gradient Descent on Streaming Data (NeurIPS 2018 · similarity 0.87)
- Non-asymptotic Analysis of Biased Adaptive Stochastic Approximation (NeurIPS 2024 · similarity 0.86)
- MARS: Unleashing the Power of Variance Reduction for Training Large Models (ICML 2025 · similarity 0.86)
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.
