Research map / Optimization / Optimizers
A Computationally Efficient Sparsified Online Newton Method
Fnu Devvrit, Sai Surya Duvvuri, Rohan Anil, Vineet Gupta, Cho-Jui Hsieh, Inderjit S Dhillon
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 |
|---|---|---|---|
| Use diagonal sparsity instead of a first-order optimizer (ablation) | SONew · MNIST Autoencoder | Train CE loss: Adam 53.591 -> diag-SONew 53.025 | Improved |
| Use tridiagonal sparsity instead of diagonal sparsity (ablation) | SONew · MNIST Autoencoder | Train CE loss: 53.025 -> 51.723 | Improved |
| Use banded sparsity with band size 4 instead of diagonal sparsity (ablation) | SONew · MNIST Autoencoder | Train CE loss: 53.025 -> 51.357 | Improved |
| Use tridiagonal sparsity instead of Adam | SONew · ImageNet ViT | iterations to target validation performance: 10% | Improved |
| Use tridiagonal sparsity instead of Adam | SONew · OGBG-molpcba GraphNetwork | iterations to target validation performance: 30% | Improved |
| Use tridiagonal sparsity instead of Adam at the same number of steps | SONew · ImageNet ViT | relative validation error: ~ 0.7% | Improved |
| Use tridiagonal sparsity instead of Adam at the same number of steps | SONew · OGBG-molpcba GraphNetwork | relative validation average precision: ~ 3.4% | Improved |
| Use tridiagonal sparsity instead of Shampoo | SONew · OGBG-molpcba GraphNetwork | average precision: 1.3% | Hurt |
| Use tridiagonal sparsity instead of AdaFactor | SONew · 1B-parameter language model | steps to same log perplexity: 26% | Improved |
| Use tridiagonal sparsity instead of AdaFactor with the same number of tokens | SONew · 1B-parameter language model | train loss: 1.7% | Improved |
Closest papers by meaning
- Error Feedback Can Accurately Compress Preconditioners (ICML 2024 · similarity 0.87)
- Exact, Tractable Gauss-Newton Optimization in Deep Reversible Architectures Reveal Poor Generalization (NeurIPS 2024 · similarity 0.86)
- Convergence Analysis of Newton's Method for Neural Networks in the Overparameterized Limit (NeurIPS 2026 · similarity 0.86)
- A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.86)
- Skip the Hessian, Keep the Rates: Globalized Semismooth Newton with Lazy Hessian Updates (NeurIPS 2026 · similarity 0.86)
- S-STE: Continuous Pruning Function for Efficient 2:4 Sparse Pre-training (NeurIPS 2024 · similarity 0.85)
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.
