Research map / Optimization / Distributed and compressed
ATOMO: Communication-efficient Learning via Atomic Sparsification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Spectral-Atomo instead of QSGD | Distributed training · VGG-11-BN on CIFAR-10 | speedup over vanilla SGD: 3.96× vs 1.68× | Helped |
| Use SVD sparsification instead of QSGD | Distributed neural-network training · VGG and ResNet-18 on SVHN and CIFAR-10 | training time: up to 2× faster | Helped |
| Use SVD sparsification instead of TernGrad | Distributed neural-network training · VGG and ResNet-18 on SVHN and CIFAR-10 | training time: up to 3× faster | Helped |
| Use a moderate SVD sparsity budget (3 or 4) rather than extreme sparsification (ablation) | Spectral-Atomo · CIFAR-10 and SVHN convergence experiments | convergence speed | Helped |
| Increase QSGD quantization from 4 bits to 8 bits (ablation) | QSGD · CIFAR-10 and SVHN convergence experiments | convergence speed | Helped |
| Use extreme sparsification with 1-bit QSGD (ablation) | QSGD · CIFAR-10 and SVHN convergence experiments | convergence speed | Hurt |
| Increase sparsification or quantization to the most aggressive setting (ablation) | Spectral-Atomo and QSGD · CIFAR-10 and SVHN convergence experiments | convergence speed | Hurt |
Closest papers by meaning
- The Convergence of Sparsified Gradient Methods (NeurIPS 2018 · similarity 0.90)
- Gradient Sparsification for Communication-Efficient Distributed Optimization (NeurIPS 2018 · similarity 0.88)
- Smoothness Matrices Beat Smoothness Constants: Better Communication Compression Techniques for Distributed Optimization (NeurIPS 2021 · similarity 0.88)
- Rethinking gradient sparsification as total error minimization (NeurIPS 2021 · similarity 0.87)
- MAST: model-agnostic sparsified training (ICLR 2025 · similarity 0.87)
- JointSQ: Joint Sparsification-Quantization for Distributed Learning (CVPR 2024 · 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.