Research map / Optimization / Distributed and compressed
Local and asynchronous SGD
Analyze asynchronous, local, and minibatch SGD for communication-efficient distributed training.
84 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Distributed and compressed, keyword cluster “decentralized · asynchronous · byzantine”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
13 papers in 2023–24 and 21 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 18.3% to 28.0%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Fast Asynchronous Parallel Stochastic Gradient Descent: A Lock-Free Approach with Convergence Guarantee (AAAI 2016 · 70 citations)
- 2017 · Can Decentralized Algorithms Outperform Centralized Algorithms? A Case Study for Decentralized Parallel Stochastic Gradient Descent (NeurIPS 2017 · 422 citations)
- 2018 · Large scale distributed neural network training through online distillation (ICLR 2018 · 151 citations)
- 2019 · Parallel Restarted SGD with Faster Convergence and Less Communication: Demystifying Why Model Averaging Works for Deep Learning (AAAI 2019 · 506 citations)
- 2020 · On the Convergence of FedAvg on Non-IID Data (ICLR 2020 · 979 citations)
- 2021 · Asynchronous Decentralized SGD with Quantized and Local Updates (NeurIPS 2021 · 22 citations)
- 2022 · Demystifying Why Local Aggregation Helps: Convergence Analysis of Hierarchical SGD (AAAI 2022 · 68 citations)
- 2023 · Stability-Based Generalization Analysis of the Asynchronous Decentralized SGD (AAAI 2023 · 13 citations)
- 2024 · Towards Stability and Generalization Bounds in Decentralized Minibatch Stochastic Gradient Descent (AAAI 2024 · 4 citations)
- 2025 · AdaGK-SGD: Adaptive Global Knowledge Guided Distributed Stochastic Gradient Descent (AAAI 2025 · 1 citation)
- 2026 · Clipping Makes Distributed and Federated Asynchronous SGD Robust to Stragglers (ICML 2026 · 0 citations)
Earliest papers here
- Fast Asynchronous Parallel Stochastic Gradient Descent: A Lock-Free Approach with Convergence Guarantee (AAAI 2016 · 70 citations)
- Asynchronous Distributed Semi-Stochastic Gradient Optimization (AAAI 2016 · 14 citations)
- Stochastic Gradient MCMC with Stale Gradients (NeurIPS 2016 · 1 citation)
What has been tried
66 papers here report 501 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 11%, hurt 16%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add the MLGK module | CIFAR-10, ResNet · TOP-1 accuracy (%) | ↑0.275 | Helped | AdaGK-SGD: Adaptive Global Knowledge Guided Distributed Stochastic Gradient Descent (AAAI 2025 · 1 citation) |
| use ℓ=2 or ℓ=3 instead of ℓ=1 | Approximate GC, G1 and G2 assignments · MSE | Helped | Leveraging partial stragglers within gradient coding (NeurIPS 2024 · 1 citation) | |
| Replace the server-side sign operation with an unbiased stochastic sign mapping | Heterogeneous majority vote · convergence rate (l2-norm) | \(\mathcal{O}(d^{3/8}T^{-1/8})\) -> \(\mathcal{O}(d^{1/4}T^{-1/4})\) | Helped | Efficient Sign-Based Optimization: Accelerating Convergence via Variance Reduction (NeurIPS 2024 · 1 citation) |
| Add global and local momentum | Pythia-70M · tokens-to-loss | 38.1B -> 19.3B | Helped | HALoS: Hierarchical Asynchronous Local SGD over Slow Networks for Geo-Distributed Large Language Model Training (ICML 2025 · 0 citations) |
| Add gradient accumulation (AccumAdam) instead of DAdam | WMT14 En-De test · METEOR | 0.5440 ± 0.0013 -> 0.5481 ± 0.0027 | Helped | From Promise to Practice: Realizing High-performance Decentralized Training (ICLR 2025 · 0 citations) |
| Use the small learning rate η = 1/(KH) instead of the two-stage learning rate | Synthetic dataset; heterogeneous MNIST · training loss convergence | Hurt | Local Steps Speed Up Local GD for Heterogeneous Distributed Logistic Regression (ICLR 2025 · 0 citations) | |
| Add OrMo's compensated parameter update rule rather than using a vanilla SGD step | CIFAR10, 16 workers (hom.) · test accuracy | 90.32 ± 0.45 -> 90.95 ± 0.27 | Helped | Ordered Momentum for Asynchronous SGD (NeurIPS 2024 · 0 citations) |
| remove the full-gradient computation at initialization (Option II) | Finite-sum nonconvex optimization · stochastic gradient complexity | n + ζ∆√n/ε2 -> (ζ∆+σ2c)√n/ε2 | Mixed | SILVER: Single-loop variance reduction and application to federated learning (ICML 2024) |
What reviewers will expect
Datasets and baselines named in the 15 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 6 of 15 papers
- CIFAR-100 · 3 of 15 papers
- MNIST · 2 of 15 papers
- synthetic two-client dataset · 2 of 15 papers
Baselines
- Mini-batch SGD · 4 of 15 papers
- DiLoCo · 2 of 15 papers
- Asynchronous SGD · 2 of 15 papers
- Delay-Adaptive ASGD · 2 of 15 papers
- Rennala SGD · 2 of 15 papers
- Malenia SGD · 2 of 15 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 34 papers; from OpenAlex and the papers' affiliations).
- Google · 3 papers
- National University of Defense Technology · 2 papers
- Princeton University · 2 papers
- New York University · 2 papers
- Meta · 2 papers
- King Abdullah University of Science and Technology · 2 papers
- KAUST · 2 papers
- Nanjing University · 2 papers
- New Jersey Institute of Technology · 2 papers
- CISPA Helmholtz Center · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Decentralized optimization algorithms (94 papers)
- Byzantine-robust learning (32 papers)
- Compressed distributed optimization (30 papers)
Elsewhere in the topic
- Gradient quantization and sparsification (57 papers)
- Communication complexity theory (35 papers)
- Error feedback theory (17 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).