Research map / Optimization / Distributed and compressed
TernGrad: Ternary Gradients to Reduce Communication in Distributed Deep Learning
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 ternary gradients instead of floating gradients | AlexNet · ImageNet, batch 1024 | top-1 accuracy: 56.62% -> 57.54% | Helped |
| Use ternary gradients instead of floating gradients | AlexNet · ImageNet, batch 512 | top-1 accuracy: 57.32% -> 57.28% | No clear effect |
| Add gradient clipping to TernGrad (ablation) | AlexNet · ImageNet, batch 256 | top-1 accuracy: 54.63% -> 57.61% | Helped |
| Use ternary gradients instead of floating gradients | CifarNet · CIFAR-10, 2 workers | accuracy: 86.56% -> 85.64% (-0.92%) | Hurt |
| Use ternary gradients instead of floating gradients | CifarNet · CIFAR-10, 16 workers | accuracy: 83.19% -> 82.80% (-0.39%) | Hurt |
| Use ternary gradients instead of floating gradients | GoogLeNet · ImageNet, batch 128 | top-5 accuracy: 88.30% -> 86.77% | Hurt |
| Use ternary gradients instead of floating gradients | GoogLeNet · ImageNet, batch 256 | top-5 accuracy: 87.82% -> 85.96% | Hurt |
| Use ternary gradients instead of floating gradients | GoogLeNet · ImageNet, batch 512 | top-5 accuracy: 89.00% -> 86.47% | Hurt |
| Use ternary gradients instead of floating gradients | AlexNet · ImageNet, 8 GPUs | training speedup: 3.04 × | Helped |
| Use ternary gradients instead of floating gradients | VggNet-A · 128 nodes, InfiniBand and NVLink cluster | training speed: double | Helped |
Closest papers by meaning
- A Communication-Efficient Distributed Gradient Clipping Algorithm for Training Deep Neural Networks (NeurIPS 2022 · similarity 0.87)
- $\textbf{A}^2\textbf{CiD}^2$: Accelerating Asynchronous Communication in Decentralized Deep Learning (NeurIPS 2023 · similarity 0.87)
- A Quadratic Synchronization Rule for Distributed Deep Learning (ICLR 2024 · similarity 0.86)
- Communication-Efficient Distributed Blockwise Momentum SGD with Error-Feedback (NeurIPS 2019 · similarity 0.86)
- Communication Efficient SGD via Gradient Sampling With Bayes Prior (CVPR 2021 · similarity 0.86)
- A Linear Speedup Analysis of Distributed Deep Learning with Sparse and Quantized Communication (NeurIPS 2018 · 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.