Research map / Architectures and efficiency / Quantization
GradiVeQ: Vector Quantization for Bandwidth-Efficient Gradient Aggregation in Distributed CNN Training
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 GradiVeQ compressed-domain aggregation instead of uncompressed RAR | ResNet-32 · CIFAR-100, CPU cluster | convergence time: 135,000 -> 76,000 seconds | Helped |
| use GradiVeQ compressed-domain aggregation instead of 4-bit-QSGD + RAR | ResNet-32 · CIFAR-100, CPU cluster | convergence time: 90,000 -> 76,000 seconds | Helped |
| use GradiVeQ compressed-domain aggregation instead of uncompressed RAR | ResNet-32 · CIFAR-100, GPU cluster | convergence time: 75,000 -> 24,000 seconds | Helped |
| use GradiVeQ compressed-domain aggregation instead of 4-bit-QSGD + RAR | ResNet-32 · CIFAR-100, GPU cluster | convergence time: 30,000 -> 24,000 seconds | Helped |
| use gradient compression rather than uncompressed aggregation | 4-bit-QSGD + RAR · CIFAR-100, ResNet-32 | top-1 accuracy: 0.676 -> 0.667 | Hurt |
| use GradiVeQ rather than uncompressed aggregation | ResNet-32 · CIFAR-100 | top-1 accuracy: 0.676 -> 0.666 | Hurt |
| replace uncompressed RAR with GradiVeQ compression and parallelized aggregation | ResNet-32 · CIFAR-100, CPU cluster | gradient aggregation time per 500 iterations: 850 -> 162 seconds; 5.25x | Helped |
| use one PCA compressor per layer (s=∞) (ablation) | GradiVeQ · CIFAR-100, ResNet-32 | accuracy degradation versus uncompressed benchmark: less than 1% | No clear effect |
Closest papers by meaning
- JointSQ: Joint Sparsification-Quantization for Distributed Learning (CVPR 2024 · similarity 0.88)
- QSGD: Communication-Efficient SGD via Gradient Quantization and Encoding (NeurIPS 2017 · similarity 0.87)
- Communication-Efficient Distributed Learning via Lazily Aggregated Quantized Gradients (NeurIPS 2019 · similarity 0.87)
- A Linear Speedup Analysis of Distributed Deep Learning with Sparse and Quantized Communication (NeurIPS 2018 · similarity 0.87)
- Double Quantization for Communication-Efficient Distributed Optimization (NeurIPS 2019 · similarity 0.87)
- A Quadratic Synchronization Rule for Distributed Deep Learning (ICLR 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.