Research map / Architectures and efficiency / Pruning and efficient tuning
Efficient Distributed Inference of Deep Neural Networks via Restructuring and Pruning
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 |
|---|---|---|---|
| Restructure and prune the trained model to reduce cross-worker connections | Sensor network, Setup 1 · Sensor network, Setup 1 | accuracy: 94.5% -> 93.5% | Hurt |
| Fine-tune the restructured and pruned model | RePurpose · Sensor network, Setup 1 | accuracy: 93.5% -> 94.4% | Helped |
| Increase the cross-communication penalty from η2 = 0.01 to η2 = 0.1 (ablation) | RePurpose · Sensor network, Setup 1 | cross-communication: 1.7%, 1.5% and 1.6% -> 0.7%, 0.1% and 0.3% | Helped |
| Use 99% rather than 50% sparsity (ablation) | RePurpose · Datacenter, N = 8K | computation speedup: 1.7× -> 10.47× | Helped |
| Use 99% rather than 50% sparsity (ablation) | RePurpose · Datacenter, N = 8K | communication speedup: 1.2× -> 1.75× | Helped |
| Use 99% rather than 50% sparsity (ablation) | RePurpose · Edge, N = 8K | computation speedup: 1.7× -> 11.01× | Helped |
| Use 99% rather than 50% sparsity (ablation) | RePurpose · Edge, N = 8K | communication speedup: 1.2× -> 3.04× | Helped |
| Increase model size from N = 1K to N = 32K (ablation) | RePurpose · Edge platform | total time speedup: 1.55× -> 3.8× | Helped |
| Fine-tune after restructuring and pruning rather than leave the model untuned (ablation) | RePurpose · MNIST Setup 2 and CIFAR10 Setup 3 | accuracy | Helped |
| Restructure neurons before pruning rather than directly sparsify without rearrangement | RePurpose · MNIST Setup 2 and CIFAR10 Setup 3 | accuracy | Helped |
| Use model distillation instead of RePurpose | Distributed inference models · MNIST Setup 2 and CIFAR10 Setup 3 | accuracy | Hurt |
| Use direct parallelization without restructuring or pruning | Distributed inference model · Distributed inference | cross-worker communication: more than 100,000 values | Hurt |
Closest papers by meaning
- SplitNet: Learning to Semantically Split Deep Networks for Parameter Reduction and Model Parallelization (ICML 2017 · similarity 0.87)
- Mesh-TensorFlow: Deep Learning for Supercomputers (NeurIPS 2018 · similarity 0.87)
- Practical Performance Guarantees for Pipelined DNN Inference (ICML 2024 · similarity 0.87)
- Exploring Hidden Dimensions in Accelerating Convolutional Neural Networks (ICML 2018 · similarity 0.87)
- Towards a Better Theoretical Understanding of Independent Subnetwork Training (ICML 2024 · similarity 0.87)
- Auto-reconfiguration for Latency Minimization in CPU-based DNN Serving (ICML 2025 · 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.