Research map / Architectures and efficiency / CNN design
E2-Train: Training State-of-the-art CNNs with Over 80% Energy Savings
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 |
|---|---|---|---|
| randomly skip mini-batches with 0.5 probability (ablation) | CNN training · CIFAR-10, ResNet-110 | accuracy: 92.75% -> 93.05% | Helped |
| randomly skip mini-batches with 0.5 probability (ablation) | CNN training · CIFAR-100, ResNet-74 | accuracy: 71.11% -> 71.37% | Helped |
| use learned input-dependent layer selection instead of random layer dropping (ablation) | ResNet training · CIFAR-10 | top-1 accuracy: +0.86% | Helped |
| use PSG instead of 8-bit fixed-point training | ResNet-74 · CIFAR-10 | energy savings: 38.62% -> 63.28% | Helped |
| use PSG instead of SignSGD | ResNet-74 · CIFAR-10 | top-1 accuracy: 92.54% -> 92.59% | Helped |
| combine SMD, SLU, and PSG rather than SMD+SLU (ablation) | E²-Train · CIFAR-10, ResNet-74, at 80% energy savings | accuracy: 91.6% -> 92.1% | Helped |
| increase SLU skipping from 20% to 60% with β=0.05 (ablation) | E²-Train · CIFAR-10, ResNet-74 | accuracy: 92.12% -> 91.36% | Hurt |
| increase SLU skipping from 20% to 60% with β=0.1 (ablation) | E²-Train · CIFAR-10, ResNet-74 | accuracy: 92.15% -> 90.94% | Hurt |
| increase SLU skipping from 20% to 60% with β=0.05 (ablation) | E²-Train · CIFAR-10, ResNet-74 | energy savings: 84.64% -> 92.81% | Helped |
| apply E²-Train rather than standard SMB training | ResNet-110 · CIFAR-100 | top-1 accuracy: 71.60% -> 71.63% | Helped |
| apply E²-Train rather than standard SMB training | ResNet-110 · CIFAR-100 | top-5 accuracy: 91.50% -> 91.72% | Helped |
| apply E²-Train rather than standard SMB training | ResNet-110 · CIFAR-10 | top-1 accuracy: 93.57% -> 93.01% | Hurt |
Closest papers by meaning
- The Hidden Joules: Evaluating the Energy Consumption of Vision Backbones for Progress Towards More Efficient Model Inference (ICML 2025 · similarity 0.85)
- E4: Energy-Efficient DNN Inference for Edge Video Analytics via Early Exiting and DVFS (AAAI 2025 · similarity 0.85)
- Energy-Based Transformers are Scalable Learners and Thinkers (ICLR 2026 · similarity 0.84)
- An Empirical Study of Distributed Deep Learning Training on Edge (Student Abstract) (AAAI 2024 · similarity 0.84)
- Towards training digitally-tied analog blocks via hybrid gradient computation (NeurIPS 2024 · similarity 0.84)
- Stepping Forward on the Last Mile (NeurIPS 2024 · similarity 0.84)
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.