Research map / Label-efficient and robust learning / Knowledge distillation
Online Ensemble Model Compression using Knowledge Distillation
Devesh Walawalkar, Zhiqiang Shen, Marios Savvides
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| train students with the online ensemble framework instead of individual baseline training | compressed students · CIFAR100 ResNet110, fifth student | accuracy: 42.26 -> 46.76 | Improved |
| train students with the ensemble framework | compressed students · CIFAR10 ResNet110, fifth student | accuracy: 92.27 -> 93.28 | Improved |
| train students with the ensemble framework | compressed students · ImageNet ResNet18, fifth student | top-1 accuracy: 55.23 -> 58.52 | Improved |
| train students with the ensemble framework | compressed students · CIFAR100 ResNet32, fourth student | accuracy: 61.85 -> 61.17 | Hurt |
| train students with the ensemble framework | compressed students · CIFAR10 and CIFAR100 | accuracy: about 1% better on average | Improved |
| train with ensemble framework on SVHN and ImageNet | fourth and fifth students · SVHN and ImageNet | accuracy: around 3% better on average | Improved |
| use ensemble teacher versus KD-ONE | ensemble teacher · CIFAR100 ResNet-32 | error rate: 26.61 -> 26.09 | Improved |
| use ensemble teacher versus KD-ONE | ensemble teacher · SVHN ResNet-32 | error rate: 1.83 -> 1.97 | Hurt |
| raise cross-entropy ratio alpha from 0.3 to 0.7 (ablation) | ensemble loss · CIFAR100 ResNet110 | weighted average accuracy: 161.71 -> 179.31 | Improved |
| raise alpha from 0.7 to 0.9 (ablation) | ensemble loss · CIFAR100 ResNet110 | weighted average accuracy: 179.31 -> 171.18 | Hurt |
| set beta=gamma=0.15 instead of beta=0.2, gamma=0.1 (ablation) | ensemble loss · CIFAR100 ResNet110 | weighted average accuracy: 164.37 -> 177.65 | Improved |
| set temperature T=2 instead of T=1 (ablation) | distillation · CIFAR100 ResNet110 | mean student accuracy: 61.78 -> 63.32 | Improved |
Closest papers by meaning
- Progressive Ensemble Distillation: Building Ensembles for Efficient Inference (NeurIPS 2023 · similarity 0.90)
- Knowledge distillation via softmax regression representation learning (ICLR 2021 · similarity 0.90)
- How to Trade Off the Quantity and Capacity of Teacher Ensemble: Learning Categorical Distribution to Stochastically Employ a Teacher for Distillation (AAAI 2024 · similarity 0.89)
- Knowledge Distillation as Semiparametric Inference (ICLR 2021 · similarity 0.88)
- Knowledge Distillation: A Survey (IJCV 2021 · similarity 0.88)
- Ensemble Distribution Distillation via Flow Matching (ICML 2025 · similarity 0.87)
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.
