Research map / Label-efficient and robust learning / Knowledge distillation
Improved Feature Distillation via Projector Ensemble
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 an ensemble of projectors instead of one projector (ablation) | Feature distillation · CIFAR-100, VGG13→VGG8 | Top-1 accuracy: 73.84 -> 74.35 | Helped |
| Use an ensemble of projectors instead of one projector (ablation) | Feature distillation · CIFAR-100, ResNet32x4→ResNet8x4 | Top-1 accuracy: 75.14 -> 76.08 | Helped |
| Add a projector to the student (ablation) | Feature distillation without a projector · CIFAR-100, VGG13→VGG8 | Top-1 accuracy: 70.74 -> 73.76 | Helped |
| Add a projector to the student (ablation) | Feature distillation without a projector · CIFAR-100, ResNet32x4→ResNet8x4 | Top-1 accuracy: 72.93 -> 73.66 | Helped |
| Increase projector depth from one layer to a 2–4-layer MLP (ablation) | Feature distillation · CIFAR-100, both reported teacher-student pairs | Top-1 accuracy: 73.76 -> 73.31; 73.66 -> 75.12 | Hurt |
| Widen the hidden layers of a 2-layer MLP projector (ablation) | Feature distillation · CIFAR-100, both reported teacher-student pairs | Top-1 accuracy: 73.31 -> 73.31 -> 73.37; 75.12 -> 75.00 -> 75.13 | No clear effect |
| Replace ReLU with GELU in the projector (ablation) | Projector ensemble · CIFAR-100, VGG13→VGG8 | Top-1 accuracy: 74.35±0.12 -> 74.39±0.18 | Helped |
| Replace ReLU with GELU in the projector (ablation) | Projector ensemble · CIFAR-100, ResNet32x4→ResNet8x4 | Top-1 accuracy: 76.08±0.33 -> 76.32±0.27 | Helped |
| Remove the projector activation (ablation) | Projector ensemble · CIFAR-100, VGG13→VGG8 | Top-1 accuracy: 74.35±0.12 -> 73.46±0.42 | Hurt |
| Remove the projector activation (ablation) | Projector ensemble · CIFAR-100, ResNet32x4→ResNet8x4 | Top-1 accuracy: 76.08±0.33 -> 75.04±0.37 | Hurt |
Closest papers by meaning
- Understanding the Role of the Projector in Knowledge Distillation (AAAI 2024 · similarity 0.91)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.86)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.86)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.86)
- How to Trade Off the Quantity and Capacity of Teacher Ensemble: Learning Categorical Distribution to Stochastically Employ a Teacher for Distillation (AAAI 2024 · similarity 0.85)
- Ensemble Distribution Distillation via Flow Matching (ICML 2025 · 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.