Research map / Label-efficient and robust learning / Knowledge distillation
Masked Autoencoders Enable Efficient Knowledge Distillers
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 |
|---|---|---|---|
| Distill from an MAE-pretrained teacher during masked pre-training rather than distilling a fine-tuned teacher | ViT-B · ImageNet-1k | top-1 accuracy (%): 82.8% vs 84.0% | Helped |
| Add masked pre-training with feature distillation instead of 100-epoch MAE alone | ViT-B · ImageNet-1k | top-1 accuracy (%): 81.6% to 84.0% | Helped |
| Use soft or hard logit distillation from the fine-tuned teacher | DeiT-B · ImageNet-1k | top-1 accuracy (%): 76.8 -> 77.5; 76.8 -> 78.3 | Hurt |
| Increase masking from 75% to 90% (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 84.0 vs 84.0 | No clear effect |
| Increase masking to 98% (FOUR visible patches) (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 84.0 -> 82.4 | Hurt |
| Align features at three-quarters depth rather than at the bottom (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 82.6 -> 84.0 | Helped |
| Align multiple layers instead of only the three-quarters-depth layer (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 84.0% vs 84.2% | No clear effect |
| Match layers at the same relative depth in teacher and student (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 83.3 -> 84.0; 82.1 -> 84.0 | Helped |
| Use a 2-layer MLP instead of a linear projection head (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 83.6 -> 84.0 | Helped |
| Increase decoder depth from 2 to 8 (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 83.7 -> 84.0 | Helped |
| Use L1 rather than L2 distance for feature alignment (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 83.3 -> 84.0; +0.7% | Helped |
| Vary the distillation-loss weight alpha from 1 to 0.5, 2, or 4 (ablation) | DMAE-B · ImageNet-1k | top-1 accuracy (%): 83.9; 84.0; 84.0 | No clear effect |
Closest papers by meaning
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.91)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.90)
- Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners (ICLR 2024 · similarity 0.89)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.88)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.88)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.88)
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.