Research map / Architectures and efficiency / Attention and transformers
Masked Image Residual Learning for Scaling Deeper Vision Transformers
Guoxi Huang, Hongtao Fu, Adrian G. Bors
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 |
|---|---|---|---|
| Use masked image residual learning instead of standard masked image reconstruction | ViT-B-48 pre-training · ImageNet-1K | fine-tuning top-1 accuracy: 85.9 -> 86.2 | Improved |
| Use MIRL rather than MAE pre-training | ViT-B transfer to MS-COCO · MS-COCO, Mask R-CNN 1× | box AP: 48.4 -> 49.3 | Improved |
| Add feature-level loss to MIRL (ablation) | ViT-B · ImageNet-1K | fine-tuning accuracy: 83.5 -> 83.6 | Improved |
| Add VGG loss to MIRL (ablation) | ViT-B · ImageNet-1K | fine-tuning accuracy: 83.5 -> 83.8 | Improved |
| Add feature-level loss to MIRL (ablation) | ViT-B · ImageNet-1K | pre-training time: 1.0× -> 1.8× | Hurt |
| Add VGG loss to MIRL (ablation) | ViT-B · ImageNet-1K | pre-training time: 1.0× -> 2.5× | Hurt |
| Increase encoder segmentation from 2 to 4 segments (ablation) | MIRL ViT-B-24 · ImageNet-1K | fine-tuning accuracy: 84.3 -> 84.7 | Improved |
| Increase encoder segmentation from 4 to 6 segments (ablation) | MIRL ViT-B-24 · ImageNet-1K | fine-tuning accuracy: 84.7 -> 84.6 | Hurt |
| Add densely interconnected decoding (DID) (ablation) | MIRL ViT-B-24 · ImageNet-1K | fine-tuning accuracy: 84.5 -> 84.7 | Improved |
| Add densely interconnected decoding (DID) (ablation) | MIRL ViT-B-48 · ImageNet-1K | fine-tuning accuracy: 85.0 -> 85.3 | Improved |
| Replace residual-prediction shortcuts with independent decoders that each reconstruct masked content (ablation) | segmented ViT pre-training · ImageNet-1K | fine-tuning accuracy: 83.5 -> 83.2 | Hurt |
| Replace main/residual targets with coarse/fine image targets (ablation) | segmented ViT-B pre-training · ImageNet-1K | fine-tuning accuracy: 83.5 -> 83.1 | Hurt |
Closest papers by meaning
- Auto-scaling Vision Transformers without Training (ICLR 2022 · similarity 0.90)
- Peeling the Onion: Hierarchical Reduction of Data Redundancy for Efficient Vision Transformer Training (AAAI 2023 · similarity 0.90)
- DeiT III: Revenge of the ViT (ECCV 2022 · similarity 0.89)
- Anti-Oversmoothing in Deep Vision Transformers via the Fourier Domain Analysis: From Theory to Practice (ICLR 2022 · similarity 0.89)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.88)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · 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.
