Research map / Architectures and efficiency / Pruning and efficient tuning
Breaking the Scale Barrier: One-Shot Knowledge Transfer via Frequency Transform
Jianlu Shen, Fu Feng, Yucheng Xie, Jiaqi Lv, Xin Geng
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 |
|---|---|---|---|
| add MLP weight transfer (vs none) (ablation) | FRONT · DeiT-Ti L12 to L4, ImageNet-1K | top-1 accuracy (%): 34.7 -> 54.8 | Improved |
| transfer all modules instead of MLP only (ablation) | FRONT · DeiT-Ti L12 to L4, ImageNet-1K | top-1 accuracy (%): 54.8 -> 55.3 | Improved |
| transfer W_qkv and W_o in addition to nothing (attention only) (ablation) | FRONT · DeiT-Ti L12 to L4, ImageNet-1K | top-1 accuracy (%): 34.7 -> 50.1 | Improved |
| add self-attention weight transfer (W_qkv, W_o) (ablation) | FRONT · BERT-B to BERT-S, GLUE | average score: 53.31 -> 63.61 | Improved |
| add MLP weight transfer only (ablation) | FRONT · BERT-B to BERT-S, GLUE | average score: 53.31 -> 58.08 | Improved |
| transfer all modules instead of attention only (ablation) | FRONT · BERT-B to BERT-S, GLUE | average score: 63.61 -> 64.63 | Improved |
| add W_qkv only transfer (ablation) | FRONT · BERT-B to BERT-S, CoLA | MCC: 8.14 -> 4.63 | Hurt |
| add frequency-regularized refinement (FRONT+) | FRONT · DeiT-Ti 12 layers, ImageNet-1K 10 epochs | top-1 accuracy (%): 65.3 -> 67.5 | Improved |
| add refinement with gamma=1/4 at r=0.67 (ablation) | FRONT · DeiT-Ti L8, ImageNet-1K | top-1 accuracy (%): 64.79 -> 68.02 | Improved |
| increase frequency ratio r from 0.33 to 1.00 without refinement (ablation) | FRONT · DeiT-Ti L8, ImageNet-1K | top-1 accuracy (%): 60.63 -> 66.19 | Improved |
| increase decay gamma from 1/4 to 1 (ablation) | FRONT+ · DeiT-Ti L8, ImageNet-1K, r=0.67 | top-1 accuracy (%): 68.02 -> 66.60 | Hurt |
| use extreme regularization lambda (1e-5 or 5e-3) (ablation) | FRONT++ · DeiT-Ti-L10 target | top-1 accuracy | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- WAVE — WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models (CVPR 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.88)
- Fantastic Gains and Where to Find Them: On the Existence and Prospect of General Knowledge Transfer between Any Pretrained Model (ICLR 2024 · similarity 0.87)
- Transferring Knowledge From Large Foundation Models to Small Downstream Models (ICML 2024 · similarity 0.86)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.86)
- FreeKD: Knowledge Distillation via Semantic Frequency Prompt (CVPR 2024 · similarity 0.86)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.86)
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.
