Research map / Label-efficient and robust learning / Recognition and long-tail
Explicit Inductive Bias for Transfer Learning with Convolutional Networks
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 |
|---|---|---|---|
| Replace weight decay toward zero with an L² penalty toward pretrained parameters | ResNet fine-tuning · MIT Indoors 67 | accuracy (%): 79.6±0.5 -> 84.2±0.3 | Helped |
| Replace weight decay toward zero with an L² penalty toward pretrained parameters | ResNet fine-tuning · Stanford Dogs 120 | accuracy (%): 81.4±0.2 -> 85.1±0.2 | Helped |
| Replace weight decay toward zero with an L² penalty toward pretrained parameters | ResNet fine-tuning · Caltech 256 -- 30 | accuracy (%): 81.5±0.2 -> 83.5±0.1 | Helped |
| Replace weight decay toward zero with an L² penalty toward pretrained parameters | ResNet fine-tuning · Caltech 256 -- 60 | accuracy (%): 85.3±0.2 -> 86.4±0.2 | Helped |
| Weight the pretrained-parameter distance with Fisher information instead of Euclidean distance (ablation) | L²-SP fine-tuning · Target classification tasks | accuracy (%): 84.2±0.3 vs 84.0±0.4; 85.1±0.2 vs 85.1±0.2; 83.5±0.1 vs 83.3±0.1; 86.4±0.2 vs 86.0±0.1 | No clear effect |
| Use L¹-SP or Group-Lasso-SP instead of L²-SP (ablation) | ResNet fine-tuning · Target classification tasks | accuracy (%) | Hurt |
| Use L¹-SP or Group-Lasso-SP with Places 365 initialization (ablation) | ResNet fine-tuning · Caltech 256 -- 60 | accuracy (%) | Hurt |
| Add the Fisher-weighted penalty to preserve source-task parameters (ablation) | L²-SP fine-tuning · Source tasks | accuracy drop (%): -5.3 -> -4.9; -4.7 -> -4.2; -4.2 -> -3.6; -3.6 -> -3.2 | Helped |
| Freeze initial network layers during L² fine-tuning (ablation) | L² fine-tuning · Stanford Dogs 120 and Caltech 256--30 | accuracy (%) | Helped |
| Use L²-SP rather than L² with partial layer freezing (ablation) | ResNet fine-tuning · Stanford Dogs 120 and Caltech 256--30 | accuracy (%) | Helped |
| Replace L² with L²-SP | ResNet-101 · Cityscapes validation | mean IoU: 68.1 -> 68.7 | Helped |
| Replace L² with L²-SP | DeepLab · Cityscapes validation | mean IoU: 72.0 -> 73.2 | Helped |
Closest papers by meaning
- Co-Tuning for Transfer Learning (NeurIPS 2020 · similarity 0.89)
- Inductive biases of multi-task learning and finetuning: multiple regimes of feature reuse (NeurIPS 2024 · similarity 0.88)
- Features are fate: a theory of transfer learning in high-dimensional regression (ICML 2025 · similarity 0.85)
- Pay Attention to Small Weights (NeurIPS 2025 · similarity 0.85)
- Training the Untrainable: Introducing Inductive Bias via Representational Alignment (NeurIPS 2025 · similarity 0.85)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · 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.