Research map / Architectures and efficiency / CNN design
The Master Key Filters Hypothesis: Deep Filters Are General
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Challenges the view that deep CNN filters become class-specific by freezing and transferring depthwise filters across datasets, architectures, and domains in DS-CNNs (ConvNeXt, HorNet, ResNet), finding they stay general.
Problem
Whether CNN filters in deeper layers become increasingly specialized (as Yosinski et al. concluded) or remain general, especially in depthwise separable CNNs.
Approach
Proposes the Master Key Filters Hypothesis: depthwise filters in DS-CNNs converge to general master filter sets regardless of dataset, task, or architecture, tested by freezing transferred depthwise filters (and pointwise filters as a contrast) across layers, datasets, and architectures.
Key results
- ConvNeXt Tiny keeps 86.9% accuracy with all filters transferred and frozen, matching baseline, and 86.2% with shuffled filters; the first 3 layers only also gives 86.9%.
- Transferring depthwise filters from ImageNet to ConvNeXt Femto on Oxford Pets yields +11.0 points with ConvNeXt Tiny source (56.0% for Femto source baseline).
- HorNet Tiny (trained on Food 101) depthwise filters transferred to ConvNeXt Femto on Pets reach 55.5%, a 3.1% gain over the selffer baseline.
- Transferring pointwise filters consistently lowered accuracy, even in selffer experiments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Transfer and freeze depthwise filters from man-made ImageNet classes to natural classes | ConvNeXt Tiny · ImageNet semantically divided (man-made to natural) | accuracy: 86.9% -> 86.9% | No clear effect |
| Shuffle transferred filters across layers (ablation) | ConvNeXt Tiny · ImageNet semantically divided (man-made to natural) | accuracy: 86.9% -> 86.2% | Hurt |
| Transfer only the first 3 layers' filters and repeat them throughout the remaining layers (ablation) | ConvNeXt Tiny · ImageNet semantically divided (man-made to natural) | accuracy: 86.9% -> 86.9% | No clear effect |
| Transfer depthwise filters from larger ImageNet-trained ConvNeXt variants | ConvNeXt Femto on Oxford Pets · Oxford Pets | accuracy: +11.0; +10.5 | Helped |
| Transfer depthwise filters across architectures from ImageNet-trained HorNet | ConvNeXt Femto on Oxford Pets · Oxford Pets | accuracy: +4.3; +7.3; +2.8 | Helped |
| Transfer and freeze depthwise filters from HorNet Tiny trained on Food 101 | ConvNeXt Femto · Oxford Pets | accuracy: 52.4% -> 55.5%; +3.1% | Helped |
| Transfer and freeze pointwise filters instead of depthwise filters (ablation) | ConvNeXt Femto · Cross-domain dataset transfers | accuracy | Hurt |
Surprising findings
- Transferring pointwise filters reduces accuracy even when filters come from the same dataset (selffer), suggesting optimization issues rather than specialization.
- Deeper, higher-accuracy ResNets retain more accuracy after full transfer, contrary to the expectation that specialization grows with depth.
Limitations
- Pointwise filter transfer results suggest convergence/optimization problems that are not fully explained.
- Cross-domain experiments mainly use the ConvNeXt Femto model with 300-epoch training and limited dataset selection.
- Claims about generality across all domains rest on a small set of visual datasets.
Experimental setup
Datasets: ImageNet Food 101 Sketch (DomainNet) CIFAR-10 STL-10 Oxford-IIIT Pets Oxford 102 Flowers
Benchmarks: ImageNet Food 101 Sketch CIFAR-10 STL-10 Oxford-IIIT Pets Oxford 102 Flowers
Metrics: top-1 accuracy accuracy retention accuracy change vs selffer baseline
Baselines: Yosinski et al. transfer results selffer baseline shuffle transferred first-3-layers transferred
Builds on: Yosinski et al. layerwise feature transfer ConvNeXt HorNet
Closest papers by meaning
- The Quest for Universal Master Key Filters in DS-CNNs (NeurIPS 2025 · similarity 0.92)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.89)
- Understanding deep learning requires rethinking generalization (ICLR 2017 · similarity 0.88)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.87)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.85)
- Normalize Filters! Classical Wisdom for Deep Vision (NeurIPS 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.