Research map / Architectures and efficiency / CNN design
The Quest for Universal Master Key Filters in DS-CNNs
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.
Shows that depthwise filters in DS-CNNs (ConvNeXtV2, Hornet) can be approximated by linear shifts of a single set of 8 universal filters found by autoencoder-based candidate generation and greedy search, and trains ConvNeXtV2 and Hornet from scratch with those 8 frozen filters plus biases.
Problem
Whether DS-CNN depthwise filters, which number in the thousands, are really distinct or converge to a small universal set of master key filters, and whether networks can be trained with only those few frozen filters.
Approach
Compress depthwise filters from many trained models with an autoencoder, sample candidate filters from its 1D code, approximate each trained filter by a linear shift (a*f+b) of a candidate, then greedily prune candidates to 8 essential filters that are then used as frozen depthwise kernels during training.
Key results
- ConvNeXt V2 Tiny with 8 frozen filters plus biases reaches 82.7% ImageNet top-1 vs 82.5% for the original model, and Hornet Tiny reaches 81.8% vs 82.3%.
- Replacing all depthwise filters of ConvNeXt V2 Huge (49,632 filters) with linear shifts of the 8 greedy-selected filters gives 82.8% vs 86.3% original, without fine-tuning.
- On Oxford Pets with ConvNeXt Femto, the 8 frozen filters reach 66.5% vs 36.3% for normal training and 56.0% for ImageNet transfer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with 8 fixed filters and learnable biases instead of fully trainable filters | ConvNeXtv2 · ImageNet-1k | Top-1 accuracy: 82.5% -> 82.7% | No clear effect |
| Train with 8 fixed filters and learnable biases instead of fully trainable filters | HorNet-Tiny · ImageNet-1k | Top-1 accuracy: 82.3% -> 81.8% | Hurt |
| Train from scratch with 8 frozen filters instead of normal training or ImageNet-transferred filters | ConvNeXt-Femto · CIFAR-10 | accuracy: 96.9 -> 96.3; 97.1 -> 96.3 | Hurt |
| Train from scratch with 8 frozen filters instead of normal training or ImageNet-transferred filters | ConvNeXt-Femto · STL-10 | accuracy: 80.4 -> 83.1; 83.2 -> 83.1 | Mixed |
| Train from scratch with 8 frozen filters instead of normal training or ImageNet-transferred filters | ConvNeXt variants · Oxford Flowers | accuracy: 63.3 -> 77.8; 72.2 -> 77.8 | Helped |
| Train from scratch with 8 frozen filters instead of normal training or ImageNet-transferred filters | ConvNeXt variants · Oxford Pets | accuracy: 38.4 -> 66.5; 58.5 -> 66.5 | Helped |
| Replace trained depthwise filters with linear-shift approximations from 50 candidate filters (ablation) | ConvNeXtv2 and HorNet models · ImageNet-1k | accuracy: 83.0% -> 75.4% (ConvNeXtv2 Tiny); 82.3% -> 79.4% (HorNet Tiny) | Hurt |
| Replace trained depthwise filters with linear-shift approximations from 25 candidate filters (ablation) | ConvNeXtv2 and HorNet models · ImageNet-1k | accuracy: 83.0% -> 66.9% (ConvNeXtv2 Tiny); 82.3% -> 78.3% (HorNet Tiny) | Hurt |
| Replace trained depthwise filters with linear-shift approximations from 10 candidate filters (ablation) | ConvNeXtv2 and HorNet models · ImageNet-1k | accuracy: 83.0% -> 1.0% (ConvNeXtv2 Tiny); 82.3% -> 66.3% (HorNet Tiny) | Hurt |
| Use 8 filters selected by greedy search instead of 8 random filters (ablation) | ConvNeXtv2 and HorNet models · ImageNet-1k | accuracy: 0.10% -> 76.7% (ConvNeXtv2 Tiny); 0.96% -> 76.0% (HorNet Tiny) | Helped |
| Replace trained filters with 8 greedy-search filter transformations (ablation) | ConvNeXtv2 and HorNet models · ImageNet-1k | accuracy: 83.0% -> 76.7% (ConvNeXtv2 Tiny); 82.3% -> 76.0% (HorNet Tiny) | Mixed |
Surprising findings
- Eight random filters in place of the learned ones collapse accuracy to about 0.1%, while the greedy-selected 8 retain most accuracy.
- The 8 filters found on ConvNeXt V2 Tiny transfer without modification to the Hornet architecture.
Limitations
- Greedy search and candidate selection were done on ConvNeXt V2 Tiny, and the 8 filters were not optimized further.
- No error bars or repeated runs due to ImageNet training cost.
- ImageNet results use the original training hyperparameters without tuning for frozen filters.
Experimental setup
Datasets: ImageNet-1K CIFAR-10 STL-10 Oxford 102 Flowers Oxford-IIIT Pets
Benchmarks: ImageNet-1K CIFAR-10 STL-10 Oxford 102 Flowers Oxford-IIIT Pets
Metrics: top-1 accuracy
Baselines: ConvNeXt V2 original trained models Hornet original model ImageNet filter transfer random 8 filters uniform 50/25/10 candidate filters
Builds on: Master Key Filters Hypothesis ConvNeXt V2 Hornet ConvNeXt Gaussian scale-space theory
Closest papers by meaning
- The Master Key Filters Hypothesis: Deep Filters Are General (AAAI 2025 · similarity 0.92)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.90)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.85)
- Normalize Filters! Classical Wisdom for Deep Vision (NeurIPS 2025 · similarity 0.85)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.84)
- Characterizing Universal Object Representations Across Vision Models (NeurIPS 2026 · similarity 0.84)
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.