Research map / Architectures and efficiency / Pruning and efficient tuning
SPADE: Sparsity-Guided Debugging for Deep Neural Networks
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.
Proposes SPADE, a sample-specific one-shot pruning preprocessing step that sparsifies a dense pretrained network per input so that existing saliency and neuron-visualization methods give more accurate, human-useful explanations of the original model.
Problem
Interpretability methods (saliency maps, neuron visualizations) are confounded by polysemantic neurons and superimposed features; existing sparsity-based fixes require retraining or alter the model's behavior, so their explanations no longer reflect the original dense model.
Approach
Treat sparsification as a per-sample interpretation preprocessing step: augment the input, then layer-by-layer one-shot prune the pretrained network (OBC or SparseGPT solvers) so each layer's sparse output matches its dense output on that sample, and run any interpretability method on the resulting trace, without retraining or changing inference on the sample.
Key results
- On ResNet50/ImageNet, SPADE raises average Trojan saliency AUC from 91.01% (dense) to 96.55%, versus 91.50% for Sparse FC.
- In a human study (24 evaluators, 746 tasks), overall success identifying the Score-CAM-matched region rose from 56.7% with dense visualizations to 69.8% with SPADE.
- FastSPADE runs in about 3 seconds per ResNet50 example on a single GPU, and a 21,121-sample ImageNet evaluation took about 120 GPU-hours.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add sample-specific SPADE pruning before saliency interpretation | ResNet50 saliency methods · ImageNet-1K Trojan patches | Average AUC (%): 91.01 -> 96.55 | Helped |
| Use FastSPADE sample-specific pruning before saliency interpretation | ResNet50 saliency methods · ImageNet-1K Trojan patches | Average AUC (%): 91.01 -> 94.34 | Helped |
| Add FastSPADE preprocessing | ResNet50 saliency methods · ImageNet-1K clean inputs | Insertion AUC (%): 44.18 -> 47.20 | Helped |
| Add FastSPADE preprocessing | ResNet50 saliency methods · ImageNet-1K clean inputs | Deletion AUC (%): 10.62 -> 9.82 | Helped |
| Preprocess the model with SPADE before neuron visualization | ResNet50 feature visualizations · ImageNet-1K human evaluation | Overall success: 56.7% -> 69.8% | Helped |
| Use a linear sparsity schedule instead of searched layer sparsities (ablation) | SPADE · ImageNet ResNet50 Trojan patches | Average AUC: 95.78 -> 93.51 | Hurt |
| Tune sparsity ratios using 30 examples instead of 100 (ablation) | FastSPADE · ImageNet ResNet50 Trojan patches | Average AUC: 94.34 -> 93.97 | Hurt |
| Tune SPADE sparsity ratios using FastSPADE (ablation) | SPADE · ImageNet ResNet50 Trojan patches | Average AUC: 96.55 -> 94.58 | Hurt |
| Use sample-specific SPADE pruning | ResNet50 saliency methods · CelebA Trojan patches | Average AUC: 80.71 -> 88.80 | Mixed |
| Use sample-specific SPADE pruning | ResNet50 saliency methods · Food-101 Trojan patches | Average AUC: 83.54 -> 95.33 | Helped |
| Use OBC-based SPADE instead of FastSPADE | MobileNet saliency methods · ImageNet | Average AUC improvement: +2.90% | Helped |
| Use sample-specific SPADE pruning | BERT Layer Integrated Gradients · SST-2 | DFFOT: 0.1835 -> 0.1766 | Hurt |
Surprising findings
- Sparse FC, a globally retrained sparse classification head, gives only a 0.49% average AUC gain, far below the per-sample SPADE gains.
- Only the accurate OBC pruner improved MobileNet; FastSPADE gave no improvement over the dense baseline on that architecture.
Limitations
- Tuning per-layer sparsity ratios is costly and may impede practical adoption.
- SPADE may systematically favor some categories of examples; the polysemanticity explanation is conjectural.
- Saliency accuracy is mainly evaluated with Trojan patches, whose applicability to dispersed-evidence images is unclear.
Experimental setup
Datasets: ImageNet-1K CelebA Food-101 SST-2 AG news
Benchmarks: ImageNet-1K Trojan patch AUC CelebA Trojan patch AUC Food-101 Trojan patch AUC Insertion/Deletion AUC on clean ImageNet
Metrics: AUROC insertion AUC deletion AUC human agreement rate with Score-CAM
Baselines: Dense model Sparse FC Correlation-aware pruning (CAP)
Builds on: OBC SparseGPT Sparse FC Captum
Closest papers by meaning
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.87)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.86)
- Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective (AAAI 2025 · similarity 0.85)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.85)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 2024 · similarity 0.85)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · 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.