Research map / Adversarial robustness and security / Attacks on text and graphs
Visualizing the Emergence of Intermediate Visual Patterns in DNNs
Mingjie Li, Shaobo Wang, Quanshi Zhang
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 |
|---|---|---|---|
| project sample features with the proposed vMF-based visualization instead of PCA, t-SNE, LLE, ISOMAP or DRPR | sample-feature visualization · Tiny ImageNet / COCO / CUB | Pearson correlation between feature strength and classification uncertainty | Improved |
| compare student DNN distilled at conv_53 against the teacher in number of reliable knowledge points | VGG-16 student via knowledge distillation · Tiny ImageNet | number of reliable knowledge points | Hurt |
| distill from penultimate fully-connected layer instead of conv_53 layer (ablation) | VGG-16 student via knowledge distillation · Tiny ImageNet | number of reliable knowledge points in conv layers | Hurt |
| distill from fully-connected layer rather than conv_53, measured at conv_53 (ablation) | VGG-16 student via knowledge distillation · Tiny ImageNet | feature similarity to teacher (orientation and strength) | Hurt |
| go from low layer (conv_33) to high layer (conv_53) | VGG-16 · Tiny ImageNet | ratio of reliable knowledge points | Improved |
| apply untargeted PGD attack (epsilon 1/255, 20 steps) and compare layers | VGG-16 · Tiny ImageNet | change in regional feature orientation and strength | Mixed |
| add alignment loss L_align (alpha=0.1) to similarity loss for regional features | regional feature projection · Tiny ImageNet | qualitative alignment with sample coordinate system | No clear effect |
| estimated regional importance w versus Shapley value | importance estimation · VGG-16 conv_53, multiple datasets | similarity between w and Shapley value | Improved |
| relate projected feature strength to estimated importance | regional features h · Tiny ImageNet, COCO, CUB | Pearson correlation of strength and importance | Improved |
| choose threshold tau=0.4 for knowledge points (larger vs smaller tau) (ablation) | knowledge point counting · VGG-16 | number of knowledge points | Mixed |
Closest papers by meaning
- Explaining Knowledge Distillation by Quantifying the Knowledge (CVPR 2020 · similarity 0.87)
- Monitoring Primitive Interactions During the Training of DNNs (AAAI 2025 · similarity 0.85)
- Don't trust your eyes: on the (un)reliability of feature visualizations (ICML 2024 · similarity 0.84)
- Layerwise Change of Knowledge in Neural Networks (ICML 2024 · similarity 0.84)
- Neuro-Visualizer: A Novel Auto-Encoder-Based Loss Landscape Visualization Method With an Application in Knowledge-Guided Machine Learning (ICML 2024 · similarity 0.83)
- Understanding Visual Feature Reliance through the Lens of Complexity (NeurIPS 2024 · similarity 0.83)
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.
