Research map / Graph learning / Knowledge and temporal graphs
Schema Inference for Interpretable Image Classification
Haofei Zhang, Mengqi Xue, Xiaokang Liu, Kaixuan Chen, Jie Song, Mingli Song
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 |
|---|---|---|---|
| Initialize IR-Atlas by averaging instance IR-Graphs and sparsify (remove low-weight vertices/edges) (ablation) | SchemaNet with DeiT-Tiny · CIFAR-100 / Caltech-101 | Top-1 accuracy: +1.34 (CIFAR-100), +4.33 (Caltech-101) | Improved |
| Initialize IR-Atlas by averaging instance IR-Graphs and sparsify (ablation) | SchemaNet with DeiT-Tiny · CIFAR-10 | Top-1 accuracy: +0.04 | Improved |
| Use raw attention for ingredient importance instead of CAM (ablation) | SchemaNet-Init with DeiT-Tiny · CIFAR-10 | Top-1 accuracy: +2.38 | Improved |
| Use raw attention for ingredient importance instead of ViT-LRP (ablation) | SchemaNet-Init with DeiT-Tiny · CIFAR-10 | Top-1 accuracy: +4.68 | Improved |
| Use raw attention for ingredient importance instead of CAM (ablation) | SchemaNet-Init with DeiT-Tiny · CIFAR-100 | Top-1 accuracy: +0.90 | Improved |
| Replace SchemaNet-Init with BagNet-style summation (BagNet baseline) on DeiT-Tiny | Visual word matching on DeiT-Tiny · CIFAR-10 | Top-1 accuracy: -25.69 vs SchemaNet-Init | Hurt |
| SchemaNet-Init (graph matching with interacting ingredients) vs Backbone-FC | DeiT-Tiny · CIFAR-10 / CIFAR-100 / Caltech-101 | Top-1 accuracy: +1.38 (CIFAR-10), +1.37 (CIFAR-100), +10.88 (Caltech-101) | Improved |
| Graph-based matching (SchemaNet) vs bag-of-visual-words with discretized words (BoVW-Deep) on DeiT-Tiny | Visual word interactions · CIFAR-100 / Caltech-101 | Top-1 accuracy: +1.66 (CIFAR-100), +22.71 (Caltech-101) using SchemaNet vs BoVW-Deep | Improved |
| Adding SchemaNet-Init graph matching vs frozen-backbone FC on ImageNet-1k | DeiT-Small · ImageNet-1k | Top-1 accuracy: +4.82 | Improved |
| Graph matching (SchemaNet-Init) vs BagNet on adversarial benchmarks | DeiT-Small trained on ImageNet-1k · ImageNet-A / ImageNet-R | Top-1 accuracy: +7.52 (ImageNet-A), +2.53 (ImageNet-R) | Improved |
| Replace SchemaNet with ViT-LRP attribution cost trade-off (ablation) | Ingredient importance computation · CIFAR-10 | running time: 53.8h vs 0.8h | Hurt |
Closest papers by meaning
- Classification by Attention: Scene Graph Classification with Prior Knowledge (AAAI 2021 · similarity 0.87)
- What Makes a Maze Look Like a Maze? (ICLR 2025 · similarity 0.85)
- Scene Graph Thinking: Reinforcing Structured Visual Reasoning for Multimodal Large Language Models (ICML 2026 · similarity 0.85)
- STELAR-VISION: Self-Topology-Aware Efficient Learning for Aligned Reasoning in Vision (AAAI 2026 · similarity 0.84)
- Concept-RuleNet: Grounded Multi-Agent Neurosymbolic Reasoning in Vision Language Models (AAAI 2026 · similarity 0.84)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 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.
