Research map / Label-efficient and robust learning / Recognition and long-tail
Labeling Neural Representations with Inverse Recognition
Kirill Bykov, Laura Kopf, Shinichi Nakajima, Marius Kloft, Marina MC Höhne
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 |
|---|---|---|---|
| use INVERT instead of Network Dissection (label-only vs segmentation masks) | neuron labeling · ImageNet ResNet18 output neurons | explanation accuracy: 19.2% -> 73.2% | Improved |
| use INVERT instead of Network Dissection | neuron labeling · ImageNet GoogleNet output neurons | explanation accuracy: 19.7% -> 82.2% | Improved |
| use INVERT instead of Network Dissection | neuron labeling · ImageNet DenseNet161 output neurons | explanation accuracy: 19.1% -> 86.9% | Improved |
| use INVERT instead of Network Dissection | neuron labeling · MaskRCNN ResNet50 FPN on MS COCO | explanation accuracy: 95.06% -> 98.77% | Improved |
| use INVERT instead of Network Dissection | neuron labeling · FCN ResNet50 on MS COCO | explanation accuracy: 95.24% -> 95.24% | No clear effect |
| increase concept-fraction constraint alpha and formula length L (ablation) | INVERT · ImageNet feature-extractor neurons (ResNet18, GoogleNet, EfficientNet B0, ViT B 16) | AUC of explanation | Mixed |
| add a Goedel AND over the top-L neurons instead of a single neuron (ablation) | training-free transfer, EfficientNet B4 · Caltech101 46 classes | accuracy: 39.96% -> 86.11% | Improved |
| add a Goedel AND over the top-L neurons (L=1 to 5) (ablation) | training-free transfer, DenseNet161 · Caltech101 46 classes | accuracy: 62.50% -> 75.00% | Improved |
| add a Goedel AND over the top-L neurons (L=1 to 10) (ablation) | training-free transfer, ViT L 16 · Caltech101 46 classes | accuracy: 47.12% -> 79.49% | Improved |
| add a Goedel AND over the top-L neurons (L=1 to 10) (ablation) | training-free transfer, AlexNet · Caltech101 46 classes | accuracy: 43.91% -> 30.88% | Hurt |
| use a single INVERT-selected neuron instead of a trained linear head | training-free transfer, DenseNet161 · Caltech101 46 classes | accuracy: 97.76% -> 62.50% | Hurt |
| compare Goedel against Product, Lukasiewicz and Yager fuzzy operators (ablation) | neuron circuits · random compositional concepts on four ImageNet models | mean AUC | Improved |
Closest papers by meaning
- Making Sense of CNNs: Interpreting Deep Representations & Their Invariances with INNs (ECCV 2020 · similarity 0.88)
- InversionView: A General-Purpose Method for Reading Information from Neural Activations (NeurIPS 2024 · similarity 0.86)
- Network Inversion of Convolutional Neural Nets (Student Abstract) (AAAI 2025 · similarity 0.85)
- Tilt your Head: Activating the Hidden Spatial-Invariance of Classifiers (ICML 2024 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.84)
- Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision (AAAI 2024 · 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.
