Research map / Label-efficient and robust learning / Recognition and long-tail
Local Concept Embeddings for Analysis of Concept Distributions in Vision DNN Feature Spaces
Georgii Mikriukov, Gesina Schwalbe, Korinna Bade
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 zeros initialization instead of ones for the concept vector (ablation) | LoCE · MS COCO 950 samples, YOLOv5 layer l3 | IoU: 0.21 -> 0.55 | Improved |
| use zeros initialization instead of ones for the concept vector (ablation) | LoCE · MS COCO 950 samples, YOLOv5 layer l3 | optimization failures %: 73.2 -> 22.2 | Improved |
| use zeros initialization instead of ones (ablation) | LoCE · MS COCO 950 samples, DETR layer l6 | optimization failures %: 96.6 -> 1.2 | Improved |
| use zeros instead of ones initialization (ablation) | LoCE · MS COCO 950 samples, MobileNetV3 layer l1 | optimization failures %: 3.8 -> 0.8 | Improved |
| increase loss resolution from 50x50 to 200x200 (ablation) | LoCE · MS COCO 950 samples, SWIN layer l3 | IoU: 0.07 -> 0.60 | Improved |
| increase loss resolution from 50x50 to 200x200 (ablation) | LoCE · MS COCO 950 samples, SWIN layer l1 | optimization failures %: 99.4 -> 71.7 | Improved |
| increase loss resolution from 100x100 to 200x200 (ablation) | LoCE · MS COCO 950 samples, EfficientNet layer l8 | IoU: 0.55 -> 0.52 | Hurt |
| increase loss resolution from 50x50 to 200x200 (ablation) | LoCE · MS COCO 950 samples, MobileNetV3 layer l6 | IoU: 0.69 -> 0.67 | Hurt |
| use image-local LoCE instead of global Net2Vec vector | LoCE · MS COCO overall, DETR deep layer | IoU: 0.56 -> 0.68 | Improved |
| use sub-global instead of global generalized vector | LoCE generalization · MS COCO overall, SWIN deep layer | IoU: 0.14 -> 0.49 | Improved |
| use sparse 16-filter Net2Vec instead of dense Net2Vec | Net2Vec · MS COCO overall, DETR deep layer | IoU: 0.56 -> 0.33 | Hurt |
| generalize LoCEs to a global vector instead of sub-global vectors | LoCE · MS COCO overall, DETR deep layer | IoU: 0.61 -> 0.45 | Hurt |
Closest papers by meaning
- Concept Activation Regions: A Generalized Framework For Concept-Based Explanations (NeurIPS 2022 · similarity 0.90)
- FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks (ICML 2025 · similarity 0.86)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.86)
- ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models (CVPR 2025 · similarity 0.86)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.85)
- ConceptBed: Evaluating Concept Learning Abilities of Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.85)
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.
