Research map / Label-efficient and robust learning / Recognition and long-tail
Studying How to Efficiently and Effectively Guide Models with Explanations
Sukrut Rao, Moritz Böhle, Amin Parchami-Araghi, Bernt Schiele
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 the Energy (EPG) loss instead of L1, PPCE or RRR* (ablation) | model guidance with bounding boxes · PASCAL VOC 2007 / MS COCO 2014 | EPG vs F1 trade-off | Improved |
| use the L1 loss instead of Energy, PPCE or RRR* (ablation) | model guidance with bounding boxes · PASCAL VOC 2007 | IoU vs F1 trade-off | Improved |
| use Energy loss rather than L1 to concentrate attributions on-object inside boxes (ablation) | model guidance · VOC test subset with segmentation masks | energy within segmentation mask vs box | Improved |
| use B-cos explanations instead of Vanilla or X-DNN attributions at input layer (ablation) | model guidance · PASCAL VOC 2007 | EPG @ F1: Vanilla 55.8 @ 69.0%, X-DNN 62.3 @ 68.9% -> B-cos 71.7 @ 79.4% | Improved |
| add model guidance with bounding boxes to X-DNN (ablation) | X-DNN ResNet-50 · PASCAL VOC 2007 | F1 and EPG: EPG +7.2 p.p., F1 +1.4 p.p. | Improved |
| model guidance overall trade-off between localization and classification | model guidance · PASCAL VOC 2007 (i.i.d. test) | F1 | Mixed |
| use only 1% of training images with box annotations (ablation) | Energy loss guidance · PASCAL VOC 2007 | EPG: +23.0 p.p. EPG at -0.3 p.p. F1 | Improved |
| dilate bounding boxes by 10, 25 or 50% during training (ablation) | Energy loss · PASCAL VOC 2007 | EPG and IoU | No clear effect |
| dilate bounding boxes by 10, 25 or 50% during training (ablation) | L1 loss · PASCAL VOC 2007 | EPG and IoU | Hurt |
| apply guidance at deeper layer Mid2 instead of input (ablation) | B-cos guidance · PASCAL VOC 2007 | input-level EPG / training time: 1.7x speed-up with similar localization gains | Improved |
| add model guidance with 1% bounding boxes under spurious correlation (ablation) | B-cos model · Waterbirds-100 | worst-group / overall accuracy: Energy conventional: worst 56.1, overall 71.2 | Improved |
Closest papers by meaning
- Learning Robust Rationales for Model Explainability: A Guidance-Based Approach (AAAI 2024 · similarity 0.85)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.84)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.84)
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.84)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.84)
- Model Steering: Learning with a Reference Model Improves Generalization Bounds and Scaling Laws (ICML 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.
