Research map / Label-efficient and robust learning / Recognition and long-tail
GCAV: A Global Concept Activation Vector Framework for Cross-Layer Consistency in Interpretability
Zhenghao He, Sanchit Sinha, Guangzhi Xiong, Aidong 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 |
|---|---|---|---|
| replace per-layer TCAV with globally fused TGCAV | GCAV · Zebra/dotted, GoogleNet | Std across layers: 0.218 -> 0.019 | Improved |
| replace per-layer TCAV with TGCAV | GCAV · Zebra/zigzagged, GoogleNet | Range ratio: 0.974 -> 0.159 | Improved |
| replace TCAV with TGCAV | GCAV · Honeycomb/honeycombed | mean score: GoogleNet 0.836 -> 0.897; ResNet50V2 0.648 -> 0.591; MobileNetV2 0.572 -> 0.511 | Mixed |
| add global fusion and alignment (TGCAV) under targeted adversarial attack on mixed5a | GCAV · dotted concept (zebra), GoogleNet mixed5a | score increase under attack: +84.61% -> +28.49% | Improved |
| add global fusion (TGCAV) under attack, mean across layers | GCAV · dotted concept (zebra), GoogleNet | mean score increase under attack: +61.77% -> +42.63% | Improved |
| add cross-layer fusion step (ablation) | GCAV · Zebra/dotted, GoogleNet | Std across layers: 0.297 -> 0.019 | Improved |
| add contrastive alignment step (ablation) | GCAV · Zebra/striped, GoogleNet | Std across layers: 0.134 -> 0.042 | Improved |
| add consistency loss (lambda_cons > 0) (ablation) | GCAV · Zebra/striped, GoogleNet | TGCAV score | Improved |
| vary embedding size (1024, 2048, 4096) (ablation) | GCAV · Zebra/striped, GoogleNet | Std across layers: 0.042 (1024), 0.069 (2048) | No clear effect |
| use lambda_var:lambda_cons 3:1 instead of 1:3 (ablation) | GCAV · Zebra/striped, GoogleNet | Std across layers: 0.192 -> 0.042 | Improved |
| apply GCAV to high-level concepts on CelebA | GCAV · CelebA human (groom) | Std across layers: 0.175 -> 0.056 | Improved |
Closest papers by meaning
- FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks (ICML 2025 · similarity 0.91)
- Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence (ICLR 2025 · similarity 0.88)
- LG-CAV: Train Any Concept Activation Vector with Language Guidance (NeurIPS 2024 · similarity 0.87)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.86)
- Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models (CVPR 2024 · similarity 0.85)
- Controlling Large Language Models Through Concept Activation Vectors (AAAI 2025 · 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.
