Research map / Architectures and efficiency / RNNs and normalization
FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FastCAV, which computes concept activation vectors as the normalized direction from the global activation mean to the concept mean, achieving up to 63.6x speedup over SVM-based CAVs with similar quality.
Problem
Computing Concept Activation Vectors (CAVs) with linear SVMs is computationally expensive and slow in high-dimensional activation spaces of modern deep networks, limiting large-scale concept-based analyses.
Approach
Replace SVM training with a closed-form direction: the normalized vector from the global mean of concept and random activations toward the concept mean. Theoretically, under Gaussian and isotropic-covariance assumptions in high dimensions, this equals the Fisher/linear SVM normal vector.
Key results
- FastCAV is up to 63.6x faster than SVM-CAV (average 46.4x) across evaluated architectures.
- FastCAV reaches comparable or slightly higher CAV accuracy than SVM-CAV on most models (e.g., ResNet50 0.89 vs 0.87), with higher intra-method similarity (ResNet50 0.816 vs 0.400).
- Inter-method cosine similarity between FastCAV and SVM-CAV CAVs is high, around 0.8 to 0.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace linear SVM CAV computation with FastCAV | CAV computation across evaluated ImageNet models · Broden concepts (ImageNet random), all layers | computation time: up to $63.6\times$ (on average $46.4\times$) faster | Helped |
| Replace linear SVM CAV computation with FastCAV | CAV computation across evaluated ImageNet models · Broden concepts (ImageNet random), all layers | accuracy | Mixed |
| Replace linear SVM CAV computation with FastCAV | CAV computation across evaluated ImageNet models · Broden concepts (ImageNet random), all layers | intra-method cosine similarity | Helped |
| Use more random sets $D_r$ | FastCAV · CAV accuracy sensitivity analysis | variance | Helped |
| Use more random sets $D_r$ | FastCAV · CAV accuracy sensitivity analysis | average accuracy | No clear effect |
| Increase concept and random examples to at least 60 each | FastCAV · CAV accuracy sensitivity analysis | accuracy | Helped |
Surprising findings
- In 43.6% of cases FastCAV gave higher CAV accuracy than SVM-CAV, while SVM-CAV won in 36.9%; vanilla LDA performed much worse in similarity.
- Most activation vectors become support vectors in high dimensions, supporting the theoretical link between SVM and mean-difference solutions.
Limitations
- Equivalence to SVM depends on strong assumptions (Gaussian, isotropic within-class covariance) that may not hold; accuracy gaps up to 40 points observed in some cases.
- Only CAV quality is evaluated; other CAV properties such as locality, consistency, and entanglement are left for future work.
Experimental setup
Datasets: ImageNet Broden MIMIC-CXR
Benchmarks: Broden concepts (ImageNet-based CAV evaluation) MIMIC-CXR medical concepts
Metrics: computation time CAV accuracy cosine similarity (intra-method robustness) inter-method cosine similarity TCAV score
Baselines: SVM-CAV sparse logistic regression logistic regression ridge classification LDA
Builds on: Concept Activation Vectors (TCAV) Automatic Concept-based Explanations (ACE) linear SVM Fisher discriminant analysis
Closest papers by meaning
- Invertible Concept-based Explanations for CNN Models with Non-negative Concept Activation Vectors (AAAI 2021 · similarity 0.90)
- Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence (ICLR 2025 · similarity 0.89)
- LG-CAV: Train Any Concept Activation Vector with Language Guidance (NeurIPS 2024 · similarity 0.86)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.86)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.86)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · 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.