GCAV: A Global Concept Activation Vector Framework for Cross-Layer Consistency in Interpretability

📝 Paper

GCAV: A Global Concept Activation Vector Framework for Cross-Layer Consistency in Interpretability
Zhenghao He, Sanchit Sinha, Guangzhi Xiong, Aidong Zhang
University of Virginia
ICCV 2025 (Accepted)

🧩 Overview

Global Concept Activation Vector (GCAV) is a novel framework designed to improve the interpretability of deep neural networks by addressing the fundamental limitations of traditional Testing with Concept Activation Vectors (TCAV).

Key Problems Solved

Traditional TCAV faces several critical issues:

• 🔄 Unstable layer selection: The same concept may have varying importance across layers
• ❌ Spurious activations: Certain layers assign high TCAV scores to irrelevant concepts
• 📊 High variance in TCAV scores: Inconsistencies in concept representation across layers

Our Solution

GCAV unifies CAVs into a single, semantically consistent representation through:

• 🎯 Contrastive learning to align concept representations across layers
• 🧠 Attention-based fusion to construct a globally integrated CAV
• 🔄 Decoder-based projection to enable TCAV evaluation on GCAV representations (TGCAV)

Key Advantages

• ✅ Cross-Layer Concept Consistency: Eliminates inconsistencies in TCAV scores across layers
• 🎯 Improved Concept Localization: Enhances interpretability by refining concept attributions
• 🛡️ Adversarial Robustness: Provides more stable explanations under adversarial perturbations
• 🔧 Architecture Agnostic: Works with multiple deep architectures (ResNet, GoogleNet, MobileNet)

🏗️ Framework Architecture

Our framework consists of three sequential training stages:

1. Layer-wise Autoencoder Training: Compress high-dimensional CAVs into unified embedding space
2. Cross-Layer Alignment: Use contrastive learning with InfoNCE and consistency loss
3. Global CAV Fusion: Attention-based fusion with variance and consistency optimization

🛠 Installation

Setup Environment

# Clone the repository
git clone https://github.com/your-username/CAVFusion.git
cd CAVFusion

# Create conda environment
conda create --name gcav python=3.9
conda activate gcav

# Install dependencies
pip install -r requirements.txt

Key Dependencies

torch>=2.1.0
tensorflow>=2.18.0
tcav>=0.2.2
scikit-learn>=1.6.0
matplotlib>=3.9.0
numpy>=1.26.4
h5py>=3.12.1

🚀 Quick Start

1. Dataset Preparation

Download the required datasets and pre-trained models:

# Download Broden dataset for concept images
python src/download_data.py

# Download pre-trained models (GoogleNet, ResNet50V2, MobileNetV2)
python src/load_models.py

2. Configuration

Edit src/configs.py to specify your experimental setup:

# Target class and concepts
target = 'zebra'
concepts = ["dotted", "striped", "zigzagged"]

# Model selection
model_to_run = 'GoogleNet'  # or 'ResNet50V2', 'MobileNetV2'

# Training parameters
embed_dim = 2048
hidden_dims = [4096]
nce_loss = 1
con_loss = 3
var_loss = 3
sim_loss = 1

3. Generate Original CAVs

python src/get_cavs_main.py

4. Train GCAV Framework

# Stage 1: Train autoencoders for dimension alignment
python src/align_dim.py

# Stage 2 & 3: Train alignment and fusion models
python src/integrate_main.py

5. Evaluate Results

# Generate TGCAV scores and visualizations
python src/evaluate.py

# Plot comparative results
python src/plot_result.py

📊 Experimental Results

Performance Improvements

Our experiments on GoogleNet, ResNet50V2, and MobileNetV2 demonstrate significant improvements. The following table shows detailed results for different target classes and concepts:

Target	Concept	Method	GoogleNet				ResNet50V2				MobileNetV2
			Mean	Std	CV	RR	Mean	Std	CV	RR	Mean	Std	CV	RR
Spider Web	cobwebbed	TCAV	0.808	0.140	0.173	0.508	0.644	0.127	0.197	0.715	0.619	0.169	0.273	0.985
		TGCAV	0.840	0.109	0.130	0.381	0.784	0.083	0.105	0.357	0.554	0.091	0.164	0.686
	porous	TCAV	0.499	0.123	0.246	0.862	0.586	0.098	0.167	0.717	0.377	0.059	0.157	0.584
		TGCAV	0.498	0.037	0.074	0.281	0.640	0.030	0.047	0.156	0.296	0.032	0.108	0.338
	blotchy	TCAV	0.459	0.210	0.457	1.460	0.574	0.107	0.186	0.871	0.438	0.133	0.304	1.005
		TGCAV	0.393	0.057	0.144	0.508	0.618	0.026	0.042	0.129	0.386	0.043	0.111	0.466
Zebra	dotted	TCAV	0.463	0.218	0.470	1.576	0.470	0.063	0.135	0.426	0.484	0.130	0.268	0.950
		TGCAV	0.373	0.019	0.051	0.161	0.479	0.046	0.096	0.376	0.408	0.021	0.051	0.196
	striped	TCAV	0.783	0.185	0.236	0.664	0.593	0.073	0.123	0.421	0.601	0.145	0.241	0.832
		TGCAV	0.699	0.042	0.060	0.200	0.597	0.050	0.084	0.318	0.600	0.017	0.029	0.083
	zigzagged	TCAV	0.647	0.232	0.359	0.974	0.578	0.051	0.088	0.277	0.618	0.095	0.153	0.615
		TGCAV	0.503	0.027	0.055	0.159	0.597	0.049	0.083	0.268	0.628	0.031	0.049	0.175

Metrics Explanation:

• Mean: Average TCAV/TGCAV scores across layers
• Std: Standard deviation (lower = more stable)
• CV: Coefficient of variation (lower = more consistent)
• RR: Range ratio (lower = less fluctuation across layers)

Key Findings

• 🎯 Reduced Variance: TGCAV significantly reduces variance in concept activation scores across layers
• 🛡️ Adversarial Robustness: 40-60% more stable under adversarial attacks compared to traditional TCAV
• 📈 Better Localization: More precise concept attribution with reduced spurious activations

Adversarial Robustness Evaluation

We evaluated the robustness of GCAV against adversarial attacks using Brown's method, targeting the "dotted" concept in the "zebra" class at the GoogleNet mixed5a layer:

Method	Mixed5a	Mixed5a Change (%)	Mean	Mean Change (%)
TCAV	0.52	--	0.46	--
TCAV (Attacked)	0.96	+84.61%	0.75	+61.77%
TGCAV	0.38	--	0.37	--
TGCAV (Attacked)	0.49	+28.49%	0.53	+42.63%

Key Robustness Insights:

• Traditional TCAV: Vulnerable to adversarial manipulation, with up to 84.61% increase in targeted layer scores
• Our TGCAV: More resilient with only 28.49% increase under the same attack
• Global Integration: By fusing information across layers, GCAV reduces the impact of single-layer perturbations

Visual Results

Cross-Layer Consistency Improvement (Figure 2)

The following figure shows TCAV score comparisons across layers on GoogleNet before (left) and after (right) applying our method:

Key Observations:

• Original TCAV (Left): High variance across layers, spurious activations in irrelevant concepts
• Our TGCAV (Right): Consistent scores across layers, reduced spurious activations
• Cross-layer stability: TGCAV eliminates layer selection ambiguity

Concept Localization Maps (Figure 3)

Our GCAV framework provides improved concept localization through Visual-TCAV concept mapping. The visualization shows three representative concepts with original method (top row) vs our method (bottom row):

Striped (Zebra)	Cobwebbed (Spider Web)	Honeycombed (Honeycomb)

Concept Localization Benefits:

• 🎯 Precise Focus: Our method focuses more precisely on characteristic patterns while reducing distractions from irrelevant regions
• 🔍 Reduced Background Noise: Less activation in background regions (sky, surrounding environment) compared to original TCAV
• 🌐 Semantic Consistency: Better alignment with expected concept locations - stripes on zebra body, web structures, honeycomb patterns
• 📊 Enhanced Interpretability: More concentrated activations on defining structural patterns, eliminating unnecessary activations in non-characteristic areas

📁 Project Structure

CAVFusion/
├── src/                      # Main source code
│   ├── configs.py           # Configuration settings
│   ├── get_cavs_main.py     # Generate original CAVs
│   ├── get_cavs.py          # CAV generation utilities
│   ├── align_dim.py         # Autoencoder training
│   ├── IntegrateCAV.py      # Core GCAV framework
│   ├── integrate_main.py    # Training pipeline
│   ├── evaluate.py          # Evaluation scripts
│   ├── plot_reslut.py       # Visualization tools
│   ├── attack.py            # Adversarial attack experiments
│   ├── concept_map.py       # Concept mapping utilities
│   ├── reconstruct.py       # CAV reconstruction
│   ├── load_models.py       # Model loading utilities
│   ├── download_data.py     # Data download script
│   ├── download_img.py      # Image download utilities
│   ├── select_img.py        # Image selection tools
│   ├── extract_500_birds.py # Bird dataset extraction
│   ├── createRandom.py      # Random dataset generation
│   └── concate.py           # Data concatenation utilities
├── imgs/                    # Images for README
│   ├── framework.png        # Framework architecture
│   ├── tcav_*.png          # Original TCAV results
│   └── tgcav_*.png         # Our TGCAV results
├── requirements.txt         # Dependencies
├── .gitignore              # Git ignore rules
├── LICENSE                 # MIT License
└── README.md               # This file

🎛️ Configuration Options

Model Selection

model_to_run = 'GoogleNet'    # GoogleNet (Inception)
model_to_run = 'ResNet50V2'   # ResNet50V2  
model_to_run = 'MobileNetV2'  # MobileNetV2

Fusion Methods

fuse_input = "aligned_cavs"    # Use aligned CAVs (recommended)
fuse_input = "input_cavs"      # Use original CAVs
fuse_input = "none_fuse"       # No fusion baseline

Training Parameters

# Alignment stage loss weights
nce_loss = 1      # InfoNCE loss weight
con_loss = 3      # Consistency loss weight

# Fusion stage loss weights  
var_loss = 3      # Variance loss weight
sim_loss = 1      # Similarity loss weight

🧪 Running Custom Experiments

Add New Concepts

1. Prepare concept images in data/concepts/your_concept/
2. Update src/configs.py:

   target = 'your_target_class'
   concepts = ["your_concept1", "your_concept2"]
   

3. Run the full pipeline as described in Quick Start

Experiment with Different Architectures

Our framework supports any CNN architecture. To add a new model:

1. Implement model wrapper in tcav/model.py
2. Define bottleneck layers in src/configs.py
3. Update model loading logic

📈 Evaluation Metrics

Stability Metrics

• Coefficient of Variation (CV): Measures relative variability
• Standard Deviation (Std): Absolute measure of dispersion
• Range Ratio (RR): Normalized range of activation scores

Robustness Evaluation

• Adversarial Stability: Performance under adversarial attacks
• Cross-Layer Consistency: Correlation between layer-wise activations
• Concept Localization: Precision of concept attribution

📄 Citation

If you find this work useful, please cite our paper:

@inproceedings{he2025gcav,
  title={GCAV: A Global Concept Activation Vector Framework for Cross-Layer Consistency in Interpretability},
  author={He, Zhenghao and Sinha, Sanchit and Xiong, Guangzhi and Zhang, Aidong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025}
}

📞 Contact

• Zhenghao He: zhenghao@virginia.edu
• Institution: University of Virginia

🙏 Acknowledgments

• Built upon the TCAV framework by Kim et al.
• Uses the Broden dataset for concept visualization
• Supported by the University of Virginia Computer Science Department

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

⭐ Star this repository if you find it helpful!