🧠 Evaluation and Comparison of XAI Methods

This project evaluates multiple Explainable AI (XAI) techniques to understand how convolutional neural networks (CNNs) make classification decisions.
Using a pretrained VGG16 model on the MexCulture142/validation dataset, the framework compares different Explainability methods that produce saliency maps highlighting which regions in an image most influence the model’s prediction.

Overview

The system provides tools to:

• Generate saliency maps using multiple XAI methods.
• Compare those maps with human gaze fixation data.
• Evaluate both faithfulness (model-based) and human-alignment (ground-truth-based) metrics.
• Visualize and analyze method performance across architectural categories.

XAI Methods

Method	Description
FEM	Gradient-free, feature-based explanation highlighting statistically rare and strong activations.
Grad-CAM	Gradient-weighted, class-specific visualization showing which regions influence a model’s decision.
LIME	Perturbation-based local surrogate model estimating regional feature importance.
RISE	Randomized input sampling approach estimating pixel importance probabilistically.

Evaluation Metrics

Without Ground Truth

• AUC Deletion → Measures drop in confidence as key pixels are removed (lower = better faithfulness).
• AUC Insertion → Measures increase in confidence as key pixels are added (higher = better fidelity).

With Ground Truth (Human Gaze)

• PCC (Pearson Correlation Coefficient) → Linear correlation with gaze-based saliency maps.
• SIM (Similarity) → Measures spatial overlap between generated and gaze maps.

Quantitative Results

Key Findings

• FEM achieves the best PCC and SIM scores, aligning most closely with human attention.
• LIME and RISE perform best in AUC Insertion, reflecting strong model fidelity.
• Each method balances interpretability and accuracy differently.

Cross-Method Insights

Method	Strengths	Limitations
FEM	Excellent localization and faithfulness; ideal for complex textures.	Slightly computationally intensive.
Grad-CAM	Clear, intuitive activations.	Can miss fine details and over-smooth results.
LIME	Strong for structured geometry.	Struggles with texture-rich or irregular scenes.
RISE	Consistent across samples, robust black-box method.	Coarser visual precision due to random masking.

How to Run the Project

1- Run Evaluation

Execute the evaluation pipeline to generate saliency maps and raw XAI results:

python Evaluation.py

This will save output data to:

xai_results.pkl

2- Analyze and Summarize

Process and visualize the results:

python Analyze_xai_results.py

This script produces:

• Quantitative summaries and evaluation metrics
• Visualization plots under plots/

Project Structure

 XAI_Evaluation
├── Evaluation.py              # Runs evaluation and saves metrics
├── Analyze_xai_results.py     # Summarizes and visualizes results
├── XAI_Metrics.py             # Implements AUC Deletion, AUC Insertion, PCC, SIM
├── FEM.py                     # Feature-based gradient-free explanation
├── GradCAM.py                 # Gradient-weighted activation mapping
├── LIME.py                    # Local surrogate model explanations
├── RISE.py                    # Randomized input sampling explanation
├── plots/                     # Generated visualizations and curves