A Comparative Analysis of Explainable AI (XAI) Techniques for Transparent and Reliable Image Classification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Black-box machine learning models lack transparency, hindering trust and adoption.
• Explainable AI (XAI) methods are crucial for providing human-understandable justifications for model decisions.
• Selecting the appropriate XAI technique is vital for specific machine learning tasks.

Purpose of the Study:

• To evaluate and compare the performance of four XAI techniques: PEEK, LRP, GRAD-CAM, and LIME.
• To assess these XAI methods on image classification tasks using the Oxford IIT pet dataset.
• To analyze the performance, robustness, generalizability, noise stability, and computational efficiency of each XAI method.

Main Methods:

• Utilized the Oxford IIT pet dataset (7390 images) for training custom Convolutional Neural Network (CNN) and VGG16 models.
• Analyzed saliency maps generated by PEEK, LRP, GRAD-CAM, and LIME to identify critical image regions for classification.
• Employed noise analysis, robustness checks, run-time measurements, and faithfulness metrics for comprehensive evaluation.

Main Results:

• XAI methods generally identified intuitive, critical features (e.g., outlines, faces, eyes) for accurate classification.
• Significant variation and marginal consensus were observed among XAI methods in pinpointing these critical features.
• Grad-CAM showed strong robustness and stability with the VGG16 model, but inconsistent performance with a shallow CNN.

Conclusions:

• While XAI methods offer insights into black-box model reasoning, their consensus on critical features is limited.
• Grad-CAM shows promise for robust explanations in complex models like VGG16.
• Further research is needed to improve consensus and tailor XAI methods for reliable image classification explanations.