Understanding the black-box: towards interpretable and reliable deep learning models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep learning (DL) models offer advanced capabilities but often function as "black boxes", raising concerns about reliability.
• Explainable AI (XAI) is crucial for enhancing model transparency and interpretability, essential for real-time applications.

Purpose of the Study:

• To investigate the reliability and truthfulness of DL models in image classification.
• To develop and validate fruit classification models using transfer learning and XAI techniques.

Main Methods:

• Employed three pre-trained models (VGG16, MobileNetV2, ResNet50) with transfer learning for a 131-class fruit classification task.
• Utilized Local Interpretable Model-Agnostic Explanations (LIME), a popular XAI technique, to inspect model predictions and reliability.

Main Results:

• Transfer learning achieved optimized classification results with approximately 98% accuracy.
• LIME validation demonstrated that model predictions were interpretable and based on relevant image features.

Conclusions:

• The research validates the effectiveness of transfer learning and XAI in creating reliable and trustworthy DL models.
• Findings provide insights into interpreting complex AI models, increasing their accountability and trustworthiness for practical deployment.