🧠 Explainable AI for Image Classification
Seeing Through the Black Box: Explainability Techniques Applied to a Custom CNN
Authors: Matheus Braga (mbb4) · Philippe Menge (pmal)
📌 About
This project explores Explainable AI (xAI) methods applied to a Convolutional Neural Network (CNN) built with PyTorch. The goal is not only to achieve good classification performance, but to understand why the model makes each decision.
The model was optimized using Optuna (hyperparameter tuning) and evaluated with multiple explainability techniques.
🔍 Explainability Techniques
| Saliency Maps |
| Grad-CAM |
| LIME |
| RISE |
| Occlusion Sensitivity |
| Rejection |
🏗️ Model Architecture
Custom CNN with Residual Blocks
Batch Normalization + Dropout
Optimized with Optuna (automated hyperparameter search)
Training with early stopping
📊 Evaluation
Confusion Matrix (before and after rejection)
Per-class metrics (Precision, Recall, F1)
Precision-Recall curves