Explainable Deep Learning on CIFAR-10: Grad-CAM & LIME

This project implements and compares two state-of-the-art explainable AI (XAI) methods—Grad-CAM and LIME—on image classification with a fine-tuned ResNet-50 model using the CIFAR-10 dataset. The code and report provide a full pipeline from data preprocessing and model training to XAI visualization and quantitative comparison.

Project Overview

• Goal: Train a robust image classifier on CIFAR-10 and analyze model decisions using Grad-CAM and LIME.
• Notebook: All code is in xAI_gradcam_lime.ipynb.
• Report: See the PDF for detailed analysis and results.

Model & Dataset

• Model: ResNet-50 (ImageNet-pretrained, fine-tuned for CIFAR-10)
  • Final layer replaced with Dropout + Linear for regularization
  • All layers fine-tuned (not just the head)
• Dataset: CIFAR-10 (60,000 images, 10 classes)
  • Images resized to 224x224, normalized with ImageNet stats
  • Data augmentation: random crop, horizontal flip, color jitter
  • Train/val/test split: 40k/10k/10k

Training & Regularization

• Loss: CrossEntropyLoss with label smoothing
• Optimizer: Adam (weight decay for L2 regularization)
• Scheduler: ReduceLROnPlateau (auto LR reduction)
• Early Stopping: Based on validation accuracy and per-class accuracy (prevents class collapse)
• Gradient Clipping: Prevents exploding gradients

Explainable AI (XAI) Methods

• Grad-CAM: Visualizes class-discriminative regions using gradients from the last convolutional layer
• LIME: Generates local explanations by perturbing superpixels and fitting a local surrogate model
• Comparison: Side-by-side visualizations and quantitative overlap (IoU) analysis

Evaluation & Metrics

• Accuracy: Overall and per-class accuracy on test set
• Confusion Matrix: Visualizes class confusion
• Sample Predictions: Shows correct/incorrect predictions with color-coded labels
• XAI Metrics:
  • Visual: Grad-CAM and LIME overlays for each class
  • Quantitative: IoU (Intersection over Union) between Grad-CAM and LIME masks, per-class and overall

Reproducing Results

1. Install dependencies:

   python -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt

2. Download CIFAR-10:
   • The notebook will auto-download if not present in data/.
3. Run the notebook:

   jupyter lab
   # Open xAI_gradcam_lime.ipynb and run all cells

4. Pretrained weights:
   • Provided: resnet50_cifar10_anti_overfitting.pth, resnet50_cifar10_best.pth, resnet50_cifar10_enhanced.pth
   • To retrain: set num_epochs and run training cells

Dependencies

See requirements.txt for all packages, including:

• torch, torchvision, numpy, matplotlib, seaborn, scikit-learn, tqdm, jupyter, ipywidgets
• lime, opencv-python, scikit-image

Tested Environment

• Python 3.10+
• PyTorch 2.5.1 (CPU or GPU)
• CUDA 11.8+ (optional, for NVIDIA GPUs)
• MPS support (Apple Silicon M1/M2/M3/M4)

---

For details on methodology, results, and interpretation, see the notebook and the report PDF.