CIFAR-10 CNN Image Classifier

🚀 Project Overview

• CNN architecture design from first principles
• Training and evaluation on the CIFAR-10 dataset
• Overfitting detection and mitigation
• Confusion matrix–based error analysis
• Clean, modular ML project structure

🧠 Dataset

• 60,000 color images (32×32)
• 10 classes:
  • airplane, automobile, bird, cat, deer
  • dog, frog, horse, ship, truck
• 50,000 training images
• 10,000 test images

🏗️ Model Architecture

• 3 Convolutional blocks
  • Conv2D → Batch Normalization → ReLU → MaxPooling
• Classifier
  • Dense(256) → Dropout(0.5)
  • Dense(128) → Dropout(0.3)
  • Dense(10 logits)
• Loss handled using SparseCategoricalCrossentropy(from_logits=True)

⚙️ Training Strategy

• Input normalization
• Data augmentation (horizontal flip, rotation, zoom)
• Batch Normalization for stable training
• Dropout for regularization
• Early stopping to prevent overfitting

📊 Evaluation & Results

• Best validation accuracy: ~67%
• Small train–validation gap → good generalization
• Performance analyzed using a confusion matrix

• Strong performance on classes like automobile, frog, ship, and truck
• Expected confusion between visually similar classes (cat ↔ dog, deer ↔ horse)
• Confusion matrix used as a diagnostic tool rather than accuracy alone

🛠️ Installation

📌 Key Learnings

• CNNs outperform dense networks for image data
• Correct data pipelines are critical for stable training
• Overfitting must be diagnosed using validation metrics
• Confusion matrices provide deeper insight than accuracy alone
• Regularization and early stopping are essential for generalization

📈 Future Improvements

• Stronger data augmentation (MixUp / CutOut)
• Learning rate scheduling
• Residual connections (ResNet-style blocks)
• Transfer learning with pretrained backbones
• Per-class precision and recall analysis