🎯 Exploring CNN Architectures on CIFAR-10

In this project, I experimented with Convolutional Neural Networks (CNNs) on the CIFAR-10 dataset.
The focus was to analyze how different architectural choices and training strategies impact model performance.

---

🏗️ Network Architecture

The main CNN model consisted of:

• Conv2D (32 filters, kernel size 5×5, ReLU) + BatchNorm
• Conv2D (32 filters, kernel size 3×3, ReLU) + BatchNorm
• MaxPooling2D (2×2)
• Conv2D (68 filters, kernel size 3×3, ReLU) + BatchNorm
• Conv2D (68 filters, kernel size 3×3, ReLU) + BatchNorm
• MaxPooling2D (2×2)
• Flatten → Dense (128 units, ReLU) → Dropout(0.3) → Dense (10 units, Softmax)

---

⏱️ Training Experiments

I tested the model under different setups (50 epochs each):

• ✅ With Pooling Layer → Faster training and higher accuracy.
• ❌ Without Pooling Layer → Slower training, lower accuracy.
• ⚡ Stride Comparison (2 vs 4) in pooling:
  • Stride = 2 → better accuracy and stable training.
  • Stride = 4 → information loss, lower accuracy.
• 📉 Schedulers (Exponential vs OneCycle):
  • Exponential Decay → slower but stable convergence.
  • OneCycle Policy → faster convergence, improved final accuracy.

---

✅ Key Takeaways

• Pooling layers play a critical role in reducing training time and improving accuracy.
• Proper stride selection (e.g., 2 instead of 4) significantly improves performance.
• The OneCycle Scheduler outperformed Exponential Decay in both speed and accuracy.
• CNNs proved to be efficient and effective on CIFAR-10 classification tasks.

---

🔗 Excited to keep exploring Deep Learning and Computer Vision.
I’d love to hear your thoughts and experiences with CNNs on image datasets!