🧠 MNIST CNN Model with Hyperparameter Tuning

This project demonstrates a complete deep learning workflow for classifying handwritten digits from the MNIST dataset using a Convolutional Neural Network (CNN).
It covers every stage — from data preprocessing to hyperparameter optimization — to achieve a high-performing and well-regularized model.

---

🚀 Project Overview

The objective of this notebook is to build, train, evaluate, and optimize a CNN model capable of recognizing handwritten digits (0–9) from grayscale 28×28 pixel images.
Through systematic training and tuning, the model’s performance is enhanced for better accuracy and generalization.

---

🧩 Steps Performed

1. Load the MNIST Dataset

• Loaded using TensorFlow’s built-in tf.keras.datasets.mnist.
• Split into 60,000 training and 10,000 testing images.

2. Data Preprocessing

• Normalized pixel values to the range [0, 1].
• Reshaped images to (28, 28, 1) to fit CNN input dimensions.
• Converted labels into integer-encoded classes.

3. Build the CNN Model

• Defined a baseline CNN using tensorflow.keras.Sequential.
• Layers included:
  • Conv2D and MaxPooling2D for feature extraction.
  • Flatten and Dense for classification.
• Designed to balance simplicity, performance, and interpretability.

4. Model Compilation

• Optimizer: Adam
• Loss Function: Sparse Categorical Crossentropy
• Metric: Accuracy

5. Train and Validate the Model

• Trained the model on the training set and validated on the test set.
• Visualized accuracy and loss trends across epochs.

6. Evaluate and Visualize Results

• Evaluated model performance on unseen test data.
• Displayed predictions vs. actual labels for sample images.
• Provided insights into strengths and misclassifications.

---

🔍 Hyperparameter Tuning

To further improve performance, Keras Tuner’s Random Search was used to explore various configurations of the CNN model.

Tuning Strategy

• Defined a hyperparameter search space for:
  • Number of filters and kernel sizes in Conv2D layers
  • Learning rate of the optimizer
  • Dropout rate for regularization
  • Batch size and dense layer units
• Added Batch Normalization and Dropout layers for better regularization.
• Introduced an additional Conv2D layer to explore deeper networks.

Tuning Process

• Implemented a modular function build_cnn_model(hp) to dynamically construct models based on hyperparameters.
• Used RandomSearch to identify the optimal configuration based on validation accuracy.

Best Model Selection

• Retrieved best hyperparameters from the tuner.
• Rebuilt and retrained the CNN using the optimal configuration.
• Evaluated on the full training and test datasets.

---

🧾 Results and Insights

• The tuned CNN achieved significant improvement in accuracy and generalization.
• Both models were compared in terms of training behavior and evaluation metrics.

Key Findings:

• Regularization (Dropout, BatchNorm) effectively reduced overfitting.
• Learning rate tuning stabilized convergence.
• Model depth enhanced feature extraction without vanishing gradients.

---

🧮 Technologies Used

• 🐍 Python
• 🧠 TensorFlow / Keras
• 🎯 Keras Tuner
• 📊 NumPy, Matplotlib, Seaborn for data analysis and visualization

---

📊 Performance Summary

Model Version	Test Accuracy	Validation Loss	Key Features
Baseline CNN	~98%	Moderate	Basic CNN architecture
Tuned CNN	>99%	Lower	BatchNorm, Dropout, optimized hyperparams

---

🧠 Key Learnings

• Systematic hyperparameter tuning greatly boosts CNN performance.
• Regularization is critical for preventing overfitting.
• Visualization helps identify convergence issues and failure patterns.
• The MNIST dataset remains a powerful benchmark for experimentation.

---

📌 Future Enhancements

• Extend to Fashion-MNIST or CIFAR-10 for more complex tasks.
• Use Bayesian Optimization or Hyperband for efficient tuning.
• Visualize feature maps to interpret learned representations.
• Apply transfer learning or quantization for deployment efficiency.

---

---

🏁 Conclusion

This project showcases the complete deep learning pipeline — from building a baseline CNN to performing rigorous hyperparameter optimization.
The final tuned model delivers high accuracy on the MNIST dataset and serves as a foundation for more advanced computer vision research.

---

📜 Author

ARIF RABBANI
Software Engineering Student | Machine Learning Enthusiast