CNN

🖼️ MNIST Digit Classification with Convolutional Neural Network (CNN)

This project builds a Convolutional Neural Network (CNN) using TensorFlow/Keras to classify handwritten digits from the MNIST dataset. CNNs are the go-to architecture for image data — they learn spatial patterns using filters that slide across the image.

---

📁 Dataset Overview

The MNIST dataset contains 70,000 grayscale images of handwritten digits (0-9):

• Training set: 60,000 images
• Test set: 10,000 images
• Each image is 28x28 pixels (grayscale)
• 10 classes: digits 0 through 9

---

🎯 Objective

To build a CNN that learns spatial features (edges, curves, shapes) from handwritten digit images, achieving higher accuracy than a plain feedforward network by leveraging the 2D structure of the data.

---

🧠 Why CNN?

• CNNs preserve spatial structure — unlike ANNs that flatten images and lose spatial info
• Convolutional layers learn local patterns: edges, curves, loops — the building blocks of digits
• Pooling layers reduce dimensionality while keeping important features
• CNNs achieve >99% accuracy on MNIST — state-of-the-art for simple image classification

---

🧪 Tools & Libraries Used

• Python
• Jupyter Notebook
• pandas, numpy for data handling
• matplotlib, seaborn for visualization
• tensorflow / keras for building and training the CNN

---

📊 Model Architecture

• Conv2D Layer 1: 32 filters, 3x3 kernel, ReLU activation
• MaxPooling2D: 2x2 pool size
• Conv2D Layer 2: 64 filters, 3x3 kernel, ReLU activation
• MaxPooling2D: 2x2 pool size
• Flatten: Convert 2D feature maps to 1D
• Dense Layer: 128 neurons, ReLU activation
• Dropout: 30% for regularization
• Output Layer: 10 neurons, Softmax activation

Total parameters: ~100,000

---

📊 Process Overview

1. Data Loading & Exploration:

   • Loaded MNIST from tf.keras.datasets
   • Visualized sample digits with their labels
   • Checked image dimensions and class balance

2. Data Preprocessing:

   • Reshaped images to (28, 28, 1) — adding channel dimension
   • Normalized pixel values to [0, 1]
   • One-hot encoded labels

3. Modeling:

   • Built CNN with Conv2D + MaxPooling2D layers
   • Used categorical crossentropy loss and Adam optimizer
   • Trained for 10 epochs

4. Evaluation & Visualization:

   • Evaluated accuracy and loss on test set
   • Plotted training vs validation curves
   • Visualized correct and incorrect predictions

---

📌 Key Results

• CNN achieved >99% accuracy on MNIST — significantly better than a plain ANN
• The model learned meaningful filters: edge detectors, curve detectors, etc.
• Training converged quickly — fewer epochs needed than ANN
• Fewer parameters than the ANN but better accuracy — showing the power of convolutions

---

📂 Folder Structure

ML-DL-Projects/
├── Deep-Learning/
│   └── CNN/
│       ├── dataset/           # (MNIST loaded via tf.keras)
│       ├── images/
│       │   ├── training_history.png
│       │   └── sample_predictions.png
│       ├── notebooks/
│       │   └── CNN_Example.ipynb
│       └── README.md
├── uv.lock
├── pyproject.toml

---

🚀 Getting Started

To run this project locally:

Step 1: Set up environment

cd ML-DL-Projects
uv sync

Step 2: Run notebook

jupyter notebook

Then open CNN_Example.ipynb.

---

💡 Future Ideas

• Add more convolutional layers (deeper network) and see how it affects accuracy
• Experiment with different filter sizes and numbers
• Use data augmentation (rotation, zoom, shift) to improve robustness
• Apply the same CNN architecture to Fashion MNIST for a harder challenge

---

📬 Contact

Feel free to reach out if you have questions or want to collaborate!