Project 1: Handwritten Digit Recognition (MNIST)

Build a neural network to classify handwritten digits (0-9).

Starter code: Run starter.py in this folder (MNIST loads via Keras when implemented).

Difficulty

Intermediate

Time Estimate

3-5 days

Skills You'll Practice

• Neural Networks
• Image Processing
• Deep Learning Basics
• Hyperparameter Tuning

Learning Objectives

By completing this project, you will learn to:

• Build and train neural networks
• Preprocess image data
• Tune hyperparameters effectively
• Evaluate deep learning models
• Use data augmentation
• Achieve high accuracy on a held-out test set (often above 98% with a strong model. Treat any target as an example, not a guarantee)

Dataset

MNIST Dataset

• Built into Keras/TensorFlow: keras.datasets.mnist
• 60,000 training images, 10,000 test images
• 28x28 grayscale images of handwritten digits (0-9)

Project Steps

Step 1: Load and Explore Data

• Load MNIST dataset
• Visualize sample images
• Check data distribution
• Normalize pixel values

Step 2: Build Neural Network

• Create simple MLP (Multi-Layer Perceptron)
• Add hidden layers
• Choose activation functions
• Initialize model

Step 3: Train Model

• Compile model (loss, optimizer, metrics)
• Train on training set
• Monitor training progress
• Use validation split

Step 4: Evaluate Model

• Evaluate on test set
• Visualize predictions
• Analyze misclassified examples
• Calculate accuracy

Step 5: Improve Model

• Add more layers
• Try different architectures
• Use dropout for regularization
• Tune hyperparameters
• Use data augmentation

Step 6: Advanced Techniques

• Try CNN (Convolutional Neural Network)
• Compare MLP vs CNN
• Achieve >99% accuracy
• Visualize learned features

Expected Deliverables

1. Jupyter Notebook with:

   • Data exploration
   • Model architecture
   • Training process
   • Evaluation results
   • Comparison of different models

2. Trained Model:

   • Saved model file
   • Can make predictions on new images

Evaluation Metrics

• Accuracy: Primary metric (aim for >99%)
• Confusion Matrix, see which digits are confused
• Per-class Accuracy: Accuracy for each digit

Tips

• Start with simple MLP, then try CNN
• Normalize pixel values to [0, 1]: Use early stopping to prevent overfitting
• Try different optimizers (Adam, SGD)
• Data augmentation can help improve accuracy
• Visualize misclassified examples to understand errors

Resources

• MNIST Dataset
• Keras MNIST Tutorial
• TensorFlow MNIST Guide

Extensions

• Try different architectures (ResNet, etc.)
• Build web app to upload and classify images
• Try on other digit datasets (USPS, SVHN)
• Implement from scratch (without frameworks)

Next Steps

After completing this project:

• Move to CIFAR-10 for color images
• Try more complex architectures
• Experiment with transfer learning
• Move to Project 2: Customer Churn