A modular Python package that implements and compares three neural network approaches for handwritten digit classification on the MNIST dataset. Includes comprehensive evaluation, sensitivity analysis, and visualization tools. Achieves 99.71% accuracy with the Deep CNN model.

• Three model architectures with different accuracy/speed tradeoffs
• Feature selection pipeline reducing input from 784 to 196 dimensions
• Comprehensive evaluation — confusion matrices, classification reports, error analysis
• Sensitivity analysis — gradient importance, perturbation robustness, occlusion maps
• Interactive demo notebook for hands-on experimentation

graph TB
    subgraph Input
        A[MNIST 28x28 Images]
    end

    subgraph "Basic MLP"
        B1[Flatten] --> B2[Dense 128/256/512] --> B3[Softmax 10]
    end

    subgraph "Optimized Model"
        C1[Feature Selection 784→196] --> C2[Dense 256 + Dropout] --> C3[Dense 128 + Dropout] --> C4[Softmax 10]
    end

    subgraph "Deep CNN"
        D1[3x Conv2D Blocks + BatchNorm] --> D2[Dense 512 → 256] --> D3[Softmax 10]
    end

    A --> B1
    A --> C1
    A --> D1

mnist-handwritten-digit-recognition-project/
├── src/mnist_recognition/       # Python package
│   ├── config.py                # Hyperparameters
│   ├── data/                    # Loading, preprocessing, augmentation, feature selection
│   ├── models/                  # Basic MLP, Optimized, Deep CNN
│   ├── evaluation/              # Metrics, error analysis, report generation
│   └── visualization/           # Training curves, comparisons, sensitivity maps
├── notebooks/
│   └── demo.ipynb               # Interactive demo notebook
├── pyproject.toml               # Package configuration
├── requirements.txt             # Dependencies
├── CONTRIBUTING.md              # Contribution guidelines
├── CODE_OF_CONDUCT.md           # Community guidelines
└── LICENSE                      # MIT License

• Python 3.8+
• pip

GPU support is optional but recommended for training.

# Clone the repository
git clone https://github.com/ChanMeng666/mnist-handwritten-digit-recognition-project.git
cd mnist-handwritten-digit-recognition-project

# Create a virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install the package
pip install -e .

Run the demo notebook for a guided walkthrough:

jupyter notebook notebooks/demo.ipynb

Or use the package directly in Python:

from mnist_recognition.data.loader import load_mnist
from mnist_recognition.data.preprocessing import preprocess_data
from mnist_recognition.data.augmentation import create_data_generators
from mnist_recognition.models.deep_cnn import create_and_train_deep_model
from mnist_recognition.config import DEFAULT_CONFIG

# Load and preprocess
(x_train_raw, y_train_raw), (x_test_raw, y_test_raw) = load_mnist()
(x_train, y_train), (x_val, y_val), (x_test, y_test) = preprocess_data(
    x_train_raw, x_test_raw, y_train_raw, y_test_raw
)

# Create data generators
config = DEFAULT_CONFIG.copy()
train_gen, val_gen = create_data_generators(
    x_train, y_train, x_val, y_val, config["batch_size"]
)

# Train the Deep CNN model
model, results = create_and_train_deep_model(config, train_gen, val_gen)

A multi-layer perceptron that flattens the 28x28 image and passes it through a single dense hidden layer. Tests three configurations (128, 256, 512 neurons) and selects the best.

Uses ANOVA F-value feature selection to reduce input dimensionality from 784 to 196 features, then trains a compact two-layer network with dropout. Achieves a 79% reduction in parameters while maintaining good accuracy.

Three convolutional blocks, each with two Conv2D layers, batch normalization, ReLU activation, max pooling, and dropout. Followed by dense layers (512 → 256 → 10). Uses gradient clipping and learning rate scheduling.

Contributions are welcome! See CONTRIBUTING.md for guidelines.

This project is licensed under the MIT License.

Chan Meng

• GitHub: @ChanMeng666
• LinkedIn: chanmeng666
• Website: chanmeng.org

Chan Meng
Need a custom app like this one? I build them — let's talk.