🧠 MNIST Classification with Fully Connected Neural Network (FCNN)

---

📈 Live Results

You can view the notebook with all outputs and results on Kaggle: https://www.kaggle.com/code/evangelosgakias/fcnn-image-classification-tensorflow

---

📑 Table of Contents

• Overview
• Project Structure
• Features
• Usage
• Results
• Sample Visualizations
• Future Improvements
• Contributing
• License
• Contact

---

📋 Overview

This project demonstrates how to build, train, and evaluate a fully connected neural network (Multi-Layer Perceptron, MLP) for image classification using the MNIST dataset. The implementation leverages TensorFlow and Keras to construct a deep learning model that learns to recognize handwritten digits (0–9). The notebook also includes systematic hyperparameter tuning using Keras Tuner to optimize model performance.

• Dataset: MNIST (60,000 training images, 10,000 test images, 28x28 grayscale, 10 classes)
• Goal: Classify handwritten digits with high accuracy using a robust, regularized MLP
• Skills Showcased: Data preprocessing, model design, training, evaluation, hyperparameter tuning, visualization, and analysis

---

🏗️ Project Structure

.
├── FNN.ipynb         # Jupyter notebook with the complete implementation
├── requirements.txt  # Python dependencies
├── LICENSE           # MIT License
├── README.md         # Project documentation (this file)
└── hyperparameter_tuning/
    └── mnist_tuning/ # Keras Tuner search results

---

🚀 Features

• Data Preparation:
  • Automatic download and loading of the MNIST dataset
  • Normalization of pixel values to [0, 1]
  • Train/validation/test split (80%/20%/test)
  • Visualization of sample images
• Model Architecture:
  • Input flattening (28x28 → 784)
  • Two dense layers with ReLU activation and L2 regularization
  • Dropout layers for regularization
  • Output layer with softmax activation for multi-class classification
• Training Process:
  • Adam optimizer, sparse categorical cross-entropy loss
  • Training with validation monitoring
  • Visualization of accuracy and loss curves
• Evaluation & Visualization:
  • Metrics: Accuracy, Loss
  • Confusion matrix
  • Sample predictions with true vs. predicted labels
  • Visualization of misclassified examples
• Hyperparameter Tuning:
  • Keras Tuner for dense units, dropout rates, L2 values, and optimizer
  • Comparison of original and fine-tuned models
  • Visual and tabular performance comparison

---

⚡ Usage

Local Setup

1. Clone the repository:

   git clone https://github.com/EvanGks/mnist-digit-classification-fcnn.git
   cd mnist-digit-classification-fcnn

2. Create and activate a virtual environment:
   • On Windows:

     python -m venv .venv
     .venv\Scripts\activate

   • On macOS/Linux:

     python3 -m venv .venv
     source .venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

4. Open the Jupyter Notebook:

   jupyter notebook FNN.ipynb

5. Run the notebook cells in order:
   • The notebook is organized sequentially with explanations and visualizations.
   • All outputs and plots will be displayed inline.

---

📊 Results

The model achieves the following performance (see Kaggle notebook for full details):

Model	Test Accuracy	Test Loss
Original Model	~98%	~0.07
Fine-Tuned Model	~98.2%	~0.06

• Validation accuracy and loss closely track training metrics, indicating good generalization.
• Confusion matrices show most predictions are correct, with errors concentrated in visually ambiguous digits.
• Fine-tuned model offers a slight but consistent improvement over the baseline.

---

🖼️ Sample Visualizations

• Training and Validation Accuracy/Loss Curves:
  • Show convergence and help detect overfitting.
• Confusion Matrix:
  • Visualizes correct vs. incorrect predictions for each digit.
• Sample Predictions:
  • Grid of test images with true and predicted labels, color-coded for correctness.
• Misclassified Examples:
  • Highlights challenging cases for the model, useful for error analysis.

See the notebook and Kaggle live version for all plots and outputs.

---

🛠️ Future Improvements

• Experiment with deeper or alternative architectures (e.g., CNNs)
• Apply data augmentation to increase robustness
• Explore additional regularization (early stopping, L1/L2 variants)
• Try different optimizers and learning rate schedules
• Integrate TensorBoard for richer training visualization
• Expand to other datasets or multi-task learning

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, open an issue first to discuss what you would like to change.

---

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

📬 Contact

For questions or feedback, please reach out via:

• GitHub: EvanGks
• X (Twitter): @Evan6471133782
• LinkedIn: Evangelos Gakias
• Kaggle: evangelosgakias
• Email: evangks88@gmail.com

---

Happy Coding! 🚀