🔢 MNIST Handwritten Digit Recognition using Deep Learning

A complete Deep Learning project for handwritten digit recognition using the MNIST dataset. The project covers the entire workflow, from dataset exploration to model optimization with Keras Tuner, concluding with an interactive visualization of the trained neural network.

---

🎬 Neural Network Animation

The animation illustrates the complete inference pipeline:

• Real class highlighted in green
• Predicted class highlighted in cyan
• Hidden layers represented in red
• Prediction confidence
• Loss and Accuracy curves
• Classification probabilities

---

📌 Project Overview

The objective of this project is to demonstrate the complete development cycle of a Convolutional Neural Network (CNN) for handwritten digit recognition.

The project includes:

• Dataset exploration
• CNN implementation using TensorFlow/Keras
• Performance evaluation
• Data Augmentation
• Hyperparameter optimization using Keras Tuner
• External image classification
• Interactive neural network visualization

---

📂 Repository Structure

.
├── app.py
├── data/
├── images/
├── models/
├── notebooks/
├── requirements.txt
├── src/
└── tuner/

---

📚 Notebooks

Notebook	Description
01_about_the_base.ipynb	MNIST dataset exploration
02_mnist_first_model.ipynb	First CNN model
03_mnist_data_augmentation.ipynb	CNN with Data Augmentation
04_mnist_keras_tuner.ipynb	Hyperparameter optimization using Keras Tuner

---

🧠 Model Architecture

The final model consists of:

• Input Layer (28×28×1)
• Rescaling
• Random Rotation
• Random Translation
• Random Zoom
• Conv2D
• MaxPooling2D
• Dropout
• Conv2D
• MaxPooling2D
• Dropout
• Flatten
• Dense (Softmax)

---

🚀 Technologies

• Python
• TensorFlow
• Keras
• Keras Tuner
• NumPy
• Pandas
• Matplotlib
• Streamlit

---

📈 Training Curves

Loss Function

The training and validation loss decrease consistently throughout the optimization process, indicating stable convergence.

---

Accuracy

The model reaches an accuracy close to 99%, demonstrating excellent generalization capability on unseen samples.

---

🔍 Predictions on MNIST Test Images

The trained model successfully classifies handwritten digits with very high confidence.

---

🔍 Predictions on External Images

The model is also capable of correctly classifying external handwritten digits after proper preprocessing.

---

⚙️ Installation

Clone the repository

git clone https://github.com/your_username/mnist-deep-learning-project.git

Create a virtual environment (optional)

python -m venv .venv

Install the dependencies

pip install -r requirements.txt

---

▶️ Running the Interactive Application

streamlit run app.py

The application displays:

• Training curves
• Input image
• Neural network representation
• Prediction confidence
• Classification probabilities

---

📊 Final Results

Metric	Value
Test Accuracy	98.8%
Test Loss	0.040

---

🎯 Main Concepts Covered

• Deep Learning
• Convolutional Neural Networks (CNN)
• Data Augmentation
• Hyperparameter Optimization
• TensorFlow/Keras
• Keras Tuner
• Model Evaluation
• Computer Vision
• Streamlit Dashboard
• Neural Network Visualization

---

👨‍💻 Author

Dr. Flavio R. Rusch

Data Scientist | Machine Learning | Deep Learning | Statistical Physics | Computational Modeling

GitHub: https://github.com/ruschh

LinkedIn: (add your profile here)

---

⭐ If you found this project useful, consider giving it a star.x