This repository implements a neural network for the classification of handwritten digits. The neural network is designed to work with the MNIST dataset and is capable of recognizing and classifying digits from 0 to 9 with high accuracy. It is designed to be flexible and easy to use, with a focus on understanding and implementing the fundamental concepts of neural networks

The neural network is implemented in the neural_network.py file. It includes backpropagation, feedforward, cost calculation, and training functions.

• main.py: The main script that trains and tests the neural network models with different configurations and selects the best model based on test accuracy.

• neural_network.py: This file contains the implementation of the Neural Network model.

• data.py: Contains functions to load and preprocess the MNIST dataset for training and testing.

• utils.py: Utility functions for activation functions (ReLU, sigmoid, tanh), their derivatives, and parameter initialization.

To run this project, you will need to have Python and the following libraries installed:

• NumPy
• Matplotlib
• Keras (for loading the MNIST dataset)

Install these dependencies using pip:

pip install numpy matplotlib keras

• project was built with Python 3.11.0

clone this repository to your local machine:

git clone https://github.com/yaniv-simmer/machine-learning-neural-networks-exercises.git 

Run the main.py script to train and test different neural network configurations. The script will print the best model's accuracy and display a loss plot.

python main.py

You can customize the neural network's configuration by modifying the following parameters in main.py:

• hidden_layer_dimensions_lst: List of hidden layer dimensions.

• activation_functions_lst: List of activation functions for each layer.

• derivative_functions_lst: List of derivative functions for each layer.

• training_iterations: Number of training iterations.

• learning_rate: Learning rate.

• regularization_coefficient_lst: List of regularization coefficients.

The best model's configuration and test accuracy are printed to the console. The loss during training is also visualized with a loss plot.

Contributions, bug reports, and feature requests are welcome. Feel free to open an issue or submit a pull request.

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