Neural Network Digits Classifier

This repository contains my implementation of a neural network for classifying handwritten digits, built as part of my learning journey through neural networks and deep learning.

Table of Contents

1. Project Overview
2. Learning Resources
3. Installation
4. Usage
5. Model Overview
6. Results

Project Overview

The goal of this project is to build a simple neural network from scratch, capable of classifying handwritten digits (0-9) from the popular MNIST dataset. The project was inspired by a couple of excellent resources that guided me through the theoretical and practical aspects of neural networks.

Features:

• Implementation of a basic feedforward neural network.
• Backpropagation for training.
• The network is trained to classify digits from the MNIST dataset.
• No external deep learning libraries (like TensorFlow or PyTorch) were used — the implementation was done from scratch using NumPy.
• Option to save & load model

Learning Resources

I used the following resources to guide me through understanding and building neural networks:

1. Book: Neural Networks and Deep Learning by Michael Nielsen

   • This book was instrumental in understanding the mathematics and concepts behind neural networks and their implementation from scratch.

2. Video Series: 3Blue1Brown's YouTube Playlist on Neural Networks

   • This series visually explains neural networks and backpropagation in an intuitive way, helping me understand the logic and flow of a neural network.

Installation

To run this project locally, you'll need Python and a few essential libraries:

Prerequisites

• Python 3.x
• numpy
• matplotlib (optional, for visualizing results)

Installation

1. Clone the repository:

   git clone https://github.com/wojciechtrapkowski/digits_classifier
   cd neural-network-digits-classifier

2. Install the required libraries:

   pip install numpy matplotlib

Usage

1. Run program:

   python3 main.py

Model Overview

Architecture:

• Input Layer: 784 neurons (28x28 pixel input images).
• Hidden Layer: Configurable, but typically a layer with 30 neurons.
• Output Layer: 10 neurons (one for each digit 0-9).

Training:

• The network is trained using stochastic gradient descent with backpropagation.
• Activation function used: Sigmoid for neurons.

Results

After training for several epochs, the model achieves approximately 95% accuracy on the MNIST test set.