https://github.com/noahfeldt/ann-from-scratch

Implementation of a multi layer perceptron artificial neural network from scratch that is tested using the MNIST dataset.
https://github.com/noahfeldt/ann-from-scratch

artificial-intelligence artificial-neural-networks image-classification machine-learning ml mlp mlp-classifier mnist mnist-classification mnist-dataset mnist-handwriting-recognition multi-layer-perceptron neural-network nn numpy python reinforcement-learning

Last synced: about 1 month ago
JSON representation

Implementation of a multi layer perceptron artificial neural network from scratch that is tested using the MNIST dataset.

• Host: GitHub
• URL: https://github.com/noahfeldt/ann-from-scratch
• Owner: NoahFeldt
• License: mit
• Created: 2023-08-13T17:06:06.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-09-28T19:14:45.000Z (about 1 year ago)
• Last Synced: 2024-01-26T09:16:56.879Z (11 months ago)
• Topics: artificial-intelligence, artificial-neural-networks, image-classification, machine-learning, ml, mlp, mlp-classifier, mnist, mnist-classification, mnist-dataset, mnist-handwriting-recognition, multi-layer-perceptron, neural-network, nn, numpy, python, reinforcement-learning
• Language: Python
• Homepage:
• Size: 380 KB
• Stars: 2
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# ANN-From-Scratch

Implementation of a multi layer perceptron artificial neural network from scratch that is tested using the MNIST dataset.

## Prerequisites

The [NumPy](https://github.com/numpy/numpy) module is used for numerical vector and matrix calculations:

```bash

pip install numpy

```

The [Keras](https://github.com/keras-team/keras) module is used to import the MNIST dataset:

```bash

pip install keras

```

The [tqdm](https://github.com/tqdm/tqdm) module is used for creating the progress bar:

```bash

pip install tqdm

```

## Design choices 

The neural network implementation uses the following design choices:

* Sigmoid activation function.

* Xavier Glorot initialization of the weights.

* Zero initialization of the biases.

* Mean squared error cost function.

## Usage

The neural network implementation, can be found in the `ann.py` module where the `NeuralNetwork` class exists.

To test the neural network on the MNIST dataset, run the `mnist.py` file. This script will train and test a neural network with the given parameters.

## Results

The neural network reaches an accuracy of about 90.5 % on MNIST dataset with the parameters used in the `mnist.py` file.