Skibidi Network: A From-Scratch Neural Network for MNIST Digit Recognition 🤖✨

This repository presents an implementation of a simple neural network built entirely from scratch using Python and NumPy. The network is designed to perform classification on the MNIST dataset, which consists of 28×28 pixel grayscale images of handwritten digits.

Description

The neural network in this repository features a straightforward architecture with the following components:

• Input Layer:
  Accepts 784 features corresponding to the flattened 28×28 pixel images.

• Hidden Layer:
  Consists of 64 neurons. Each neuron applies the hyperbolic tangent (tanh) activation function. Weights are initialized randomly from the interval ([-0.5, 0.5]) and each neuron receives a bias of 0.5.

• Output Layer:
  Comprises 10 neurons, one for each digit from 0 to 9. The tanh activation function is applied at this layer, with a bias of 0.7 assigned to each neuron.

Note: Although using the tanh activation function and Mean Squared Error (MSE) loss is not the best approach for digit classification (a softmax activation with cross-entropy loss is generally preferred), we are using them here solely for educational purposes. 🎓

Features

• From-Scratch Implementation:
  Demonstrates the fundamental principles of neural network design without relying on high-level machine learning libraries.

• Custom Weight Initialization:
  Weights are randomly initialized within a specified interval, and biases are set to predetermined values.

• Forward and Backward Propagation:
  Implements the forward pass with tanh activation and backpropagation using the derivative of tanh.

• MNIST Digit Recognition:
  Trains the network on the MNIST dataset for the purpose of digit classification.

• Visualization Tools:
  Includes code to plot the training loss over epochs and to visualize sample predictions on test images.

Libraries Used

• TensorFlow:
  Utilized exclusively for loading the MNIST dataset.

• NumPy:
  Employed for matrix operations and numerical computations throughout the network.

• Matplotlib:
  Used for creating visualizations of the training loss and sample predictions.

Requirements

• Python 3.x
• NumPy
• Matplotlib
• TensorFlow (only for loading the MNIST dataset)

Installation

Clone the repository and navigate to the project directory:

git clone https://github.com/a1harfoush/NN-from-scratch.git
cd NN-from-scratch

Install the required dependencies:

pip install numpy matplotlib tensorflow

Usage

Run the main script to train the neural network and visualize the results:

python nn_mnist.py

Upon execution, the script will:

• Train the network on the MNIST dataset.
• Output the training progress and final test accuracy.
• Plot the training loss over epochs 📈.
• Display a selection of test images with their predicted and true labels.

Results & Visualization

After training, you will see:

• Test Accuracy:
  The script will output the overall test accuracy (e.g., Test accuracy: 0.76). This provides a quantitative measure of the network's performance.

• Visualization:

  • A Training Loss Plot that shows how the loss decreases over epochs.

    

  • Sample Predictions: A visual gallery of selected test images alongside their true and predicted labels.

    

Feel free to update this section with your own results and screenshots once you've run the network!

Acknowledgements

• The MNIST dataset is provided by Yann LeCun and collaborators.
• This project is intended for educational purposes, illustrating the fundamental concepts of neural network implementation and training.