NNN | Number's Neural Network

Welcome to the Number's Neural Network (NNN) project! This is a basic neural network built from scratch to recognize hand-drawn digits (0-9) using a custom dataset. The project includes a Python implementation for training and a TypeScript port for visualization. Let's dive into the world of digits and neural networks!

Check out the Hosted Demo (99% Train Accuracy, 55% Test Accuracy - Overfitted 🫠)

Project Structure

• main.py: Contains the neural network architecture and functions to load the pre-trained model.
• dataset: Directory containing the custom dataset of hand-drawn digits.
• nnn-web: Directory containing the TypeScript port for visualizing the neural network.

Getting Started

Neural Network Architecture Overview

Our neural network is a multi-class classification model with one-hot encoding. Here's a quick rundown of the architecture:

• Input Layer: 100 neurons (because the grid I'm using is 10*10 | make sure ur dataset is compatible with this)
• Hidden Layers: Three hidden layers with 20, 20, and 10 neurons respectively. We like to keep things interesting.
• Output Layer: 10 neurons, one for each digit (0-9). We use the softmax activation function here to ensure our outputs sum up to 1 (because math is fun).
• Cost Function: Categorical Cross-Entropy (CCE). It's like a harsh teacher that penalizes wrong answers more severely.

Python Implementation

1. Setup the Dataset Path: Ensure the dataset is placed in the dataset directory (mine is already there).

2. Define Neural Network Architecture: The architecture is defined in main.py:

   layer_1 = 20
   layer_2 = 20
   layer_3 = 10
   output_layer = 10
   n = [100, layer_1, layer_2, layer_3, output_layer]

3. Load Pre-trained Model: Use the load_model function to load some pre-trained model parameters (if u save one):

   def load_model(file_name):
       global W1, W2, W3, W4, b1, b2, b3, b4
       try:
           with open(file_name, 'rb') as file:
               model_params = pickle.load(file)

4. Train and make predictions: Use the train and predict functions in main.py to use the use the NN.

5. Have fun with the NN: Use the NNN Colab Notebook to easily play around and explore the implementation.

TypeScript Visualization

The TypeScript port available in the nnn-web directory provides a visualization of the neural network. Because who doesn't love a good visual?

1. Navigate to the nnn-web directory:

   cd nnn-web

2. Install Dependencies:

   npm install

3. Run the Development Server:

   npm run dev

4. Open the Application: Open http://localhost:3000 in your browser to see the visualization.

For more details on the TypeScript visualization, refer to the NNN_WEB README.

Conclusion

That's it! You've got a neural network that can recognize hand-drawn digits and a cool visualization to show off to your friends. Happy coding!

P.S. If your neural network starts talking to you, it might be time to take a break. Just saying.