MNIST Handwritten Digits Classification

This project implements a simple neural network for classifying handwritten digits from the MNIST dataset. The dataset contains grayscale images of digits (0-9), and the model predicts the correct digit.

---

Docker Implementation with Prometheus and Grafana

I have also made a Docker implementation of the same, along with Prometheus and Grafana for monitoring and visualization. You can view those results here - https://github.com/Torajabu/multiclass-classifcation-simple-neural-network--mnist-handwritten-dataset/tree/main/docker%2C%20Prometheus%2C%20Grafana%20results and https://drive.google.com/file/d/1YsrsR8fbUTyxVbacPidSqIoNr2YBY2O9/view?usp=sharing

Steps to Use:

1. Pull the Docker Image:

   docker pull rajab97/neural_networks
   
   
   

Features

• Dataset: MNIST Handwritten Digits (normalized to [0, 1]).
• Model Architecture:
  • Input Layer: 784 neurons (flattened 28x28 images).
  • Hidden Layers:
    • Layer 1: 25 neurons with ReLU activation.
    • Layer 2: 15 neurons with ReLU activation.
  • Output Layer: 10 neurons (one for each digit).
• Loss Function: Sparse Categorical Crossentropy.
• Optimizer: Adam optimizer.
• Metrics: Accuracy.
• Visualizations:
  • Training and validation loss plots.
  • Predicted vs. actual labels for random images.
  • Error analysis.

---

How to Run

1. Clone the repository:

   git clone https://github.com/Torajabu/multiclass-classifcation-simple-neural-network--mnist-handwritten-dataset.git
   

""" Navigate to the project directory: cd multiclass-classifcation-simple-neural-network--mnist-handwritten-dataset

Install dependencies:

pip install tensorflow numpy matplotlib sklearn

Run the script:

python mnist_neural_network.py

Results

Model Performance:
    Training and validation loss reduce consistently over epochs.
    Random predictions on MNIST images with correct/incorrect classification are visualized.
Misclassified Images:
    The total number of errors is displayed along with the incorrectly classified images.

For detailed results, check the Results Folder. Notes

Acknowledgment: The MNIST dataset is publicly available and was created by Yann LeCun and colleagues. For more information, visit MNIST.
Model Customization:
    Add more layers or neurons to improve performance.
    Experiment with learning rates or activation functions (tanh, etc.).
    Use dropout layers to prevent overfitting