MNIST Classifier from Scratch (NumPy + Custom Backpropagation)

This project implements a neural network from scratch using only NumPy, trained on the MNIST handwritten digits dataset. It does not use any high-level deep learning frameworks like Keras or PyTorch for the model logic — everything is built manually to learn the internals of how neural networks work.

---

Key Features

• Built using only NumPy (except for data loading via TensorFlow/Keras)
• Implements:
  • Fully connected layers
  • ReLU activation
  • Softmax output
  • Cross-entropy loss (with optional label smoothing)
  • Manual backpropagation
• Trains and evaluates on the MNIST dataset
• Includes both Jupyter Notebook and .py script versions
• Clean and modular OOP structure

---

Files in This Repo

File	Description
mnist_numpy_nn.ipynb	Interactive notebook version with code + output
mnist_numpy_nn.py	Standalone script version (for command-line execution)
README.md	This readme
LICENSE	MIT license (open-source)

---

How It Works

1. Data Loading:
   Uses tf.keras.datasets.mnist.load_data() to fetch and normalize the dataset.

2. Model Architecture:

   • Input layer: 784 units (28×28 flattened)
   • Hidden layer: 128 units, ReLU activation
   • Output layer: 10 units (Softmax for classification)

3. Training:

   • Mini-batch gradient descent
   • Manual weight updates via gradients
   • Label smoothing is optionally applied for better generalization

4. Evaluation:

   • Accuracy is computed on both the validation and test datasets

---

Install using:

pip install numpy tensorflow

sample output : - Validation Accuracy: 92.34% Test Accuracy: 92.19%