---

🧠 Neural Network from Scratch — Handwritten Digits Classification

This repository contains a simple neural network built from scratch (using only NumPy) to classify handwritten digits from the scikit-learn Digits Dataset. No high-level libraries like TensorFlow or PyTorch are used — this project is purely for learning the fundamentals of neural networks such as forward propagation, backward propagation, and gradient descent.

---

🚀 Features

• Implements a 2-layer feedforward neural network
• ReLU activation in the hidden layer
• Softmax activation in the output layer
• One-hot encoding for labels
• Manual implementation of gradient descent
• Achieves ~95% accuracy on the test set

---

📂 Project Structure

neural_network_digits/
│
├── neural_network.py   # Main Python file (your code)
├── README.md           # Project documentation
└── requirements.txt    # Python dependencies (optional)

---

⚙️ How It Works

1️⃣ Dataset Preparation

• Loads the digits dataset (8x8 grayscale images, pixel values from 0–16)
• Normalizes features by dividing by 16
• Splits into train (80%) and test (20%) sets

2️⃣ Model Architecture

Layer	Type	Units	Activation
Input Layer	Dense	64 (8×8 image)	—
Hidden Layer	Dense	10	ReLU
Output Layer	Dense	10	Softmax

---

3️⃣ Training Process

• Forward Propagation → compute predictions
• Backward Propagation → calculate gradients manually
• Gradient Descent → update weights and biases iteratively

---

🖥️ Usage

1. Clone the Repository

git clone https://github.com/vinayakkamatcodes/neural_netwrk.git
cd neural_netwrk

2. Install Dependencies

pip install numpy scikit-learn

3. Run the Training Script

python neural_network.py

---

📊 Sample Output

--- Starting Training ---
Iteration: 0
Accuracy: 11.25%
Iteration: 50
Accuracy: 72.80%
Iteration: 100
Accuracy: 85.90%
...
--- Training Complete ---

--- Testing on Test Set ---
Test Accuracy: 95.40%

---

🧠 Key Learnings

• Matrix multiplication for forward propagation
• Chain rule for backpropagation
• Gradient updates and learning rate tuning
• How neural networks "learn" from data

---

📌 Future Improvements

• Add more hidden layers for deeper architecture
• Implement mini-batch gradient descent
• Add regularization (L2 or dropout) to prevent overfitting
• Visualize loss vs. iterations

---