🧠 Neural Network Digit Recognizer (NumPy Only)

This project implements a neural network for digit classification using the MNIST dataset. The entire training pipeline — from architecture and forward propagation to backpropagation and optimization — is developed from scratch using only the NumPy library.

---

📌 Overview

• Dataset: MNIST (loaded via torchvision.datasets)
• Implementation: Python, NumPy (no high-level ML frameworks)
• Accuracy: ~82–91% on 75 unseen test samples
• Architecture: 784–20–10 fully connected layers

---

⚙️ Features

• Manual Feedforward & Backpropagation

  • Custom matrix-based forward propagation
  • Hand-coded backward pass using gradient descent with Adam optimizer

• Optimizations

  • L2 regularization
  • Exponential learning rate decay
  • Early stopping based on validation cost variance
  • Weight & bias checkpointing for best-performing epoch

• Initialization

  • Xavier initialization for hidden layers
  • He initialization for output layer

• Activation & Output

  • Leaky ReLU for hidden layers
  • Softmax for final classification

• Visualization

  • Cost curves for training, validation, and test sets
  • Tracking of variance to identify overfitting
  • Printout of final accuracies per dataset

---

📊 Performance

• Training Accuracy: Varies across epochs; up to ~90%
• Validation Accuracy: Tracked to prevent overfitting
• Test Accuracy: ~82–91% on 75 random MNIST digits

---

📁 Project Structure

digit-recognizer-numpy/ ├── digit_recognizer.py # Full training + testing pipeline with hardcoded neural network ├── README.md # Project documentation └── /data # Automatically created folder by torchvision for MNIST dataset

---

▶️ How to Run

git clone https://github.com/aprotsenko24/digit-recognizer-numpy.git
cd digit-recognizer-numpy
pip install numpy matplotlib torchvision
python digit_recognizer.py

---

Let me know if you'd like to include example plots (like loss curves) or add a section on limitations and future improvements!