MNIST Neural Network From Scratch

---

This project is a complete implementation of a neural network for MNSIT digit recognition, built entirely from scratch without using any deep learning frameworks.

It combines Core Neural Network Maths, V****ectorized Numerical Computing, and a User Interface that enables real-time digit drawing, prediction visualization, and dataset augmentation through user feedback.

The goal of this project is to demonstrate a strong understanding of Neural Network internals, Backpropagation, and End-to-End Machine Learning System Design, rather than relying on high-level libraries.

---

Project Overview

• Fully implemented neural network using NumPy
• Manual forward and backward propagation
• Interactive training, testing, and inference workflow
• Human-in-the-loop dataset expansion
• Persistent model parameters with resume capability

---

Model Architecture

Input Layer

• 784 neurons (28 × 28 pixel grayscale image)

Hidden Layer

• 64 neurons
• ReLU activation
• He initialization

Output Layer

• 10 neurons
• Softmax activation

Loss Function

• Categorical Cross-Entropy

Optimization

• Batch Gradient Descent

---

Key Features

Neural Network from Scratch

• No use of TensorFlow, PyTorch, Keras, or similar frameworks
• Explicit implementation of forward propagation, backpropagation, and parameter updates
• Fully vectorized operations for efficiency and numerical stability

Training and Evaluation

• Configurable number of training steps and learning rate
• Automatic checkpointing of model parameters
• Resume training from saved parameters
• Overall accuracy reporting
• Per-digit accuracy analysis with visualization

Interactive Digit Recognition Interface

• Mouse-based digit drawing on a 28 × 28 grid
• Real-time predictions with class confidence visualization
• Optional heatmap view of pixel intensity
• MNIST-style preprocessing pipeline to ensure compatibility with training data

Human-in-the-Loop Learning

• Users can mark predictions as correct or incorrect
• Incorrect predictions allow manual label correction
• New samples are appended directly to the dataset
• Enables continuous dataset improvement and retraining

---

Digit Preprocessing Pipeline

The drawing interface applies a preprocessing pipeline designed to closely match MNIST data characteristics:

1. Gaussian smoothing
2. Noise thresholding
3. Bounding box extraction
4. Aspect-Ratio preserving resizing
5. Centering within a 28 × 28 canvas
6. Center-of-Mass alignment
7. Normalization to the [0, 1] range

---

Project Structure

.
├── nn/
│   ├── main.py        # Application entry point
│   ├── menu_ui.py     # Menu-Driven GUI
│   ├── ui.py          # Digit drawing and Interface
│   ├── model.py       # Neural Network implementation
│   ├── train.py       # Training Logic
│   ├── test.py        # Evaluation and Visualization
│   └── data.py        # Data loading and preprocessing
├── dataset/
│   ├── data.csv       # Training dataset (MNIST-Style)
│   └── parameters.npz # Saved Model Parameters
├── requirements.txt
└── README.md

---

How to Run

Install dependencies:

pip install -r requirements.txt

Run the application:

python main.py

The menu allows you to:

• Train the neural network
• Test model performance
• Draw digits and observe predictions

---