DigitSuo

$TETSUO on Solana

Contract Address: 8i51XNNpGaKaj4G4nDdmQh95v4FKAxw8mhtaRoKd9tE8

---

A neural network–based digit recognition project using the MNIST dataset.

Installation

1. Clone the Repository

   git clone [email protected]:tetsuo-ai/DigitSuo.git
   cd DigitSuo

2. Install Dependencies

   Ubuntu/Debian:

   sudo apt-get install gcc libncurses-dev zlib1g-dev doxygen graphviz

   Fedora/RHEL:

   sudo dnf install gcc ncurses-devel zlib-devel doxygen graphviz

Building the Project

Use the provided Makefile for the following targets:

• Recognition Interface:

  make

  Compiles source files in src/ into the executable digit_recognition.

• Training Program:

  make train

  Compiles train.c with OpenMP support into the executable train.

• Documentation:

  make docs

  Generates HTML documentation in docs/html/ using Doxygen.

• Clean Build Artifacts:

  make clean

• Delete Debug Log:

  make clean_debug

• Delete Debug Log:

  make clean_docs

Usage

Training the Network

Run the training program:

./train

This will:

• Load and preprocess the MNIST dataset (files: train-images-idx3-ubyte.gz and train-labels-idx1-ubyte.gz).
• Apply data augmentation.
• Train the neural network.
• Save optimized weights to src/weights.h.

Using the Recognition Interface

Run the recognition interface:

./digit_recognition

Controls

• Mouse/Arrow Keys: Draw digits
• Number Keys (0-9): Load example digits
• Enter: Submit for recognition
• C: Clear drawing
• Q: Quit application

Technical Details

Neural Network Architecture

Input Layer (784 neurons)
    │
    │   Dense connections with ReLU activation
    │   Weights initialized using He initialization
    ▼
Hidden Layer (256 neurons)
    │
    │   Dense connections with Softmax activation
    │   Xavier/Glorot initialization
    ▼
Output Layer (10 neurons)

Training Pipeline

1. Data Preprocessing

   • Load the MNIST dataset with zlib decompression.
   • Apply data augmentation:
     • Random rotation (±10°)
     • Random shifts (±5 pixels)
     • Gaussian blur (σ = 0.3)
   • Generate balanced mini-batches.

2. Optimization

   • Mini-batch gradient descent with momentum.
   • Learning rate decay schedule.
   • Early stopping with patience.
   • OpenMP for parallel processing.

Performance Metrics

• Training accuracy: >98% on the MNIST test set.
• Inference latency: <10ms per prediction.
• Memory footprint: ~2MB during inference.
• Training time: ~5 minutes on a modern CPU.

Documentation

Generating Documentation

Generate the documentation using:

make docs

Viewing Documentation

Start the documentation server:

python3 serve_docs.py

Access the documentation at http://localhost:8000.

Project Structure

.
├── Doxyfile                  # Doxygen configuration file
├── doxygen-custom.css        # Custom CSS for documentation
├── Makefile                  # Build configuration
├── README.md                 # Project documentation
├── docs/                     # Documentation directory (generated)
├── serve_docs.py             # Script to serve documentation
├── train-images-idx3-ubyte.gz # MNIST training images (compressed)
├── train-labels-idx1-ubyte.gz # MNIST training labels (compressed)
├── train.c                   # Training program source
└── src/                      # Source code for recognition interface
    ├── main.c
    ├── draw_interface.c
    ├── neural_net.c
    ├── neural_net.h
    ├── draw_interface.h
    ├── utils.c
    ├── utils.h
    └── weights.h

Customization

Adjust neural network parameters in train.c:

#define HIDDEN_SIZE 256      // Number of hidden neurons
#define BATCH_SIZE 64        // Training batch size
#define NUM_EPOCHS 10        // Maximum training epochs
#define BASE_LR 0.1f         // Initial learning rate