MNIST Neural Network in Rust 🦀
A fun side project where I built a neural network from scratch in Rust to recognize handwritten digits!
What is this?
Just a simple 3-layer neural network that learns to read handwritten digits (0-9) from the famous MNIST dataset. No fancy ML libraries - everything coded from scratch!
Input (784) → Hidden (128) → Output (10)
Gets around ~97% accuracy, which is pretty cool for something built from the ground up.
The Data
MNIST is a classic dataset with 70,000 images of handwritten digits:
- 60,000 for training
- 10,000 for testing
- Each image is 28×28 pixels
What the digits look like
Pretty straightforward - just grayscale images of people's handwriting.
Want to see more? Run: python3 visualize_mnist.py
How it works
Basic neural network stuff:
// Push data through the network
hidden = relu(input · weights1 + bias1)
output = softmax(hidden · weights2 + bias2)
// Learn from mistakes (backpropagation)
// Update weights to get better predictions
Training setup:
- 10 epochs
- Learning rate: 0.1
- Batch size: 32
- Cross-entropy loss
Results
Training Progress
The network learns pretty quickly! Loss drops fast in the first few epochs, then gradually improves.
Epoch 1: 91% accuracy (not bad for random start!)
Epoch 5: 98% accuracy
Epoch 10: 99% accuracy on training data
Final test: 97.5% on unseen digits
So it correctly identifies about 9,750 out of 10,000 handwritten digits it's never seen before.
Getting Started
Train the Model
# Download MNIST dataset
./download_mnist.sh
# Train the model
cargo run --release
After training, the model will be saved to the model/ directory.
The Code
Super simple structure:
// Load data
let train_data = load_mnist_data("train")?;
// Make a neural network
let mut nn = NeuralNetwork::new(784, 128, 10);
// Train it
nn.train(train_data, epochs=10);
// Test it
nn.test(test_data);
- Downloads last month
- 21