ZeroZen

A neural network from scratch in Rust. No frameworks, no magic — made with just raw math, gradients, and Zen-like focus.

---

Why Make This? 🤔

Ever wondered what goes on inside the black box? I did. So I built zerozen — a neural network from scratch to deeply understand how it works under the hood.

Also, it beautifully combines:

• My love for mathematics (linear algebra and calculus)
• My love for computer science
• My curiosity to peek behind the curtain

Why Rust? 🦀

Because... why not? I briefly considered C, but for reasons only my past self knows, I gracefully switched to Rust.

Why Not Python? 🐍

I wanted to refresh my Rust skills and keep dependencies minimal — yes, not even numpy. We’re going full monk mode here: no distractions, just raw vectors and pure determination.

---

Features ✨

• Zero dependencies (except rand because I like surprises)
• Minimal yet functional neural network core
• Logging during training
• Loss function history tracking
• Fully modular and customizable layers
• Written in pure Rust 🦀

---

Running ZeroZen 🏃‍♂️

The Zen Nix Way 🌿

If you’re a NixOS user (or have Nix+flakes setup):

# Enter a development shell with all dependencies
nix develop

# Build the executable
nix build

# Run the binary
nix run github:nikhilcodes777/zerozen

The Cargo Way 📦

cargo run --example xor

---

Usage Example 🚀

Basic XOR

fn main() -> Result<()> {
    let mut net = Network::builder() // Builder pattern
        .layer(LayerConfig {
            neurons: 3,
            activator: ActivationKind::LeakyReLU(0.01), // Leaky ReLU with customizable leak
        })
        .layer(LayerConfig {
            neurons: 1,
            activator: ActivationKind::Sigmoid,
        })
        .loss(LossKind::MeanSquaredError) // Customizable loss function
        .learning_rate(0.5)
        .epochs(20 * 1000)
        .with_logging(true, 500) // Log every 500 epochs
        .build(2)?; // Builds a network with 2 input features

    let input = matrix!(0.0, 0.0;
                        1.0, 0.0;
                        0.0, 1.0;
                        1.0, 1.0;);

    let targets = matrix!(0.0;
                          1.0;
                          1.0;
                          0.0;);

    println!("Training Started");
    net.train(&input, &targets)?;
    println!("Training Completed");

    let predictions = net.forward(&input)?;
    println!("Inputs\n{input}\nTargets\n{targets}\nPredictions\n{predictions}");

    println!("{net}");
    Ok(())
}

MNIST image upscaler using coordinate-based learning

---

License 📄

This project is licensed under the MIT License.

---

Built with ❤️, lots of ☕, and a healthy dose of "how hard could it be?" attitude.