A Minimal Feed-Forward Neural Network made in Rust

This repository contains a minimal but fully functioning Feed-Forward Neural Network (FNN) implemented in the Rust programming language. The neural network uses batch gradient descent for backpropagation. It is single-threaded and uses CPU only for computations.

This implementation uses the mean squared error loss function and includes two types of activation functions: Sigmoid and Rectified Linear Unit (ReLU).

Goal

The goal of this project is to get a better understanding of neural networks by creating one from scratch.

Getting Started

To get a copy of this project up and running on your local machine, you will need Rust.

Clone this repository:

git clone https://github.com/w4g1/neural-network-rust.git

Go into the repository:

cd neural-network-rust

Run XOR example:

cargo run --example xor --release

Epoch: 250000 Error: 0.004982472275727541
Epoch: 500000 Error: 0.0022680697570409874
Epoch: 750000 Error: 0.0014475361058490137
Epoch: 1000000 Error: 0.0010574201380490365
Epoch: 1250000 Error: 0.0008307775961258309
Training completed in 844.0897ms
0 XOR 0 = 0.0364493729489178
0 XOR 1 = 0.9629761743234105
1 XOR 0 = 0.9597184054455132
1 XOR 1 = 0.040673502104589074

Usage

This implementation abstracts loss and activation functions into separate traits, providing flexibility:

trait LossFunction {
    fn compute(&self, target: f64, output: f64) -> f64;
    fn derivative(&self, target: f64, output: f64) -> f64;
}

trait ActivationFunction {
    fn compute(&self, input: f64) -> f64;
    fn derivative(&self, output: f64) -> f64;
}

A Neuron struct supports arbitrary number of inputs and keeps track of its output, error, and the activation function it should use. Various other structural elements such as Connection, Layer, NeuralNetworkConfig, TrainConfig, and NeuralNetwork are defined to accommodate a fully working feed forward neural network.

To build a simple XOR logic function learning neural network, use the following:

let mut network = NeuralNetwork::new(
    vec![
        Layer::new(2, &Activation::Sigmoid(Sigmoid)),
        Layer::new(5, &Activation::Sigmoid(Sigmoid)),
        Layer::new(1, &Activation::Sigmoid(Sigmoid)),
    ],
    NeuralNetworkConfig {
        loss_function: &Loss::MeanSquaredError(MeanSquaredError),
    },
);

To train the network, use the train method on the instance and pass in training data and Configuration for the training.

let config = TrainConfig {
    learning_rate: 0.5,
    steps: -1,
    epochs: -1,
    error_threshold: 0.001,
    eval_frequency: 100_000,
};

network.train(&dataset, config);

License

This project is open-source software licensed under the MIT license.

Contributions