README.md
In eduardokapp/r_neural_network: Kapp Neural Nets

Overview

This is my implementation of a neural network in R, structured as an R package that gives you the power to create and train neural networks. It is mostly based around my core understanding of what a neural network is and is a personal-project only (unless..?).

The idea here is to implement a neural network structure with no external tools or packages, so one can actually understand what happens under the hood in a network or in its learning process.

For the implementation, I used three sources as references: - Michael Nielsen's Neural Networks and Deep Learning - 3Blue1Brown's Neural Networks Series - sentdex's Neural Networks from Scratch in Python

Features

Currently, this package supports network creation, network training and predictions with a network object. The training is done using the MSE cost function and the sigmoid activation function.

A network is described as overall information (input size, output size and number of layers) but also includes every layer it has. You can easily access and check a specific layer weights, biases by doing:

# say you want to check layer 3
some_layer <- my_network$layers[[3]]
# weights
some_layer$weights

# biases
some_layers$biases

To find out more about it, read the docs.

Add more cost functions and activation functions
Add more control over the training process

Examples

A very simple example of creating and training a network:

# Generate random data
n_samples <- 1000
inputs <- matrix(runif(n_samples * 10, 0, 1), nrow = 10, ncol = n_samples)
outputs <- matrix(runif(n_samples * 2, 0, 1), nrow = 2, ncol = n_samples)

# Create network object
network <- create_network(10, n_layers = 3, layer_sizes = c(24, 24, 2))

# Train the network
trained_network <- train_network(
  network,
  as.matrix(inputs),
  as.matrix(outputs),
  50,
  100,
  0.1
)

# Make predictions
network_outputs <- predict(trained_network, inputs)