Proxima

Proxima is a fully connected neural network micro library written in javascript for browsers and nodejs.

Support Proxima development by donating or becoming a sponsor.

Installation

npm install proxima --save

Usage

var Proxima = require('../proxima') // only for nodejs

var xor_training_data = [
    {inputs: [0, 1], targets: [1]},
    {inputs: [1, 0], targets: [1]},
    {inputs: [0, 0], targets: [0]},
    {inputs: [1, 1], targets: [0]}
]

var hyperParameters = {
    neural_network: [2,3,1],
    learning_rate: 0.5,
    max_iterations: 15000,
    cost_threshold: 0.005,
    log_after_x_iterations: 0,
}

var p = new Proxima(hyperParameters)
p.train(xor_training_data)
console.log(p.predict([0,1]))
console.log(p.predict([1,0]))
console.log(p.predict([0,0]))
console.log(p.predict([1,1]))

The above program will print something like this to the console

training: 32.45ms error: 0.004967480795865376 iterations: 1301
training: 37.64ms
[ 0.9003734502498547 ]
[ 0.9000429398063956 ]
[ 0.09926572660491863 ]
[ 0.09967427748286291 ]

Configuration options

var hyperParameters = {
    neural_network: [2,3,1],    // 3 layered neural network with 1 input layer with 2 nodes, 1 hidden layer with 3 nodes and 1 output layer with 1 node
    learning_rate: 0.5,         // η Defaults to 0.5
    max_iterations: 15000,      // Maximum training iterations if the cost_threshold in not reached
    cost_threshold: 0.005,      // Stops training when the result of the cost/loss function is less, defaults to 0.05  
    log_after_x_iterations: 0,
}

What's behind Proxima

Activation functions

• Sigmoid (default)
• Tanh (todo)
• ReLu (todo)
• Leaky ReLu (todo)
• Swish (todo)

Cost functions

• Squared Error (default)
• Mean Squared Error (todo)
• Root Mean Square (todo)
• The Sum of Square Errors (todo)

Gradient descent methods

• Stochastic gradient descent (default)
• Mini-batch gradient descent (todo)
• Batch gradient descent (todo)

Gradient descent optimization algorithms

• Regular gradient descent (default)
• Momentum based gradient descent (todo)
• Implement other gradient descent optimization algorithms (todo)