Layer

Basics

Memory

A "basic layer" is composed of three types of Matrices which store what the neural network understand, its memory.

• weights - how a layer forward propagates, or predicts. Usually weights initialize as random numbers and are
• errors - how a network knows how far it was from an input or target during back propagation
• deltas - how a network knows to adjust its weights during back propagation

Action

A layer has three different operations for it to "learn"

• predict - usually referred to by non-mortals as "forward propagation", this is where weights are used
• compare - the first of two steps in "back propagation", this compares what a network predicted to a target to calculate deltas and errors
• learn - the second step in "back propagation", this step used to update the weights from what was measured from deltas and errors during compare

Layer Composition

A layer can be very simple, like Random or Add, but "compound layers" can also be described as "layers of layers". Layer Example:

import { FeedForward, layer } from 'brain.js';
const { input, output, add, random } = layer;

function mySuperLayer(input) {
  return add(random(), input);
}

Usage example:

const net = new FeedForward({
  inputLayer: () => input(),
  hiddenLayers: [
    input => mySuperLayer(input)
  ],
  outputLayer: input => output(input)
});

In this example both add and random are composed together, ie layer composition. This simple means of composing layers and in turn networks works with both simple (feed forward) or more complex (recurrent) networks.