Neural network output values over 1?

What is wrong?

after training on iris data https://archive.ics.uci.edu/ml/datasets/Iris/ output of one iris is massively higher than the others

How do we replicate the issue?

download the csv of data

const brain = require('brain.js');
const net = new brain.NeuralNetwork();
const fs = require('fs');

const data = fs.readFileSync('./iris.csv').toString();
const trainingData = data.split('\r\n').map((d) => {
    const [slength, swidth, plength, pwidth, flower] = d.split(',');
    return {
        input: {
            slength: slength / 10,
            swidth: swidth / 10,
            plength: plength / 10,
            pwidth: pwidth / 10,
        },
        output: {
            [flower]: 1,
        },
    };
});
const trainResults = net.train(trainingData);
console.log(trainResults);
const result = net.run({ 
    slength: 0.51,
    swidth: 0.35,
    plength: 0.13999999999999999,
    pwidth: 0.1 });

console.log(result);

How important is this (1-5)?

3

Expected behavior

{ 'Iris-setosa': 0.9973204731941223,
  'Iris-versicolor': 0.003389077726751566,
  'Iris-virginica': 2.462011443240403e-13 }

should be more like

{ 'Iris-setosa': 0.9973204731941223,
  'Iris-versicolor': 0.003389077726751566,
  'Iris-virginica': 0.003 }

Other Comments

training results

{ error: 0.008838672393814171, iterations: 20000 }

[robertleeplummerjr]

The other option you can go with is formatting the output as a string, such as using .toFixed(14).

Example:

2.462011443240403e-13.toFixed(14); // -> "0.00000000000025"

[robertleeplummerjr]

It’s exponent values. An effort to make it more computationally easy and/or visually friendly. 13 zeros vs 2 zeros. Happens to the best of us.