无法将11行玩具神经网络代码转换为JavaScript

Question

I am giving a short presentation on neural networks Tuesday to my fellow student web developers. I was hoping to translate this code (under Part 1, a tiny toy neural network: 2 layer network) to JavaScript so it would be more recognizable to my audience.

import numpy as np

# sigmoid function
def nonlin(x,deriv=False):
    if(deriv==True):
        return x*(1-x)
    return 1/(1+np.exp(-x))

# input dataset
X = np.array([  [0,0,1],
                [0,1,1],
                [1,0,1],
                [1,1,1] ])

# output dataset            
y = np.array([[0,0,1,1]]).T

# seed random numbers to make calculation
# deterministic (just a good practice)
np.random.seed(1)

# initialize weights randomly with mean 0
syn0 = 2*np.random.random((3,1)) - 1

for iter in xrange(10000):

    # forward propagation
    l0 = X
    l1 = nonlin(np.dot(l0,syn0))

    # how much did we miss?
    l1_error = y - l1

    # multiply how much we missed by the 
    # slope of the sigmoid at the values in l1
    l1_delta = l1_error * nonlin(l1,True)

    # update weights
    syn0 += np.dot(l0.T,l1_delta)

print "Output After Training:"
print l1

Here's my JavaScript code as it stands now. I just de-ES6ified it to get it to run in my IDE:

const _ = require('lodash')
const m = require('mathjs')

const sigmoid = function(z) { return 1.0 / (1.0 + Math.exp(-z)) }

const sigmoid_prime = function(z) { return sigmoid(z) * (1 - sigmoid(z)) }

var X = m.matrix([ [0,0,1],[0,1,1],[1,0,1],[1,1,1] ])
var y = m.transpose(m.matrix(([[0,1,1,0]])))

var syn0 = m.random([3, 1], -1, 1)

var l0, l1, l1_delta, l1_error

_.range(10000).forEach(function() {

    l0 = X;
    l1 = m.map(m.multiply(l0, syn0), sigmoid)
    l1_error = m.subtract(y, l1)
    l1_delta = m.dotMultiply(l1_error, m.map(l1, sigmoid_prime))
    syn0 = m.multiply(m.transpose(l0),l1_delta)
})

console.log("Output After Training:")
console.log(l1)

As you can see I'm using mathjs as a substitute for numpy. I have tried to look carefully at the documentation for mathjs and numpy and not mix up my matrix multiplication and my elementwise multiplication, but something is very broken and I get .5 for every output. I have stepped through my program in the debugger and compared values side by side in a python scratch file, starting python out with the values for syn0 that the JavaScript program generated, and it seems like it's here, the backpropagation line, that they slightly diverge (and maybe diverge more over iterations): l1_delta = m.dotMultiply(l1_error, m.map(l1, sigmoid_prime)) . But I can't see why.

EDIT: I should have updated my code before I posted to reflect that in the last version I changed the y definition to var y = m.matrix([ [0], [0], [1], [1]]) and it slightly modified the problem, in that the output switched from being all .5's to floats slightly off of .5.

SECOND EDIT: Brent rightly points out in comments that I have a bug, in that to imitate the code I'm porting from my sigmoid prime function only needs to be z*(1-z). I had missed that wrinkle. Sadly, this doesn't make a difference. Console logging the stringified function and the value of syn0 in the last iteration:

sigmoid prime is function (z) {return sigmoid(z) * (1 - sigmoid(z))}
syn0 is Matrix {
  _data: 
   [ [ 0.21089543115482337 ],
     [ -0.010100491415226356 ],
     [ -0.021376195229226028 ] ],
  _size: [ 3, 1 ],
  _datatype: undefined }

Now changing the function:

sigmoid prime is function (z) { return z * (1 - (z)) }
syn0 is Matrix {
  _data: 
   [ [ 0.2235282818415481 ],
     [ -0.010714305064562765 ],
     [ -0.022890185954402634 ] ],
  _size: [ 3, 1 ],
  _datatype: undefined }

Answer 1

It looks like you're very close, this is a nice port.

I think this is a small bug in your translation of the nonlin function. In the case where the deriv parameter is true, the equation is x * (1 - x) . In your version you are using sigmoid(x) * (1 - sigmoid(x)) . I don't think you need to be calling sigmoid from within sigmoid_prime .

I hope that helps!