I've been struggling with Neural Networks for a while now. I get the math behind backpropagation.

Still as reference I'm using the formulas from here.

Screenshot of the formulas

The Network learns XOR:

Prediction After Training: [0.0003508415406266712] Expected: [0.0]

But basically doesn't learn anything on the MNIST dataset:

Error after n trainings examples:

 - 0      Total Net Error:  4.3739634316135225 
 - 10000  Total Net Error:  0.4876292680858326
 - 20000  Total Net Error:  0.39989816082272944
 - 30000  Total Net Error:  0.49507443066631834
 - 40000  Total Net Error:  0.5483594859079792
 - 50000  Total Net Error:  0.5135921029479789
 - 59000  Total Net Error:  0.4686434346776871

[Prediction] [Expected]

 - [0.047784337754445516]               [0]
 - [0.09444684951344406]                [0]
 - [0.0902378720783441]                 [0]
 - [0.09704810171673675]                [0]
 - [0.02940947956812051]                [0]
 - [0.12494839048272757]                [1]
 - [0.1512762065177885]                 [0]
 - [0.055847446615593155]               [0]
 - [0.22983410239796548]                [0]
 - [0.09162426430286492]                [0]

The Old Code ( can be ignored )

I've broken the network down as much as possible. With no matrix or vector multiplication. Here the code for the different classes:

Main File:    

    # Load Trainigs Data
    rawImages, rawLabels, numImagePixels = get_data_and_labels("C:\\Users\\Robin\\Documents\\MNIST\\Images\\train-images.idx3-ubyte", "C:\\Users\\Robin\\Documents\\MNIST\\Labels\\train-labels.idx1-ubyte")

    # Prepare Data
    print("Start Preparing Data")
    images = []
    labels = []
    for i in rawImages:
        insert = []
        for pixel in i:
            insert.append(map0to1(pixel, 255))
    for l in rawLabels:
        y = [0] * 10
        y[l] = 1
    print("Finished Preparing Data")

    # Create Network
    mnistNet = Network((numImagePixels, 16, 16, 10))

    # Train
    print("Start Training")
    for index in range(len(images)):
        netError = mnistNet.train(images[index], labels[index])
        if index % 10000 == 0:
            print(index, " Total Net Error: ", netError)

    prediction = mnistNet.predict(images[0])
    print("After Training The Network Predicted:", prediction, "Expected Was:", labels[0])

class Network:

    def __init__(self, topology):
        # Make Layer List
        self.layerList = []
        # Make Input Layer
        self.layerList.append(Layer(0, topology[0], 0))
        # Make All Other Layers
        for index in range(1, len(topology)):
            self.layerList.append(Layer(index, topology[index], topology[index-1]))

    def predict(self, x):
        # Set x As Value Of Input Layer
        # Feed Through Network
        for index in range(1, len(self.layerList)):
        # Return The Output Of The Last Layer
        return self.layerList[-1].getA()

    def train(self, x, y):
        # Feed Through Network
        prediction = self.predict(x)
        # Container For The Calculated Layer Errors
        errorsPerLayer = []
        # Calculate Error Of The Output Layer
        errorOutputLayer = listSubtract(prediction,y)
        # Add The Error To The Container
        # Calculate The Total Error Of The Network
        totalError = calcTotalError(errorOutputLayer)
        # Calculate The Error Of The Hidden Layers
        for layerNum in range(len(self.layerList)-2, 0, -1):
            # Get The Error Of The Next Layer
            errorOfNextLayer = errorsPerLayer[0]
            # Forward The Calculation To The Next Layer, Which Returns The Weighted Error, By Giving It It's Error
            weightedError = self.layerList[layerNum+1].calculateWeightedError(self.layerList[layerNum].getNeuronNum(), errorOfNextLayer)
            # Forward The Calculation To The Current Layer, Which Returns The Error Of The Layer, By Giving It The Number Of Neurons In The Current Layer And The Weighted Error Of The Next Layer
            currentLayerError = self.layerList[layerNum].calculateError(weightedError)
            # Add The Just Calculated Error To The List
            errorsPerLayer.insert(0, currentLayerError)
        # Insert 0 As Error For The Input Layer, It's Not Important But That Way It's Size Matches The One Of The Layer List
        errorsPerLayer.insert(0, 0)
        # Update Weights And Biases
        for layerNum in range(1, len(self.layerList)):
            # Get The Output Of The Previous Layer
            aOfPrevLayer = self.layerList[layerNum-1].getA()
            # Forward The Error Of The Current Layer And The Output Of The Previous Layer To The Current Layer For Calculating Delta W
            self.layerList[layerNum].updateWeightsAndBiases(errorsPerLayer[layerNum], aOfPrevLayer)

        #print("The Network Predicted: ", prediction, " Expected Was: ", y, " The Error Of The Output Layer Is: ", errorOutputLayer)
        # Return The Total Error Of The Network For Usage Outisde This Class
        return totalError

    def getNetworkInfo(self):
        for layer in self.layerList:


class Layer:

    def __init__(self, layerNum, numNeurons, numNeuronsPrevLayer):
        self.neurons = []
        # Set The Number Of The Layer
        self.layerNum = layerNum
        # Create The Neurons
        for index in range(numNeurons):
        # Print Info
        print("Layer ", layerNum, " makes ", numNeurons, " Neurons", len(self.neurons))

    def feedForward(self, aPrevLayer):
        # Give It To The Neurono For Processing
        for neuron in self.neurons:

    def calculateWeightedError(self, numNeuronsCurrentLayer, errorOfNextLayer):
        # The Container For The Weighted Error Of The Next Layer
        weightedError = []
        # The Calulation For Every Neuron Of The Current Layer One After Another
        for neuronNum in range(numNeuronsCurrentLayer):
            eSum = 0
            # The Error Of The Neuron With The Neuron For Later Calculation
            for e, n in zip(errorOfNextLayer, self.neurons):
                # Forward The Calculation To The Current Neuron, By Giving It It's Error And The Connecting Neuron Num
                eSum += n.weightError(e, neuronNum)
            # Add The Summed And Weighted Error

        # Return The Error Of The Current Layer
        return weightedError

    def calculateError(self, weightedError):
        # The Container For The Error Of The Current Layer
        errorOfCurrentLayer = []
        # The Weighted Error For The Neuron With The Neruon
        for wE, n in zip(weightedError, self.neurons):
            # Add The Product Of The Weighted Error With The Z Of The Current Neuron Run To Sigmoid Prime
            errorOfCurrentLayer.append(wE * sigmoidPrime(n.getZ()))
        # Return The Error Of The Current Layer
        return errorOfCurrentLayer

    def updateWeightsAndBiases(self, errorOfCurrentLayer, aOfPrevLayer):
        # The Error For The Neuron With The Neuron
        for e, n in zip(errorOfCurrentLayer, self.neurons):
            # Error Of Current Layer Is Equal To The Delta Of The Bias So Apply That
            # Forward The Error And All The Activity Of The Previous Layer To The Current Neuron To Update It's Weights
            n.updateWeights(e, aOfPrevLayer)

    def setInput(self, x):
        # Set It To Every Neuron
        for neuron, val in zip(self.neurons, x):

    def getA(self):
        aOfLayer = []
        for neuron in self.neurons:
        return aOfLayer
    def getNeuronNum(self):
        return len(self.neurons)

    def getLayerInfo(self):
        return "Layer( %i ), has %i Neurons" % (self.layerNum, len(self.neurons))


class Neuron:

    def __init__(self, numNeuronsPrevLayer):
        self.a = 0
        self.z = 0
        self.b = 0.5
        if numNeuronsPrevLayer != 0:
            self.w = np.random.uniform(low = 0, high = 0.5, size=(numNeuronsPrevLayer,))

    def feedForward(self, aPrevLayer):
        # Reset Z
        self.z = 0
        # Calculate Z
        for w, a in zip(self.w, aPrevLayer):
            self.z += w*a
        # Add Bias
        self.z += self.b
        # Calculate A
        self.a = sigmoid(self.z)

    def weightError(self, e, neuronNum):
        # Weight Error With The Connecting Weight
        return e * self.w[neuronNum]

    def updateWeights(self, e, aOfPrevLayer):
        # The Weight With The Matching Activity Of The Previous Layer
        for index in range(len(self.w)):
            # The Delta Of The Weight Is The Error Of That Neuron Mutliplied With The Through The Weight Connected Activiy Of The Previous Layer
            self.w[index] -= e * aOfPrevLayer[index]

    def updateBias(self, e):
        # E Is The Delta Of The Bias1
        self.b -= e

    def setInput(self, x):
        self.z = x
        self.a = x

    def getA(self):
        return self.a
    def getZ(self):
        return self.z
    def getB(self):
        return self.b
    def getW(self):
        return self.w


**helper functions**

    def map0to1(val, valMax):
        return val/valMax

    def calcTotalError(errorOutputLayer):
        totalError = 0
        for e in errorOutputLayer:
            totalError += e**2
        totalError *= 0.5
        return totalError

    def listSubtract(list1, list2):
        subbed = []
        for l1, l2 in zip(list1, list2):
        return subbed

The New Code

Main File

import random
from network import *
from mnistreader import *

def map0to1(val, valMax):
    return val/valMax

trainX = [ [0.0,0.0], [1.0,0.0], [0.0,1.0], [1.0,1.0] ]
trainY = [ [ 0.0], [ 1.0 ], [ 1.0 ], [ 0.0 ] ]

# Create Network
xorNet = Network((2,2,1))

# Train
for index in range(100000):
    randIndex = random.randint(0, 3)
    xorNet.train(trainX[randIndex], trainY[randIndex])

print("Prediction After Training:", xorNet.predict(trainX[0]), "Expected:", trainY[0])
print("Prediction After Training:", xorNet.predict(trainX[1]), "Expected:", trainY[1])
print("Prediction After Training:", xorNet.predict(trainX[2]), "Expected:", trainY[2])
print("Prediction After Training:", xorNet.predict(trainX[3]), "Expected:", trainY[3])

mnistNet = Network((784, 30, 10))

# Load Trainigs Data
rawImages, rawLabels, numImagePixels = get_data_and_labels("C:\\Users\\Robin\\Documents\\MNIST\\Images\\train-images.idx3-ubyte", "C:\\Users\\Robin\\Documents\\MNIST\\Labels\\train-labels.idx1-ubyte")

# Prepare Data
print("Start Preparing Data")
images = []
labels = []
for i in rawImages:
    insert = []
    for pixel in i:
        insert.append(map0to1(pixel, 255))
for l in rawLabels:
    y = [0] * 10
    y[l] = 1
print("Finished Preparing Data")

# Define Variables
learningRate = 0.0001
error = 10

# Training
while error > 0.1:
    for tNum in range(len(images)):
        error = mnistNet.train(images[tNum], labels[tNum], learningRate)
    print("Error:", error, "\n Prediction:\n", mnistNet.predict(images[1]), "\nExpected:", rawLabels[1], "\n\n")

# Test Prediction
print("For", rawLabels[1], "Predicted\n", mnistNet.predict(images[1]))

Network Class

import numpy as np

from layer import *
from transferfunction import *

class Network:

    def __init__(self, shape):
        # Save The Shape Of The Nework
        self.shape = shape
        # Create A List Of Layers
        self.layers = []
        # Create Input Layer
        self.layers.append(Layer((shape[0],), layerType = 'Input'))
        # Create Hidden Layers
        for numNeurons, numNeuronsPrevLayer in zip(shape[1:], shape[:-2]):
            self.layers.append(Layer((numNeurons, numNeuronsPrevLayer), layerType = 'Hidden'))
        # Create Output Layer
        self.layers.append(Layer((shape[-1], shape[-2]), layerType = 'Output'))

    def predict(self, x):
        # X Is A Row So Shape It To Be A Column
        x = np.array(x).reshape(-1, 1)
        # Set X To Be The Ouput Of The Input Layer
        # Feed Through Other Layers
        for layerNum in range(1,len(self.layers)):
        # Return The Output Of The Output Layer
        return self.layers[-1].getOutput()

    def train(self, x, y, learningRate):
        1. Feed Forward
        2. Calculate Error
        3. Calulate Deltas
        4. Apply Deltas

        Error Output Layer = f'(z) * (prediction - target)
        Error Hidden Layer = f'(z) * ( transposed weights next layer DOT error next layer )
        Delta Bias         = learning rate * error
        Delta Weights      = learning rate * ( error DOT transposed activity previous layer )


        # Feed Through Network
        prediction = self.predict(x)
        # Y Is A Row So Shape It To Be A Column
        y = np.array(y).reshape(-1, 1)

        # Calculate Error
        error = prediction - y
        # Calculate Total Error
        totalError = 0.5 * np.sum(error**2)

        # Create Container For The Deltas
        deltas = []
        # Calculate Delta For Output Layer
        deltas.append( np.multiply(sigmoidPrime(self.layers[-1].getZ()), error) )
        # Calculate Deltas For Every Hidden Layer
        for layerNum in range(len(self.layers)-2, 0, -1):
            # Compute The Weighted Error Of The Next Layer
            weightedErrorOfNextLayer = np.dot(self.layers[layerNum+1].getW().T, deltas[0])
            # Compute The Delta And Add It In Front
            deltas.insert(0, np.multiply(sigmoidPrime(self.layers[layerNum].getZ()), weightedErrorOfNextLayer) )
        # Insert Placeholder To Make Delta List As Big As The Layer List
        deltas.insert(0, 0)

        # For Numerical Grandient Checking Do
        numericalGradients = self.performNumericalGradientChecking(x,y)

        # Update Weights And Biases
        for layerNum in range(1, len(self.layers)):
            self.layers[layerNum].updateBias(deltas[layerNum], learningRate)
            self.layers[layerNum].updateWeight(deltas[layerNum], self.layers[layerNum-1].getOutput(), learningRate, numericalGradients[layerNum-1])

        # Show Information
        #print('Network Predicted: \n', prediction, '\nTarget:\n', y, '\nError: ', totalError)
        return totalError

    def performNumericalGradientChecking(self, x, y):
        # Container Saving All Current Weight Values
        weightSave = []
        # Save All The Weight Values
        for layerNum in range(1, len(self.layers)):

        # Define Epsilon To Be A Small Number
        epsilon = 1e-4

        # Gradient Container
        numericalGradients = []

        #Perform The Check For Every Layer Therfore Every Set Of Weights
        for layerNum in range(1, len(self.layers)):

            # Feed Forward With Changed Weights(+epsilon), And Compute Cost
            self.layers[layerNum].setW(weightSave[layerNum-1] + epsilon)
            prediction = self.predict(x)
            loss2 = 0.5 * np.sum ((prediction - y)**2)

            # Feed Forward With Changed Weights(-epsilon), And Compute Cost
            self.layers[layerNum].setW(weightSave[layerNum-1] - epsilon)
            prediction = self.predict(x)
            loss1 = 0.5 * np.sum ((prediction - y)**2)

            # Reset Weight

            # Calculate Numerical Loss
            numericalGradient = (loss2 - loss1) / (2 * epsilon)

            # Add The Numerical Grandient

        return numericalGradients

    def __str__(self):
        strBuff = ''
        for layer in self.layers:
            strBuff += layer.getInfo()
        return strBuff

Layer Class

import numpy as np

from transferfunction import *

class Layer:

    def __init__(self, neurons, layerType = 'Hidden'):
        ''' Neurons Is A Tuple Consisting Of [0]=NumNeurons And [1]=numNeuronsPrevLayer '''

        # Remember The Type Of The Layer
        self.layerType = layerType
        # Remember How Many Neurons This Layer Has
        self.neuronCount = neurons[0]
        # Create Layer Based On Type
        if layerType.lower() == 'input':
            # Create Container For Input Data
            self.a = []
        elif layerType.lower() == 'hidden' or layerType.lower() == 'output':
            # Create Container For Activation
            self.z = []
            # Create Container For Neurons Input
            self.a = []
            # Create Weights
            self.w = np.random.uniform(low = 0.0, high = 0.4, size=(neurons[0], neurons[1]))
            # Create Prev Delta Weight For Momentum
            self.momentum = 0.3
            self.prevDelta = np.full((neurons[0], neurons[1]), 0)
            # Create Bias
            self.b = np.full((neurons[0],1), 0, dtype=float)
            print('Wrong Type Of Layer Specified')

    def feedForward(self, aPrevLayer):
        self.z = np.dot(self.w, aPrevLayer) + self.b
        self.a = sigmoid(self.z)

    def updateBias(self, e, learningRate):
        self.b -= learningRate * e
    def updateWeight(self, e, aPrevLayer, learningRate, numericalGradient):
        # Calulate The Delta Of The Weights
        deltaW = np.dot(e, aPrevLayer.T)

        # Compare DeltaW With The Numerical Grandient
        check = np.linalg.norm(deltaW - numericalGradient) / np.linalg.norm(deltaW + numericalGradient)
        # DEBUG

        # The Weight Change Is The Delta With The Addition Of The Momentum
        self.w -= ( learningRate * deltaW ) + ( self.momentum * self.prevDelta)
        # Save The Current DeltaW
        self.prevDelta = deltaW

    def getW(self):
        return self.w
    def getZ(self):
        return self.z
    def getOutput(self):
        return self.a

    def setOutput(self, x):
        self.a = x
    def setW(self, w):
        self.w = w

    def getInfo(self):
        if self.layerType.lower() == 'input':
            return 'Input Layer With ' + str(self.neuronCount) + ' Neurons\n'
            return self.layerType + ' Layer With ' + str(self.neuronCount) + ' Neurons And Weights Of Shape: ' + str(self.w.shape) + ' With Biases Of Shape: ' + str(self.b.shape) + '\n'

So my Question simply is: "What's wrong?"


  1. Error gets stuck at 0.45
  2. When using a hidden layer with 800 neurons I get the warning division by zero and all becomes NaN's from sigmoid prime
  3. Numerical gradient checking for the hidden Layer is: 1.0 and for the output layer: 0.995784895209. I know that is supposed to be a very small number. But on the second trainings example it creates a overflow error and becomes NaN's

Major Edit I'm truly grateful for all suggestions so far, I've updated the question using a vectorized form so it's easier to get an overview of what I'm doing here. I tried gradient checking now too, not sure if I implemented it right (used the tutorial by Welch Labs (https://youtu.be/pHMzNW8Agq4)) I hope the Code is readable

  • 1
    $\begingroup$ Where is your main program telling eg. how many neurons your network has and other initialization parameters you use for the whole program? $\endgroup$
    – mico
    Commented Jan 14, 2018 at 14:42
  • $\begingroup$ @Robin why are you using MLP? Do you know its weakness for image data? $\endgroup$ Commented Jan 25, 2018 at 14:08
  • $\begingroup$ @Media yes I'm using a fully connected feed forward neural network the most standard as it can get. I know that a CNN is better but I thought if other NNs can learn to classify MNIST data with the most simple methods so should I. But The Error keeps converging on 0.45 no matter what I change. Interested in your thoughts. $\endgroup$
    – Robin
    Commented Jan 25, 2018 at 14:15
  • $\begingroup$ Look, this error that you are facing is not logical. I don't know if you are training the whole data using MLP or not. I've tried to do that using MLP and the result was far away from yours. But about MLPs, you have to consider a point. After training an MLP, if you try to visualize the net, you will see that what it has learned is like mask. MLPs are not useful for learning images, because what they do is not learning features of shapes, unlike CNNs, they just try to make a mask. If you want I have a ready code for MLPs and I can upload it for you somewhere. But [TO be continued]... $\endgroup$ Commented Jan 25, 2018 at 14:21
  • $\begingroup$ ... (!!) but consider that you are not doing well for your data. I suggest two things. make a simple MLP on tensorflow or keras and see that It performs well. then try visualize the net to fully understand what MLPs can learn. If you want I guess I can find a video for which can help you the problem. $\endgroup$ Commented Jan 25, 2018 at 14:24

3 Answers 3


Your network has 28 x 28 = 784 (normal MNIST size) inputs, 16 + 16 hidden nodes and 10 outputs. This is not enough for an enough accurate model as a result.

This question suggests to use 256 x 256 hidden nodes and Wikipedia page on MNIST gives for 2-layer reference the values: 784-800-10 meaning 800 x 10 nodes. Wikipedia gives error rate 0.016 for that 800 x 10 solution.

Sidenote: For 6-layer Deep Neural Network the numbers are : 784-2500-2000-1500-1000-500-10 so the new numbers I gave aren't that big. Of course the error rate on DNN is 0.0035 so it needs those layers.


Do I have to change the learning rate which I now defined to be 0.3?

Answer in the referred question says 0.0001 is more appropriate value.

Edit of question author I marked this as solved because now it is working. I implemented it again and after tweaking the hyperparameters it worked. So this is, in a way the solution to my problem. Because I asked why this problem occurs in the first place, my implementation error aside.

  • $\begingroup$ Thx for now. I will make the changes and try it out as soon as I get a chance to. Keep you posted. Really appreciate the help. That topic interests me a lot but couldn't crack it so far. $\endgroup$
    – Robin
    Commented Jan 14, 2018 at 21:08
  • $\begingroup$ so yeah it's better but kinda not working -> see edit $\endgroup$
    – Robin
    Commented Jan 17, 2018 at 15:12
  • 1
    $\begingroup$ @Robin look my edit. $\endgroup$
    – mico
    Commented Jan 17, 2018 at 16:38
  • $\begingroup$ Damm I'm sorry didn't use a learning rate before and just searched the internet for common ones. Next time I'll read it better $\endgroup$
    – Robin
    Commented Jan 17, 2018 at 18:45
  • $\begingroup$ so the network performed better than ever before but is stuck -> edit $\endgroup$
    – Robin
    Commented Jan 19, 2018 at 12:50

I guess you are doing something wrong in your code. I guess its better to use gradient checking approach for figuring out whether the whole code has any problem or not.

Based on the comments, if I want to show you exactly what happens, first take a look at here which professor Hinton himself explains that what MLPs learn is like learning masks instead of learning the features of the inputs. For illustrating more you can take a look at here which shows that MLPs can learn what and I hope that you can expand it to higher dimensions, for your case 784 dimensions. If you use MLPs alone, you will face to the problem of something which is like masking. What CNNs try to find is feature. They try to find features which can be better explained by finding hierarchical features. You can also take a look at here which explains that using CNNs you try to find features, and after convolution layers, you use dense layers, MLPs which they try to classify the features.

MLPs are good for classifying MNIST data set but they are weak for generalizing unseen data. If your code works fine after using gradient checking try to use different hyper parameters which I guess our friend has explained so much well for you.


You need to increase no of hidden units. The more the number, more extract feature your network builds. Be wise with it, as too many features will make it overfit and it won't generalize on new data.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.