# How to Implement the Backpropagation Algorithm From Scratch In Python

The backpropagation algorithm is the classical feed-forward artificial neural network.

It is the technique still used to train large deep learning networks.

In this tutorial, you will discover how to implement the backpropagation algorithm from scratch with Python.

After completing this tutorial, you will know:

• How to forward-propagate an input to calculate an output.
• How to back-propagate error and train a network.
• How to apply the backpropagation algorithm to a real-world predictive modeling problem.

Let’s get started.

• Update Nov/2016: Fixed a bug in the activate() function. Thanks Alex!
• Update Jan/2017: Changed the calculation of fold_size in cross_validation_split() to always be an integer. Fixes issues with Python 3.
• Update Jan/2017: Updated small bug in update_weights(). Thanks Tomasz!

How to Implement the Backpropagation Algorithm From Scratch In Python
Photo by NICHD, some rights reserved.

## Description

This section provides a brief introduction to the Backpropagation Algorithm and the Wheat Seeds dataset that we will be using in this tutorial.

### Backpropagation Algorithm

The Backpropagation algorithm is a supervised learning method for multilayer feed-forward networks from the field of Artificial Neural Networks.

Feed-forward neural networks are inspired by the information processing of one or more neural cells, called a neuron. A neuron accepts input signals via its dendrites, which pass the electrical signal down to the cell body. The axon carries the signal out to synapses, which are the connections of a cell’s axon to other cell’s dendrites.

The principle of the backpropagation approach is to model a given function by modifying internal weightings of input signals to produce an expected output signal. The system is trained using a supervised learning method, where the error between the system’s output and a known expected output is presented to the system and used to modify its internal state.

Technically, the backpropagation algorithm is a method for training the weights in a multilayer feed-forward neural network. As such, it requires a network structure to be defined of one or more layers where one layer is fully connected to the next layer. A standard network structure is one input layer, one hidden layer, and one output layer.

Backpropagation can be used for both classification and regression problems, but we will focus on classification in this tutorial.

In classification problems, best results are achieved when the network has one neuron in the output layer for each class value. For example, a 2-class or binary classification problem with the class values of A and B. These expected outputs would have to be transformed into binary vectors with one column for each class value. Such as [1, 0] and [0, 1] for A and B respectively. This is called a one hot encoding.

### Wheat Seeds Dataset

The seeds dataset involves the prediction of species given measurements seeds from different varieties of wheat.

There are 201 records and 7 numerical input variables. It is a classification problem with 3 output classes. The scale for each numeric input value vary, so some data normalization may be required for use with algorithms that weight inputs like the backpropagation algorithm.

Below is a sample of the first 5 rows of the dataset.

Using the Zero Rule algorithm that predicts the most common class value, the baseline accuracy for the problem is 28.095%.

Download the seeds dataset and place it into your current working directory with the filename seeds_dataset.csv. The dataset is in tab-separated format, so you must convert it to CSV using a text editor or a spreadsheet program.

## Tutorial

This tutorial is broken down into 6 parts:

1. Initialize Network.
2. Forward Propagate.
3. Back Propagate Error.
4. Train Network.
5. Predict.
6. Seeds Dataset Case Study.

These steps will provide the foundation that you need to implement the backpropagation algorithm from scratch and apply it to your own predictive modeling problems.

### 1. Initialize Network

Each neuron has a set of weights that need to be maintained. One weight for each input connection and an additional weight for the bias. We will need to store additional properties for a neuron during training, therefore we will use a dictionary to represent each neuron and store properties by names such as ‘weights‘ for the weights.

A network is organized into layers. The input layer is really just a row from our training dataset. The first real layer is the hidden layer. This is followed by the output layer that has one neuron for each class value.

We will organize layers as arrays of dictionaries and treat the whole network as an array of layers.

It is good practice to initialize the network weights to small random numbers. In this case, will we use random numbers in the range of 0 to 1.

Below is a function named initialize_network() that creates a new neural network ready for training. It accepts three parameters, the number of inputs, the number of neurons to have in the hidden layer and the number of outputs.

You can see that for the hidden layer we create n_hidden neurons and each neuron in the hidden layer has n_inputs + 1 weights, one for each input column in a dataset and an additional one for the bias.

You can also see that the output layer that connects to the hidden layer has n_outputs neurons, each with n_hidden + 1 weights. This means that each neuron in the output layer connects to (has a weight for) each neuron in the hidden layer.

Let’s test out this function. Below is a complete example that creates a small network.

Running the example, you can see that the code prints out each layer one by one. You can see the hidden layer has one neuron with 2 input weights plus the bias. The output layer has 2 neurons, each with 1 weight plus the bias.

Now that we know how to create and initialized a network, let’s see how we can use it to calculate an output.

### 2. Forward Propagate

We can calculate an output from a neural network by propagating an input signal through each layer until the output layer outputs its values.

We call this forward-propagation.

It is the technique we will need to generate predictions during training that will need to be corrected, and it is the method we will need after the network is trained to make predictions on new data.

We can break forward propagation down into three parts:

1. Neuron Activation.
2. Neuron Transfer.
3. Forward Propagation.

#### 2.1. Neuron Activation

The first step is to calculate the activation of one neuron given an input.

The input could be a row from our training dataset, as in the case of the hidden layer. It may also be the outputs from each neuron in the hidden layer, in the case of the output layer.

Neuron activation is calculated as the weighted sum of the inputs. Much like linear regression.

Where weight is a network weight, input is an input, i is the index of a weight or an input and bias is a special weight that has no input to multiply with (or you can think of the input as always being 1.0).

Below is an implementation of this in a function named activate(). You can see that the function assumes that the bias is the last weight in the list of weights. This helps here and later to make the code easier to read.

Now, let’s see how to use the neuron activation.

#### 2.2. Neuron Transfer

Once a neuron is activated, we need to transfer the activation to see what the neuron output actually is.

Different transfer functions can be used. It is traditional to use the sigmoid activation function, but you can also use the tanh (hyperbolic tangent) function to transfer outputs. More recently, the rectifier transfer function has been popular with large deep learning networks.

The sigmoid activation function looks like an S shape, it’s also called the logistic function. It can take any input value and produce a number between 0 and 1 on an S-curve. It is also a function of which we can easily calculate the derivative (slope) that we will need later when backpropagating error.

We can transfer an activation function using the sigmoid function as follows:

Where e is the base of the natural logarithms (Euler’s number).

Below is a function named transfer() that implements the sigmoid equation.

Now that we have the pieces, let’s see how they are used.

#### 2.3. Forward Propagation

Forward propagating an input is straightforward.

We work through each layer of our network calculating the outputs for each neuron. All of the outputs from one layer become inputs to the neurons on the next layer.

Below is a function named forward_propagate() that implements the forward propagation for a row of data from our dataset with our neural network.

You can see that a neuron’s output value is stored in the neuron with the name ‘output‘. You can also see that we collect the outputs for a layer in an array named new_inputs that becomes the array inputs and is used as inputs for the following layer.

The function returns the outputs from the last layer also called the output layer.

Let’s put all of these pieces together and test out the forward propagation of our network.

We define our network inline with one hidden neuron that expects 2 input values and an output layer with two neurons.

Running the example propagates the input pattern [1, 0] and produces an output value that is printed. Because the output layer has two neurons, we get a list of two numbers as output.

The actual output values are just nonsense for now, but next, we will start to learn how to make the weights in the neurons more useful.

### 3. Back Propagate Error

The backpropagation algorithm is named for the way in which weights are trained.

Error is calculated between the expected outputs and the outputs forward propagated from the network. These errors are then propagated backward through the network from the output layer to the hidden layer, assigning blame for the error and updating weights as they go.

The math for backpropagating error is rooted in calculus, but we will remain high level in this section and focus on what is calculated and how rather than why the calculations take this particular form.

This part is broken down into two sections.

1. Transfer Derivative.
2. Error Backpropagation.

#### 3.1. Transfer Derivative

Given an output value from a neuron, we need to calculate it’s slope.

We are using the sigmoid transfer function, the derivative of which can be calculated as follows:

Below is a function named transfer_derivative() that implements this equation.

Now, let’s see how this can be used.

#### 3.2. Error Backpropagation

The first step is to calculate the error for each output neuron, this will give us our error signal (input) to propagate backwards through the network.

The error for a given neuron can be calculated as follows:

Where expected is the expected output value for the neuron, output is the output value for the neuron and transfer_derivative() calculates the slope of the neuron’s output value, as shown above.

This error calculation is used for neurons in the output layer. The expected value is the class value itself. In the hidden layer, things are a little more complicated.

The error signal for a neuron in the hidden layer is calculated as the weighted error of each neuron in the output layer. Think of the error traveling back along the weights of the output layer to the neurons in the hidden layer.

The back-propagated error signal is accumulated and then used to determine the error for the neuron in the hidden layer, as follows:

Where error_j is the error signal from the jth neuron in the output layer, weight_k is the weight that connects the kth neuron to the current neuron and output is the output for the current neuron.

Below is a function named backward_propagate_error() that implements this procedure.

You can see that the error signal calculated for each neuron is stored with the name ‘delta’. You can see that the layers of the network are iterated in reverse order, starting at the output and working backwards. This ensures that the neurons in the output layer have ‘delta’ values calculated first that neurons in the hidden layer can use in the subsequent iteration. I chose the name ‘delta’ to reflect the change the error implies on the neuron (e.g. the weight delta).

You can see that the error signal for neurons in the hidden layer is accumulated from neurons in the output layer where the hidden neuron number j is also the index of the neuron’s weight in the output layer neuron[‘weights’][j].

Let’s put all of the pieces together and see how it works.

We define a fixed neural network with output values and backpropagate an expected output pattern. The complete example is listed below.

Running the example prints the network after the backpropagation of error is complete. You can see that error values are calculated and stored in the neurons for the output layer and the hidden layer.

Now let’s use the backpropagation of error to train the network.

### 4. Train Network

The network is trained using stochastic gradient descent.

This involves multiple iterations of exposing a training dataset to the network and for each row of data forward propagating the inputs, backpropagating the error and updating the network weights.

This part is broken down into two sections:

1. Update Weights.
2. Train Network.

#### 4.1. Update Weights

Once errors are calculated for each neuron in the network via the back propagation method above, they can be used to update weights.

Network weights are updated as follows:

Where weight is a given weight, learning_rate is a parameter that you must specify, error is the error calculated by the backpropagation procedure for the neuron and input is the input value that caused the error.

The same procedure can be used for updating the bias weight, except there is no input term, or input is the fixed value of 1.0.

Learning rate controls how much to change the weight to correct for the error. For example, a value of 0.1 will update the weight 10% of the amount that it possibly could be updated. Small learning rates are preferred that cause slower learning over a large number of training iterations. This increases the likelihood of the network finding a good set of weights across all layers rather than the fastest set of weights that minimize error (called premature convergence).

Below is a function named update_weights() that updates the weights for a network given an input row of data, a learning rate and assume that a forward and backward propagation have already been performed.

Remember that the input for the output layer is a collection of outputs from the hidden layer.

Now we know how to update network weights, let’s see how we can do it repeatedly.

#### 4.2. Train Network

As mentioned, the network is updated using stochastic gradient descent.

This involves first looping for a fixed number of epochs and within each epoch updating the network for each row in the training dataset.

Because updates are made for each training pattern, this type of learning is called online learning. If errors were accumulated across an epoch before updating the weights, this is called batch learning or batch gradient descent.

Below is a function that implements the training of an already initialized neural network with a given training dataset, learning rate, fixed number of epochs and an expected number of output values.

The expected number of output values is used to transform class values in the training data into a one hot encoding. That is a binary vector with one column for each class value to match the output of the network. This is required to calculate the error for the output layer.

You can also see that the sum squared error between the expected output and the network output is accumulated each epoch and printed. This is helpful to create a trace of how much the network is learning and improving each epoch.

We now have all of the pieces to train the network. We can put together an example that includes everything we’ve seen so far including network initialization and train a network on a small dataset.

Below is a small contrived dataset that we can use to test out training our neural network.

Below is the complete example. We will 2 neurons in the hidden layer. It is a binary classification problem (2 classes) so there will be two neurons in the output layer. The network will be trained for 20 epochs with a learning rate of 0.5, which is high because we are training for so few iterations.

Running the example first prints the sum squared error each training epoch. We can see a trend of this error decreasing with each epoch.

Once trained, the network is printed, showing the learned weights. Also still in the network are output and delta values that can be ignored. We could update our training function to delete these data if we wanted.

Once a network is trained, we need to use it to make predictions.

### 5. Predict

Making predictions with a trained neural network is easy enough.

We have already seen how to forward-propagate an input pattern to get an output. This is all we need to do to make a prediction. We can use the output values themselves directly as the probability of a pattern belonging to each output class.

It may be more useful to turn this output back into a crisp class prediction. We can do this by selecting the class value with the larger probability. This is also called the arg max function.

Below is a function named predict() that implements this procedure. It returns the index in the network output that has the largest probability. It assumes that class values have been converted to integers starting at 0.

We can put this together with our code above for forward propagating input and with our small contrived dataset to test making predictions with an already-trained network. The example hardcodes a network trained from the previous step.

The complete example is listed below.

Running the example prints the expected output for each record in the training dataset, followed by the crisp prediction made by the network.

It shows that the network achieves 100% accuracy on this small dataset.

Now we are ready to apply our backpropagation algorithm to a real world dataset.

### 6. Wheat Seeds Dataset

This section applies the Backpropagation algorithm to the wheat seeds dataset.

The first step is to load the dataset and convert the loaded data to numbers that we can use in our neural network. For this we will use the helper function load_csv() to load the file, str_column_to_float() to convert string numbers to floats and str_column_to_int() to convert the class column to integer values.

Input values vary in scale and need to be normalized to the range of 0 and 1. It is generally good practice to normalize input values to the range of the chosen transfer function, in this case, the sigmoid function that outputs values between 0 and 1. The dataset_minmax() and normalize_dataset() helper functions were used to normalize the input values.

We will evaluate the algorithm using k-fold cross-validation with 5 folds. This means that 201/5=40.2 or 40 records will be in each fold. We will use the helper functions evaluate_algorithm() to evaluate the algorithm with cross-validation and accuracy_metric() to calculate the accuracy of predictions.

A new function named back_propagation() was developed to manage the application of the Backpropagation algorithm, first initializing a network, training it on the training dataset and then using the trained network to make predictions on a test dataset.

The complete example is listed below.

A network with 5 neurons in the hidden layer and 3 neurons in the output layer was constructed. The network was trained for 500 epochs with a learning rate of 0.3. These parameters were found with a little trial and error, but you may be able to do much better.

Running the example prints the average classification accuracy on each fold as well as the average performance across all folds.

You can see that backpropagation and the chosen configuration achieved a mean classification accuracy of 95.238% which is dramatically better than the Zero Rule algorithm that did slightly better than 28.095% accuracy.

## Extensions

This section lists extensions to the tutorial that you may wish to explore.

• Tune Algorithm Parameters. Try larger or smaller networks trained for longer or shorter. See if you can get better performance on the seeds dataset.
• Additional Methods. Experiment with different weight initialization techniques (such as small random numbers) and different transfer functions (such as tanh).
• More Layers. Add support for more hidden layers, trained in just the same way as the one hidden layer used in this tutorial.
• Regression. Change the network so that there is only one neuron in the output layer and that a real value is predicted. Pick a regression dataset to practice on. A linear transfer function could be used for neurons in the output layer, or the output values of the chosen dataset could be scaled to values between 0 and 1.
• Batch Gradient Descent. Change the training procedure from online to batch gradient descent and update the weights only at the end of each epoch.

Did you try any of these extensions?

## Review

In this tutorial, you discovered how to implement the Backpropagation algorithm from scratch.

Specifically, you learned:

• How to forward propagate an input to calculate a network output.
• How to back propagate error and update network weights.
• How to apply the backpropagation algorithm to a real world dataset.

Do you have any questions?

## Want to Code Algorithms in Python Without Math?

#### Code Your First Algorithm in Minutes

…with step-by-step tutorials on real-world datasets

Discover how in my new Ebook:
Machine Learning Algorithms From Scratch

It covers 18 tutorials with all the code for 12 top algorithms, like:
Linear Regression, k-Nearest Neighbors, Stochastic Gradient Descent and much more…

### 164 Responses to How to Implement the Backpropagation Algorithm From Scratch In Python

1. Talk Data To Me November 7, 2016 at 9:28 pm #

That’s what I was looking for. Write a neural network without any libraries (scikit, keras etc.) Thnak you very much!

2. MO November 8, 2016 at 9:26 am #

where can i see your data set, i want to see how it looked like

• Jason Brownlee November 8, 2016 at 10:01 am #

Hi MO.

The small contrived dataset used for testing is listed inline in the post in section 4.2

The dataset used for the full example is on the UCI ML repository, linked in the section titled “Wheat Seeds Dataset”. Here is the direct link:
http://archive.ics.uci.edu/ml/datasets/seeds

3. prakash November 11, 2016 at 12:40 am #

in two class classification for 0 the expected value is [1,0] for 1 its is [0,1].
how will be the output vectors for more than two class??

• Jason Brownlee November 11, 2016 at 10:02 am #

Hi prakash,

For multi-class classification, we can extend the one hot encoding.

Three class values for “red”, “green” “blue” can be represented as an output vector like:
1, 0, 0 for red
0, 1, 0 for green
0, 0, 1 for blue

I hope that helps.

4. Rakesh November 13, 2016 at 3:41 pm #

Hi, Jason.
You’ve mentioned that there are 3 output classes.
How do we check the values which come under the 3 classes / clusters?
Could we print the data which fall under each class?

• Jason Brownlee November 14, 2016 at 7:35 am #

Hi Rakesh,

The data does belong to 3 classes. We can check the skill of our model by comparing the predicted classes to the actual/expected classes and calculate an accuracy measure.

5. Alex November 16, 2016 at 12:35 pm #

I’m confused why the activation method iterates from 0 to len(inputs) – 1 instead of from 0 to len(weights) – 1. Am I missing something?

• Jason Brownlee November 17, 2016 at 9:47 am #

Hi Alex,

The length of weights is the length of the input + 1 (to accommodate the bias term).

We add the bias term first, then we add the weighted inputs. This is why we iterate over input values.

Does that help?

• Alex November 17, 2016 at 12:29 pm #

When I step through the code above for the ‘forward_propagate’ test case, I see the code correctly generate the output for the single hidden node but that output doesn’t get correctly processed when determining the outputs for the output layer. As written above in the activate function ‘for i in range(len(inputs)-1):’, when the calculation gets to the activate function for the output node for class=0, since ‘inputs’ has a single element in it (the output from the single hidden node), ‘len(inputs) – 1’ equals 0 so the for loop never executes. I’m assuming the code is supposed to read ‘for i in range(len(weights) -1):’ Does that make sense?

I’m just trying to make sure I don’t fundamentally misunderstand something and improve this post for other readers. This site has been really, really helpful for me.

• Jason Brownlee November 18, 2016 at 8:27 am #

I’m with you now, thanks for helping me catch-up.

Nice spot. I’ll fix up the tutorial.

Update: Fixed. Thanks again mate!

6. Tomasz Panek November 21, 2016 at 1:23 am #

# Update network weights with error
def update_weights(network, row, l_rate):
for i in range(len(network)):
inputs = row
if i != 0:
inputs = [neuron[‘output’] for neuron in network[i – 1]]
for neuron in network[i]:
for j in range(len(inputs)-1):
neuron[‘weights’][j] += l_rate * neuron[‘delta’] * inputs[j]
neuron[‘weights’][-1] += l_rate * neuron[‘delta’]

In this fragment:
for j in range(len(inputs)-1):
neuron[‘weights’][j] += l_rate * neuron[‘delta’] * inputs[j]
neuron[‘weights’][-1] += l_rate * neuron[‘delta’]

If inputs length = 1, you are not updating weights, it’s correct? You are updating only bias, because in hidden layer is only one neuron.

7. Tomasz November 21, 2016 at 1:34 am #

Hello. In method update_weight you are doing for j in range(len(inputs) – 1). If inputs lenght = 1, you aren’t updating weights. It’s correct? Hidden layer have one neuron so in output layer weights aren’t updated

• Jason Brownlee November 22, 2016 at 6:54 am #

Hi Tomasz,

The assumption here is that the input vector always contains at least one input value and an output value, even if the output is set to None.

You may have found a bug though when updating the layers. I’ll investigate and get back to you.

• Jason Brownlee January 3, 2017 at 10:17 am #

Thanks Tomasz, this was indeed a bug.

I have updated the update_weights() function in the above code examples.

8. Michael December 13, 2016 at 4:15 am #

Hi, Thanks for the tutorial, I’m doing a backpropagation project at the moment so its been really useful.

I was a little confused on the back-propagation error calculation function. Does “if i != len(network)-1:” mean that if the current layer isn’t the output layer then this following code is run or does it mean that the current layer is an output layer?

• Jason Brownlee December 13, 2016 at 8:08 am #

The line means if the index i is not equal to the index of the last layer of the network (the output layer), then run code inside the block.

9. Michael January 5, 2017 at 7:53 am #

I have another question.
Would it be possible to extend the code from this tutorial and create a network that trains using the MNIST handwritten digit set? using a input unit to represent each pixel in the image. I’m also not sure whether/how I could use feature extractors for the images.

I have a project where I have to implement the Backpropagation algorithm with possibly the MNIST handwritten digit training set.

I hope my question makes sense!

• Jason Brownlee January 5, 2017 at 9:42 am #

Sure Michael, but I would recommend using a library like Keras instead as this code is not written for performance.

Load an image as a long list of pixel integer values, convert to floats and away you go. No feature extraction needed for a simple MLP implementation. You should get performance above 90%.

10. Calin January 6, 2017 at 10:40 pm #

Hi Jason,

Great post!

I have a concern though:

In train_network method there are these two lines of code:

expected = [0 for i in range(n_outputs)]
expected[row[-1]] = 1

Couldn’t be the case that expected[row[-1]] = 1 will throw IndexError, as n_outputs is the size of the training set which is a subset of the dataset and row basically contains values from the whole dataset?

• Jason Brownlee January 7, 2017 at 8:37 am #

Hi Calin,

If I understand you correctly, No. The n_outputs var is the length of the number of possible output values.

Maybe put some print() statements in to help you better understand what values variables have.

• Calin January 7, 2017 at 9:48 pm #

Hmm..I ran the entire code (with the csv file downloaded from http://archive.ics.uci.edu/ml/datasets/seeds), added some breakpoints and this is what I got after a few iterations:

n_outputs = 168
row[-1] = 201

which is causing IndexError: list assignment index out of range.

• Adriaan January 11, 2017 at 4:27 am #

I’ve got the same error, That my list assignment index is out of range

• Jason Brownlee January 11, 2017 at 9:29 am #

Sorry to hear that, did you try running the updated code?

• Ivan January 16, 2017 at 10:28 am #

This is error of csv read. Try to reformat it with commas. For me it worked

• Jason Brownlee January 16, 2017 at 10:45 am #

What was the problem and fix exactly Ivan?

• Bob February 5, 2017 at 10:59 am #

The data file (http://archive.ics.uci.edu/ml/machine-learning-databases/00236/seeds_dataset.txt) has a few lines with double tabs (\t\t) as the delimiter — removing the double tabs and changing tabs to commas fixed it.

Thanks for the good article.

• Jason Brownlee February 6, 2017 at 9:42 am #

Thanks for the note Bob.

• Adriaan January 11, 2017 at 5:50 am #

I’ve had the same error at the ‘train_network’ function. Is your dataset fine? I’ve had some problems because the CSV file wasn’t loaded correctly due to my regional windows settings. I’ve had to adjust my settings and everything worked out alright.

11. Stanley January 8, 2017 at 3:15 pm #

Thanks for such a good article.

Just one question: in the equation “weight = weight + learning_rate * error * input”, why there is an “input”? IMO it should be: “weight = weight + learning_rate * error”?

• Jason Brownlee January 9, 2017 at 7:47 am #

The var names and explanation are correct.

The update equation is:

For the input layer the input are the input data, for hidden layers the input is the output of the prior layer.

Jason,

Thanks for the code and post.
Why is “expected” in expected = [0 for i in range(n_outputs)] initialized to [0,0] ?
Should not the o/p values be taken as expected when training the model ?
i.e for example in case of Xor should not 1 be taken as the expected ?

• Jason Brownlee January 13, 2017 at 9:16 am #

Expected is a one-hot encoding. All classes are “0” expect the actual class for the row which is marked as a “1” on the next line.

13. Michael January 19, 2017 at 3:44 am #

Hello, I have a couple more questions. When training the network with a dataset, does the error at each epoch indicate the distance between the predicted outcomes and the expected outcomes together for the whole dataset? Also when the mean accuracy is given in my case being 13% when I used the MNIST digit set, does this mean that the network will be correct 13% of the time and would have an error rate of 87%?

• Jason Brownlee January 19, 2017 at 7:38 am #

Hi Michael,

The epoch error does capture how wrong the algorithm is on all training data. This may or may not be a distance depending on the error measure used. RMSE is technically not a distance measure, you could use Euclidean distance if you like, but I would not recommend it.

Yes, in generally when the model makes predictions your understanding is correct.

14. Bernardo Galvão January 24, 2017 at 3:51 am #

Hi Jason,

in the excerpt regarding error of a neuron in a hidden layer:

“Where error_j is the error signal from the jth neuron in the output layer, weight_k is the weight that connects the kth neuron to the current neuron and output is the output for the current neuron.”

is the k-th neuron a neuron in the output layer or a neuron in the hidden layer we’re “on”? What about the current neuron, are you referring to the neuron in the output layer? Sorry, english is not my native tongue.

Bernardo

15. anonymous February 1, 2017 at 1:42 am #

It would have been better if recall and precision were printed. Can somebody tell me how to print them in the above code.

16. kehinde kolade February 6, 2017 at 8:29 pm #

Hello Jason, great tutorial, I am developer and I do not really know much about this machine learning thing but I need to extend this your code to incorporate the Momentum aspect to the training, can you please explain how I can achieve this extension?

• Jason Brownlee February 7, 2017 at 10:14 am #

Sorry, I don’t have the capacity to write or spell out this change for you.

My advice would be to read a good book on the topic, such as Neural Smithing: http://amzn.to/2ld9ds0

17. ibrahim February 18, 2017 at 2:21 am #

Hi Jason,
I have my own code written in C++, which works similar to your code. My intention is to extend my code to convolutional deep neural nets, and i have actually written the convolution, Relu and pooling functions however i could not begin to apply the backpropagation i have used in my shallow neural net, to the convolutional deep net, cause i really cant imagine the transition of the backpropagation calculation between the convolutional layers and the standard shallow layers existing in the same system. I hoped to find a source for this issue however i always come to the point that there is a standard backpropagation algorithm given for shallow nets that i applied already. Can you please guide me on this problem?

• Jason Brownlee February 18, 2017 at 8:42 am #

I”d love to guide you but I don’t have my own from scratch implementation of CNNs, sorry. I’m not best placed to help at the moment.

I’d recommend reading code from existing open source implementations.

18. matias February 22, 2017 at 3:34 pm #

Thank you, I was looking for exactly this kind of ann algorith. A simple thank won’t be enough tho lol

19. Manohar Katam February 26, 2017 at 3:40 pm #

Great one! .. I have one doubt .. the dataset seeds contains missing features/fields for some rows.. how you are handling that …

• Jason Brownlee February 27, 2017 at 5:49 am #

You could set the missing values to 0, you could remove the rows with missing values, you could impute the missing values with mean column values, etc.

Try a few different methods and see what results in the best performing models.

• Manohar Katam March 1, 2017 at 2:59 pm #

What if I have canonical forms like “male” or “female” in my dataset… Will this program work even with string data..

• Jason Brownlee March 2, 2017 at 8:11 am #

Hi Manohar,

No, you will need to convert them to integers (integer encoding) or similar.

20. Wissal ARGOUBI February 27, 2017 at 11:12 pm #

Great job! this is what i was looking for ! thank you very much .
However i already have a data base and i didn’t know how to make it work with this code how can i adapt it on my data
Thank you

21. Shweta Gupta March 5, 2017 at 4:37 am #

Thanks for such a great article..
I have one question, in update_weights why you have used weight=weight+l_rate*delta*input rather than weight=weight+l_rate*delta?

22. Sittha March 13, 2017 at 1:23 pm #

Thanks for a good tutorial.
I have some IndexError: list assignment index out of range. And I cannot fix it with comma or full-stop separator.

• Jason Brownlee March 14, 2017 at 8:11 am #

What is the full error you are getting?

Did you copy-paste the full final example and run it on the same dataset?

• Sittha March 24, 2017 at 3:36 am #

line 151 :
expected[row[-1]] = 1
IndexError : list assignment index out of range

• Jason Brownlee March 24, 2017 at 8:00 am #

Is this with a different dataset?

• Benji Weiss May 11, 2017 at 5:31 am #

if it is a different dataset, what do i need to do to not get this error

23. Karan March 16, 2017 at 6:26 pm #

The dataset that was given was for training the network. Now how do we test the network by providing the 7 features without giving the class label(1,2 or 3) ?

• Jason Brownlee March 17, 2017 at 8:27 am #

You will have to adapt the example to fit the model on all of the training data, then you can call predict() to make predictions on new data.

• Karan March 19, 2017 at 7:43 pm #

Ok Jason, i’ll try that and get back to you! Thank you!

24. Karan March 19, 2017 at 7:48 pm #

Just a suggestion for the people who would be using their own dataset(not the seeds_dataset) for training their network, make sure you add an IF loop as follows before the 45th line :
if minmax[i][1]!=minmax[i][0]

This is because your own dataset might contain same values in the same column and that might cause a divide by zero error.

25. Li Qun March 25, 2017 at 5:45 pm #

Thanks jason for the amazing posts of your from scratch pyhton implementations! i have learned so much from you!

I have followed through both your naive bayes and backprop posts, and I have a (perhaps quite naive) question:

what is the relationship between the two? did backprop actually implement bayesian inference (after all, what i understand is that bayesian = weights being updated every cycle) already? perhaps just non-gaussian? so.. are non-gaussian PDF weight updates not bayesian inference?

i guess to put it simply : is backpropagation essentially a bayesian inference loop for an n number of epochs?

I came from the naive bayes tutorial wanting to implement backpropagation together with your naive bayes implementation but got a bit lost along the way.

sorry if i was going around in circles, i sincerely hope someone would be able to at least point me on the right direction.

• Jason Brownlee March 26, 2017 at 6:11 am #

Great question.

No, they are both very different. Naive bayes is a direct use of the probabilities and bayes theorem. The neural net is approximating a mapping function from inputs and outputs – a very different approach that does not directly use the joint probability.

26. Chiraag March 26, 2017 at 10:10 pm #

How did you decide that the number of folds will be 5 ? Could you please explain the significance of this number. Thank You.

• Jason Brownlee March 27, 2017 at 7:54 am #

In this case, it was pretty arbitary.

Generally, you want to split the data so that each fold is representative of the dataset. The objective measure is how closely the mean performance reflect the actual performance of the model on unseen data. We can only estimate this in practice (standard error?).

27. Li Qun March 27, 2017 at 10:19 pm #

Dear Jason,

thank you for the reply! I read up a bit more about the differences between Naive Bayes (or Bayesian Nets in general) and Neural Networks and found this Quora answer that i thought was very clear. I’ll put it up here to give other readers a good point to go from:

https://www.quora.com/What-is-the-difference-between-a-Bayesian-network-and-an-artificial-neural-network

TL:DR :
– they look the same, but every node in a Bayesian Network has meaning, in that you can read a Bayesian network structure (like a mind map) and see what’s happening where and why.
– a Neural Network structure doesn’t have explicit meaning, its just dots that link previous dots.
– there are more reasons, but the above two highlighted the biggest difference.

Just a quick guess after playing around with backpropagation a little: the way NB and backprop NN would work together is by running Naive Bayes to get a good ‘first guess’ of initial weights that are then run through and Neural Network and Backpropagated?

• Jason Brownlee March 28, 2017 at 8:23 am #

Please note that a Bayesian network and naive bayes are very different algorithms.

28. Melissa March 27, 2017 at 10:54 pm #

Hi Jason,
Further to this update:

Update Jan/2017: Changed the calculation of fold_size in cross_validation_split() to always be an integer. Fixes issues with Python 3.

I’m still having this same problem whilst using python 3, on both the seeds data set and my own. It returns an error at line 75 saying ‘list object has no attribute ‘sum” and also saying than ‘an integer is required.’

Any help would be very much appreciated.
Overall this code is very helpful. Thank you!

• Jason Brownlee March 28, 2017 at 8:24 am #

Sorry to hear that, did you try copy-paste the complete working example from the end of the post and run it on the same dataset from the command line?

• Melissa March 28, 2017 at 9:29 am #

Yes I’ve done that, but still the same problem!

29. david March 29, 2017 at 6:16 am #

Hello jason,

please i need help on how to pass the output of the trained network into a fuzzy logic system if possible a code or link which can help understand better. Thank you

30. Aditya April 2, 2017 at 3:57 pm #

Awesome Explanation

31. Raunak Jain April 6, 2017 at 5:20 pm #

Hello Jason
I m getting list assignment index out or range error. How to handle this error?

• Jason Brownlee April 9, 2017 at 2:37 pm #

The example was developed for Python 2, perhaps this is Python version issue?

32. Marco April 6, 2017 at 9:37 pm #

Thanks but I think python is not a good choice…

• Jason Brownlee April 9, 2017 at 2:40 pm #

I think it is a good choice for learning how backprop works.

What would be a better choice?

33. Agrawal April 6, 2017 at 9:38 pm #

Hey, Jason Thanks for this wonderful lecture on Neural Network.

As I am working on Iris Recognition, I have extracted the features of each eye and store it in .csv file, Can u suggest how further can I build my Backpropagation code.
As when I run your code I am getting many errors.
Thank you

34. Jack April 7, 2017 at 3:42 pm #

Could you please convert this iterative implementation into matrix implementation?

35. Jk April 12, 2017 at 5:04 am #

Hi Jason,

In section 4.1 , may you please explain why you used ### inputs = row[:-1] ### ?

Thanks

• Jason Brownlee April 12, 2017 at 7:58 am #

Yes. By default we are back-propagating the error of the expected output vs the network output (inputs = row[:-1]), but if we are not the output layer, propagate the error from the previous layer in the network (inputs = [neuron[‘output’] for neuron in network[i – 1]]).

I hope that helps.

• JK April 13, 2017 at 3:59 am #

Thanks for your respond. I understand what you said , the part I am no understanding is the [:-1] . why eliminating the last list item ?

• Jason Brownlee April 13, 2017 at 10:10 am #

It is a range from 0 to the second last item in the list, e.g. (0 to n-1)

36. Prem Puri April 12, 2017 at 8:18 pm #

In function call, def backward_propagate_error(network, expected):
how much i understand is , it sequentially pass upto
if i != len(network)-1:
for j in range(len(layer)):
error = 0.0
for neuron in network[i + 1]:
error += (neuron[‘weights’][j] * neuron[‘delta’])
My question is which value is used in neuron[‘delta’]

• Jason Brownlee April 13, 2017 at 10:01 am #

delta is set in the previous code block. It is the error signal that is being propagated backward.

37. Prem Puri April 14, 2017 at 3:20 am #

Thanks very much!

38. youssef oumate April 26, 2017 at 4:53 pm #

Hi Jason

Thank you very much for this awesome implementation of neural network,
I have a question for you : I want to replace the activation function from Sigmoid
to RELU . So, what are the changes that I should perform in order to get
correct predictions?

• Jason Brownlee April 27, 2017 at 8:34 am #

I think just a change to the transfer() and transfer_derivative() functions will do the trick.

• youssef oumate April 27, 2017 at 10:17 am #

Awesome !

Thank you so much

39. Yahya Alaa April 30, 2017 at 2:38 am #

Hi Jason,
Thank you very much for this wonderful implementation of Neural Network, it really helped me a lot to understand neural networks concept,

n_inputs = len(dataset[0]) – 1
n_outputs = len(set([row[-1] for row in dataset]))
network = initialize_network(n_inputs, 2, n_outputs)
train_network(network, dataset, 0.5, 20, n_outputs)

What do n_inputs and n_outputs refer to? According to the small dataset used in this section, is n_inputs only 2 and n_outputs only 2 (0 or 1) or I am missing something?

• Jason Brownlee April 30, 2017 at 5:31 am #

Input/outputs refers to the number of input and output features (columns) in your data.

• Yahya Alaa May 3, 2017 at 1:42 pm #

Is the program training the network for 500 epochs for each one of the k-folds and then testing the network with the testing data set?

• Jason Brownlee May 4, 2017 at 8:02 am #

Hi Yahya,

5-fold cross validation is used.

That means that 5 models are fit and evaluated on 5 different hold out sets. Each model is trained for 500 epochs.

I hope that makes things clearer Yahya.

• Yahya Alaa May 4, 2017 at 8:17 am #

Yes you made things clear to me, Thank you.
I have two other questions,
How to know when to stop training the network to avoid overfitting?
How to choose the number of neurons in the hidden layer?

• Jason Brownlee May 5, 2017 at 7:27 am #

You can use early stopping, to save network weights when the skill on a validation set stops improving.

The number of neurons can be found through trial and error.

• Yahya Alaa May 6, 2017 at 8:48 am #

I am working on a program that recognizes handwritten digits, the dataset is consisting of pictures (45*45) pixels each, which is 2025 input neurons, this causes me a problem in the activation function, the summation of (weight[i] * input[i]) is big, then it gives me always a result of (0.99 -> 1) after putting the value of the activation function in the Sigmoid function, any suggestions?

• Jason Brownlee May 7, 2017 at 5:31 am #

I would recommend using a Convolutional Neural Network rather than a Multilayer Perceptron.

40. morok April 30, 2017 at 3:56 am #

In section 3.2. Error Backpropagation, where did output numbers came from for testing backpropagation

‘output’: 0.7105668883115941
‘output’: 0.6213859615555266
‘output’: 0.6573693455986976

Perhaps from outputs on test forward propagation [0.6629970129852887, 0.7253160725279748] taking dd -> derivative = output * (1.0 – output), problem is they don’t match, so I’m a bit lost here…

thanks!

Awesome article!!!

• Jason Brownlee April 30, 2017 at 5:34 am #

In that example, the output and weights were contrived to test back propagation of error. Note the “delta” in those outputs.

41. Umamaheswaran May 8, 2017 at 9:49 pm #

Hi Jason,

I am using the MNIST data set to implement a handwritten digit classifier. How many training examples will be needed to get a perfomance above 90%.

42. Huyen May 9, 2017 at 6:32 pm #

Hi Jason,

Your blog is totally awesome not only by this post but also for the whole series about neural network. Some of them explained so much useful thing than others on Internet. They help me a lot to understand the core of network instead of applying directly Keras or Tensorflow.

Just one question, if I would like to change the result from classification to regression, which part in back propagation I need to change and how?

• Jason Brownlee May 10, 2017 at 8:46 am #

Thanks Huyen.

You would change the activation function in the output layer to linear (e.g. no transform).

43. TGoritsky May 12, 2017 at 12:41 am #

Hi Jason,

I am playing around with your code to better understand how the ANN works. Right now I am trying to do predictions with a NN, that is trained on my own dataset, but the program returns me one class label for all rows in a test dataset. I understand, that normalizing dataset should help, but it doesn`t work (I am using your minmax and normalize_dataset functions). Also, is there a way to return prediction for one-dimensional dataset?
Here is the code (sorry for lack of formatting):
def make_predictions():
dataset = [[29,46,107,324,56,44,121,35,1],
[29,46,109,327,51,37,123,38,1],
[28,42,107,309,55,32,124,38,1],
[40,112,287,59,35,121,36,1],
[27,43,129,306,75,41,107,38,1],
[28,38,127,289,79,40,109,37,1],
[29,37,126,292,77,35,100,34,1],
[30,40,87,48,77,51,272,80,2],
[26,37,88,47,84,44,250,80,2],
[29,39,91,47,84,46,247,79,2],
[28,38,85,45,80,47,249,78,2],
[28,36,81,43,76,50,337,83,2],
[28,34,75,41,83,52,344,81,2],
[30,38,80,46,71,53,347,92,2],
[28,35,72,45,64,47,360,101,2]]
network = [[{‘weights’: [0.09640510259345969, 0.37923370996257266, 0.5476265202749506, 0.9144446394025773, 0.837692750149296, 0.5343300438262426, 0.7679511829130964, 0.5325204151469501, 0.06532276962299033]}],
[{‘weights’: [0.040400453542770665, 0.13301701225112483]}, {‘weights’: [0.1665525504275246, 0.5382087395561351]}, {‘weights’: [0.26800994395551214, 0.3322334781304659]}]]
# minmax = dataset_minmax(dataset)
# normalize_dataset(dataset, minmax)
for row in dataset:
prediction = predict(network, row)
print(‘Expected=%d, Got=%d’ % (row[-1], prediction))

44. Tomo May 18, 2017 at 6:22 pm #

Hi Jason!
In the function “backward_propagate_error”, when you do this:

neuron[‘delta’] = errors[j] * transfer_derivative(neuron[‘output’])

The derivative should be applied on the activation of that neuron, not to the output . Am I right??

neuron[‘delta’] = errors[j] * transfer_derivative(activate(neuron[‘weights’], inputs))

And inputs is:
inputs = row[-1]
if i != 0:
inputs = [neuron[‘output’] for neuron in self.network[i-1]]

Thank you! The post was really helpful!

45. Tina May 26, 2017 at 3:49 am #

Hello Jason!

This is a very interesting contribution to the community 🙂
Have you tried using the algorithm with other activation functions?
I tried with Gaussian, tanh and sinx, but the accuracy was not that high, so I think that I omitted something. What I altered were the activation functions and the derivatives. Is there something else that needs to be changed?

• Jason Brownlee June 2, 2017 at 11:49 am #

Sigmoid was the defacto standard for many years because it performs well on many different problems.

Now the defacto standard is ReLU.

• Manu June 6, 2017 at 8:50 pm #

Sigmoid and ReLU are transfer functions right ?
Activation function is just the sum of all weights and inputs

46. vishwanathan May 27, 2017 at 8:08 pm #

Thanks for the great post. Here is some observation that I am not able to understand. In the back ward propagate you are not taking all the weights and only considering the jth. Can you kindly help understand. I was under the impression that the delta from output is applied across all the weights,
for neuron in network[i + 1]:
error += (neuron[‘weights’][j] * neuron[‘delta’])

• vishwanathan May 27, 2017 at 8:14 pm #

I understand you do not want to take in the bias weight hence the exclusion of the last weight in neuron. I kind of get stumped on bias.

47. vishwanathan May 27, 2017 at 9:12 pm #

Thanks for the great article. In the backward propagate, the delta value is applied for each weight across the neuron and the error is summed. I am curious why is the delta not applied to individual weights of the neuron and the error summed for that neuron. Can you please clarify?

48. Josue May 29, 2017 at 3:12 am #

Why don’t you split the data into TrainData and TestData, like 80% of the dataset for training and 20% for testing, because if you train with 100% of rows of the dataset and then test some rows of the dataset the accuracy will be good . But if you put new data on the seeds.csv the model will work with less accuracy, Right?

49. Josue May 29, 2017 at 11:08 am #

Thanks for the post! I have a question about cross-validation. The dataset of seeds is perfect for 5 folds but for a dataset of 211? I’ll have uniformly sized subset right? (211/5) Can you give me a suggestion how I could handle that ?

• Jason Brownlee June 2, 2017 at 12:20 pm #

One way is that some records can be discarded to give even sized groups.

50. Sebastián May 30, 2017 at 9:35 am #

Thanks so much for the tutorial. It was really helpful!

51. Manu June 10, 2017 at 9:00 pm #

Hello Jason,

any advice on how to handle multi-classifier problems when the classes have high cardinality ?
I’m thinking about input data of search engines linked to choosen urls.

• Jason Brownlee June 11, 2017 at 8:25 am #

Ouch, consider modeling it as regression instead (e.g. a rating or recommender system).

• Manuel June 13, 2017 at 1:17 am #

Ok thank you very much Jason.
But it wont work with searches unseen by the algorithm.
I red something in the books “Programming collective intelligence” about a neural net from scratch for this king of problem but I don’t understang how it works for the moments…

• Jason Brownlee June 13, 2017 at 8:23 am #

Consider focusing on one measure/metric that really matters in your domain, then try a suite of framings of the problem and different algorithms to get a feeling for what might work best.

52. Yash June 18, 2017 at 6:21 pm #

I am not able to understand the above code.So, I request you to explain me the above code

• Jason Brownlee June 19, 2017 at 8:43 am #

Which part do you not understand exactly?

53. Tathagat June 21, 2017 at 3:20 pm #

Hey Jason..am a novice in machine learning..have a small question…how can I track the timesteps involved in the algorithm with accordance with the code?

• Jason Brownlee June 22, 2017 at 6:04 am #

What do you mean by time steps?

54. bazooka June 29, 2017 at 6:52 am #

Hi, Jason. I am so confused, in the result, why there are 4 set of [output,weight,delta]

like this:
[{‘output’: 0.9999930495852168, ‘weights’: [0.9315463130784808, 1.0639526745114607, 0.9274685127907779], ‘delta’: -4.508489650980804e-09}, {‘output’: 0.9992087809233077, ‘weights’: [-2.4595353900551125, 5.153506472345162, -0.5778256160239431], ‘delta’: 1.940550145482836e-06}]
[{‘output’: 0.01193860966265472, ‘weights’: [2.3512725698865053, -8.719060612965613, 1.944330467290268], ‘delta’: -0.0001408287858584854}, {‘output’: 0.988067899681387, ‘weights’: [-2.2568526798573116, 8.720113230271012, -2.0392501730513253], ‘delta’: 0.0001406761850156443}]

after the backpropagation we find the optimal weights to get minimum error, what does these 4 group means?
E

• Jason Brownlee June 29, 2017 at 7:48 am #

That is the internal state of the whole trained network.

55. hassan June 29, 2017 at 7:30 am #

hi Jason
thanks for your code and good description here, i like it so much.
i run your example code and encounter with an error same others whom left note here
the error is:
expected[row[-1]] = 1
IndexError: list assignment index out of range

how i can fix this error?

• Jason Brownlee June 29, 2017 at 7:49 am #

The code was written for Python 2.7, confirm that this is your Python version.

Also confirm that you have copied the code exactly.

56. Jerome July 5, 2017 at 9:20 pm #

Dear Jason,

i have this question about Back Propagate Error

1- derivative sigmoid = output * (1.0 – output)
That is ok

2- error = (expected – output) * transfer_derivative(output)
Ok but it also means that error == 0 for output = 1 whatever the expected is because transfer_derivative(1) ==0

So, whatever the expected , error is nil if output is 1 …
Is there something rotten here?

Thanks

Jerome

57. wddddds July 10, 2017 at 10:01 pm #

Thank you Jason, It’s a great tutorial and really helpful for me!

But I have to say that trying to reimplement your code strongly increased my ability of debugging 🙂

58. Victor July 17, 2017 at 7:50 pm #

Hi Jason,

Thanks for sharing your code. I’m a PhD candidate in machine learning, and I have a doubt about the weights update in section 4.1:

weight = weight + learning_rate * error * input

Should not it be as follows?

weight = weight – learning_rate * error * input

Thanks again for sharing this.

Regards,
Victor.

• Víctor August 4, 2017 at 11:07 pm #

I didn’t say anything, my mistake in understanding.

Thanks again for sharing your work.

59. vishnu priya July 22, 2017 at 4:26 pm #

Hi..
Thanks for ur coding. It was too helpful. can u suggest me how to use this code for classifying tamil characters. i have tried in cnn and now i need to compare the result with bpn. can u pls suggest me.

thank you

60. vishnu priya July 23, 2017 at 4:06 pm #

Thank you sir. With this tutorial i have implemented cnn sir. but for BPN i am getting error rate 687.203 sir. i dnt know what to do sir. can u help me sir.

Thank you

61. Vishnupriya July 24, 2017 at 4:53 pm #

Classification of Tamil characters sir. I have 144 different classes. I have taken 7 glcm features of each character and I need to train this features in backpropagation and predict the character to which class it belongs.

62. codeo July 26, 2017 at 5:37 pm #

Hi, so I wasn’t following this tutorial when implementing my neural network from scratch, and mine is in JavaScript. I just need help with the theory. How do I calculate the error for each node in the net so that I can incrementally change the weights? Great tutorial btw

• codeo July 26, 2017 at 6:38 pm #

Hahaha nevermind, it was my code
Multidimensional arrays and stuff boggle the mind hah

• Jason Brownlee July 27, 2017 at 7:56 am #

Glad to hear you worked it out.

63. PRABHAKARAN M July 31, 2017 at 4:31 pm #

[ 6.38491205 5.333345 4.81565798 5.43552204 9.96445304 2.57268919 4.07671018 1.5258789 6.19728301 0 1 ]
Dear sir,
the above mentioned numerical values are extracted from the dental x-ray image using gray level co occurrence matrix [10 inputs and 1 output]. This dataset is used as a input for BPN classifier. whether the same data set as[.csv] file can be used as the input for DEEP Convolutional Neural Network technique ? and can i get the output as image ? for example if i give the dental x ray images as numerical values i have to get the caries affected teeth as the output for the given dataset.

64. PRABHAKARAN M July 31, 2017 at 4:32 pm #

can i get the example code for dental caries detection using deep Convolutional Neural Network for the given dataset as x ray images.

• Jason Brownlee August 1, 2017 at 7:52 am #

I do not have sample code for this problem, sorry.

65. John August 1, 2017 at 3:26 am #

Very nice explanation, thank you.
I have some questions.

1) weight = weight + learning_rate * error * input

Do I really need to multiply it with input ? For example here http://home.agh.edu.pl/~vlsi/AI/backp_t_en/backprop.html they dont it multiply it with input. At least, I think that…

2) Is your method same as in http://home.agh.edu.pl/~vlsi/AI/backp_t_en/backprop.html?
i think yes, but again, Im not sure and Im confused by that input multiplication.

3) What is exactly loss function in your example (I usually found some derivations of loss (cost ?) function (in other explanations), not transfer function derivation)? Im actually very confused by notation which I find around …

4) momentum and weight decay. In your example, you can implement them that you substract calculated decay and add calculated momentum (to weight update) ? Again, I found forms which substract both and weight update as w + deltaW, so again I’m mega confused by notation for backpropagation which I found…

Sorry for dumb questions, … math is not my strong side, so many things which can be inferred by math sense are simply hidden for me.

• John August 1, 2017 at 3:30 am #

*substract both and weight update as w + deltaW, so again

I found above sentence as nonsense, must be side effect of my confusion …

• Jason Brownlee August 1, 2017 at 8:12 am #

Hang in there.

Pick one tutorial and focus on it. Jumping from place to place will make things worse for sure.

• Jason Brownlee August 1, 2017 at 8:10 am #

Hi John, good questions.

According to my textbook, yes.
I can’t speak for random sites on the internet sorry.

Loss is prediction error. You can change this to other forms like MAE or MSE.

No decay or momentum in this example. Easy to add if you want. There are many ways to dial in the learning process. No hard and fast rules, just some norms that people reuse.

66. Parminder Kaur August 6, 2017 at 7:50 pm #

A VERY GOOD TUTORIAL SIR…
Sir i am implementing remote sensed image classification using BPN neural network using IDL.
I am not finding good resources on constructing features for input dataset and also number of hidden layers and number of neurons in hidden layer.
Any resources you know, can help me?

Thanks

• Jason Brownlee August 7, 2017 at 8:41 am #

The CNN will perform feature extraction automatically, you could explore using different filters on the data to see if it helps the network.

The number of layers and neurons/filters per layer must be found using trial and error. It is common to copy the designs from other papers as a starting point.

I hope that helps.

67. pero August 9, 2017 at 1:11 am #

Nice tutorial, very clean and readable code. =) thank you!

68. Vatandas August 15, 2017 at 3:28 am #

1. I expect that this code is deep learning (many hidden layer) but not. One sentence is easy (“you can add more hidden layer as explained”) but to do is not as easy as you said.

2. I think your code is wrong.
neuron[‘delta’] = errors[j] * transfer_derivative(neuron[‘output’])
but
Error = Target – ActivatedOutputNode
Delta = Error * Derivative(NONActivatedOutputNode)

I mean you use the same ‘output’ variable both error and delta. But in error it must be activated one, in delta it must be NONactivated one.