How to use the custom neural network function in the MATLAB Neural Network Toolbox

Question

I'm trying to create the neural network shown below. It has 3 inputs, 2 outputs, and 2 hidden layers (so 4 layers altogether, or 3 layers of weight matrices). In the first hidden layer there are 4 neurons, and in the second hidden layer there are 3. There is a bias neuron going to the first and second hidden layer, and the output layer.

I have tried using the "create custom neural network" function in MATLAB, but I can't get it to work how I want it to.

This is how I used the function

net1=network(3,3,[1;1;1],[1,1,1;0,0,0;0,0,0],[0,0,0;1,0,0;0,1,0],[0,0,0])
view(net1)

And it gives me the neural network shown below:

As you can see, this isn't what I want. There are only 3 weights in the first layer, 1 in the second, 1 in the output layer, and only one output. How would I fix this?

Thanks!

---

Just to clarify how I want this network to work:

• The user will input 3 numbers into the network.
• Each one of the 3 inputs is multiplied by 4 different weights, and then these numbers are sent to the 4 neurons in the first hidden layer.
• The bias node acts the same as one of the inputs, but it always has a value of 1. It is multiplied by 4 different weights, and then sent to the 4 neurons in the first hidden layer.
• Each neuron in the first hidden layer sums the 4 numbers going into it, and then passes this number through the sigmoid activation function.
• The neurons in the first hidden layer then output 4 numbers that are each multiplied by 3 different weights, and sent to the 3 neurons in the second hidden layer.
• The bias node going to the second hidden layer works the same as the first bias node
• Each neurons in the second hidden layer sums up the 5 numbers going into it and passes it through the sigmoid activation function.
• The neurons in the second layer then output two numbers that are again multiplied by weights and go to each of the outputs
• The output layer also sums all of its inputs, including its bias input, and then passes this through the sigmoid activation function to get the final two values.

Answer 1

After some time playing around I've figured out how to do it. The code I needed to use is:

net = newff([0 1; 0 1; 0 1],[4,3 2],{'logsig','logsig','logsig'})
view(net)

This creates the network I was looking for.

I was originally mistaken about the matlab representation of neural networks. The green arrows show the path of all of the numbers, not just a single number.