Initial values used in nftool

Hi,

I am using nftool to generate a neural network and using lmtrain. I want to know how do I know what the values are for the following variables:

• initial weights
• initial biases
• initial mu-value
• mu increase factor
• mu decrease factor

Furthermore, how can I change the above parameters in nftool?

For background: I want to investigate the performance of the network using different hidden neurons, but can only actually compare the networks if the initial conditions are identical.

Thanks in advance.

Answer 1

[Walter Roberson]

The default values are shown in https://www.mathworks.com/help/deeplearning/ref/trainlm.html

Initial weights and biases are random.

To be able to compare runs you need to use rng() to set the random seed.

Furthermore, how can I change the above parameters in nftool?

You cannot. nftool is a convenience interface for people who do not need fine control. If you need fine control, then you need to set up everything and call trainlm() yourself.

[David Ho]

If you would like fine control of the network parameters without starting from scratch, you can use the "Generate Code" option in the Neural Network Fitting tool to generate a script that does exactly the same training. You can then modify the parameters of the network to the values you would like.

[Walter Roberson]

Good point, @David Ho

[Marlise Jacobs]

@Walter Roberson Thank you so much. How do I integrate rng()? Currently, I call the nftool function from the command window, and then the app immediately opens and runs. So how do I call the rng-function before the nftool function runs?

[Walter Roberson]

rng(12345); nftool()

Then run one fitting (only!!) and quit nftool; repeat as needed.

But better yet would be to follow David's suggestion to use the Generate Code feature to convert the run into MATLAB code that you could then edit, including putting in an rng() call, or initializing the weights and biases yourself.