Overfitting

[Michael Paulse]

Hi Jeff

Thank you a great program and for your contribution to this forum.
I am new to NN although I fully understand the process, feedforwarding, back propagation and the math. I am concerned that since Simbrain is fully automated it may be overfitting to find solutions. How can I be sure that my outcome is indeed the "best" solution? Is there a feature in Simbrain to do this?

Thanks,
Mike

[jyoshimi]

Hi Mike good question.   Simbrain 3 uses a plain vanilla (and unoptimized) instantiation of backprop.  It's got none of the bells and whistles of modern backprop or machine learning.  A common way of preventing overfitting is by using regularization terms, which is one of the many things Simbrain 3's backprop lacks.

There might be things you can do in Simbrain e.g. from a script, that would allow you to address overfitting, but I'd have to give it some thought, and time is pretty short at my end right now.

Given how long it may take for us to release Simbrain 4, we have toyed with doing a pass on backprop in Simbrain 3, but we'd really need help, so if anyone is interested, let me know!

- Jeff

[Michael Paulse]

I appreciate the reply. Mike