How to generate output results given input variables?

classic Classic list List threaded Threaded
4 messages Options
EL
Reply | Threaded
Open this post in threaded view
|

How to generate output results given input variables?

EL
Hi Simbrain,

How do I generate empirical results given input variables?

Say I have 6 types of variable and each variable has 30 set of values.

Thanks.
Reply | Threaded
Open this post in threaded view
|

Re: How to generate output results given input variables?

jyoshimi
Administrator
I'm not sure what you mean by "empirical results".  Can you elaborate?  It sounds like you might want to do something like backprop to associate input values with output values.  For the variables if they are numerical they should be scaled or centered to lie in the range -1,1 or 0,1 and if they are not numerical they need to be coded numerically.   Then depending on what you want to do you choose the appropriate network type.

- Jeff
EL
Reply | Threaded
Open this post in threaded view
|

Re: How to generate output results given input variables?

EL
Hi Jeff,

Yes I would like to generate results based on the set of numerical variables input (input layer / layer 1) using backprop.

I have watched the Simbrain youtube videos and understand how to input the values. However, what should be the normal target values (output layer / layer 3) to be inputted, so that I can obtain results in the hidden layer / layer 2 during testing at the validate input data tab.

Pardon me as I am new to the Simbrain program.

Thanks.
Reply | Threaded
Open this post in threaded view
|

Re: How to generate output results given input variables?

jyoshimi
Administrator
Hi again,

I still don't totally understand, but I think we're getting closer.  The first thing to note is that everything will revolve around the input and output layers.   The hidden layer probably won't be a focus for your task.  Then what you have to do train the network on input / target pairs that you already have.  Once you've done that, you can test the network using new inputs for which you don't know the output.

So suppose you have this already

(1,0,0) --> (0,1)
(0,1,0) --> (1,0)

These are input / target pairs you already know.  So you train with those.  Then you can test on a new inputs, like these

(.5,0,.1)
(0,0,.2)

And see what it produces.

I'm not sure this helps.   But basically when you ask "what should be the normal target values (output layer / layer 3) to be inputted" that's based on whatever information you already have.

Think of concrete examples.  In a number recognition task, you'd show the network a bunch of number images and then the targets would be what the numbers should be classified as.   Or if you wanted to predict house price from square feet (Andrew Ng's example) you'd give the network a bunch of examples of square feet --> price pairs you already know.  So the targets come from data you already have.

- jeff