training a backprop network using odor world data

[kimira]

Hi,

Thanks for putting up Sim brain. it is a very useful tool to teach neural nets.

I am trying to play with training backprop networks using odor world. backprop 3-5-3.sim and backprop auto-3.sim are the demos. But they are networks that are already trained, and we can just run them. How was the training acheived? using what data?

Thanks

[jyoshimi]

Hello, and thanks for your query.

There is some documentation on this:

http://www.simbrain.net/Documentation/docs/Pages/Network/network/backpropnetwork.html
http://www.simbrain.net/Documentation/docs/Pages/Network/network/trainingFile.html

On the first page, see the bottom under "training."

The overall process is a bit counterintuitive (relative to the overall visual emphasis of Simbrain).   The basic method is to right click on the backprop tab, select "Train Backprop Network" and then to specify input and output files.  You have to manually create those files outside of Simbrain.  All the training happens in the dialog.   (In version 3 we are changing things so that training networks is much more flexible and intuitive).

Most of the training files I've used are in the folder {Simbrain_Home}/simulations/networks/bp/training.

To train a network to respond in some way to OdorWorld objects, you should build your Odor World environment and Simbrain network, see how the OdorWorld objects affect the input layer of the network, and then manually create your training files.   Then you can use those files to train the network.

Let me know if this helps,

- Jeff