"Carlos Aragon" wrote in message <k5n37h$ier$1@newscl01ah.mathworks.com>...
> I'm building a feedforwardnet like this:
>
> (..)
> P=[V';ia';w'];
> T=[tq'];
> net=feedforwardnet([5 25],'trainbr');
> (..)
>
> How could i train this neural net for more then one group '[V';ia';w']' ? How is the matlab structure to perform this kind of training?
>
> Note that 'P' in this case is a 10006x3 matrix that i extract from a motor model.
The issue here is that after training with set1, the weights will forget set1
while they are learning set 2. There are a variety of ways to mitigate forgetting.
The best is to use a modication of NEWRB that allows the input of an initial
hidden layer. Then
1. After training with set1, use those weights as initial weights for training with set2 + set1.
2. After training with set1, use those weights as initial weights for training with set2 and a "characteristic subset" of set1. The drawback is how to define that characteristic.
The reason this works is that each hidden node basis function has local region of influence and a 1-to-1 correspondence with a previous worst classified training vector.
Hope this helps.
Greg![]()
> I'm building a feedforwardnet like this:
>
> (..)
> P=[V';ia';w'];
> T=[tq'];
> net=feedforwardnet([5 25],'trainbr');
> (..)
>
> How could i train this neural net for more then one group '[V';ia';w']' ? How is the matlab structure to perform this kind of training?
>
> Note that 'P' in this case is a 10006x3 matrix that i extract from a motor model.
The issue here is that after training with set1, the weights will forget set1
while they are learning set 2. There are a variety of ways to mitigate forgetting.
The best is to use a modication of NEWRB that allows the input of an initial
hidden layer. Then
1. After training with set1, use those weights as initial weights for training with set2 + set1.
2. After training with set1, use those weights as initial weights for training with set2 and a "characteristic subset" of set1. The drawback is how to define that characteristic.
The reason this works is that each hidden node basis function has local region of influence and a 1-to-1 correspondence with a previous worst classified training vector.
Hope this helps.
Greg