Hello,
I have tried to construct coder-decoder with patternet in order to reduce 1x400 vector but without any luck. What I trying to do: Create a neural network with 400 inputs neurons and 400 output neurons and X hidden neurons in the last hidden layer (but consider now just one hidden layer). It should be kind of decimation or PCA of my vector. The goal is to find the number of hidden neurons which will satisfy my error requirements to output data.
What I have done is I took 400x20000 matrix as input matrix and the same matrix as targets. When I start training on patternet it will print me that Performance goal is set and the tool didn't do any epochs at all. Performance function was set to cross-entropy and training algorithm SCG. I tried to use MSE as cost fcn. This actually started training, but performance did not lower over epochs at all.
Do you have any suggestions how could I accomplish that?
Thanks
