'tansig' is the correct default transfer function for both hidden and output layers in patternnet because the default normalization for input and output are mapminmax. It is still correct if you override with mapstd.
However, if you remove the output normalization and the target matrix contains unit vector columns with a single 1, then logsig and softmax are appropriate for estimating input-conditional class posterior probabilities. Both yield consistent estimates. However, the softmax probabilities estimates always sum to 1. If you are using logsig and the unit sum property is desirable, you can always divide the logsig estimates by their sum.
If you are still having problems, post your code.
Trainmask is a vector with ones at the training set indices and zeros otherwise. Multiplication using .* singles out training set inputs and outputs. Similarly for the other masks.
Search ANSWERS and NEWSGROUP for sample code.
Hope this helps.
Thank you for formally accepting my answer.
Greg
