While building a training set I come across an issue.. different sized samples.

Hello.

I am building a training set for a neural network toolbox. In this dataset, there are 4 classes. Two of the classes contain 91 samples with 181 elements in each sample. The other two classes contain 91 samples with 91 elements in each sample. My objective is to stack these classes on top of each other in one big matrix, but I can't do that because of the size difference.

I thought about making all datasets 181*91 matrices with NAN to fill empty slots, but I am wondering if that will make training too easy for the network?

To explain further, I am afraid that would automatically teach the network that if the dataset contains a lot of NANs, then it must be one of two classes, rather than learning characteristics of the actual dataset.

Ideas? Thanks!

[Greg Heath]

Your use of the terms "sample" and "elements" is confusing and misleading.

You need N examples of I-dimensional input vectors paired with N O-dimensional output vectors so that

 [ I N ] = size(input)

 [ O N ] = size(target)

For classifiers O is the number of classes.

Is your 181 a vector dimension or the number of vectors?

Greg