How can you classify multidimensional data to 2 classes using neural networks

I have an input matrix (24,000x64) in which there are 100 feature samples of 240 trials across 64 channels. I would like to classify all trials each consisting of 100 samples to 2 classes.

What would be an efficient way to do this using neural networks?

How should I present input into the network?

What topology should I use?

How to prepare the target vector? Thank you

Answer 1

[Walter Roberson]

You need to reshape to a 240 * (100*64) matrix (or the transpose of it). You do not actually have 100 sample features: you have 100*64 sample features because the features are different for every channel. Your target vector would be length 240 (or possibly 240 x 2 depending which network you train with.)

For information on using networks for classifications see http://www.mathworks.com/help/nnet/gs/classify-patterns-with-a-neural-network.html

[Siddharth Talwar]

Thank you, this seems to be fine. But, I do want to ask, how can I present 100*64 samples into a network. Can one make a neural network in which each input neuron takes in multiple samples in one iteration. For example, 64 input neurons taking in 100 samples each, and processing and updating the hidden network continuously and classifying a trial into a class, as 6400 inputs do not make much sense. Can there be any other approach to it?

[Walter Roberson]

6400 inputs is fine as far as NN care.

You can use cell arrays and implicit time series; see the cell array format description at http://www.mathworks.com/help/nnet/ref/train.html . However, the implication of using those is that the input for each of the channels is not simultaneously present, which does not sound like what you are dealing with: it sounds like you are reading 64 different probes each of which has 100 associated values at the same time. I could be wrong though as you did not really explain that part.

If 6400 input values seems to be too much then you should be considering feature reduction.

[Greg Heath]

240 vectors only define, at most, a 239-dimensional space.

Therefore you should reduce the input dimensionality to less than or equal to 239.

Hope this helps.

Greg

[Siddharth Talwar]

Thank you for you help! :)