Selecting neural network for a particular classification problem

Hi,

I've been trying to fit a data with neural network, and optimizing the testing accuracy. The input feature space is 1,990 x 40,000 (1,990 is reduced feature space, further reduction is not possible) which I've to classify in one of the 25 classes.

The data plot (train and test) is as shown below:

The issue is training accuracy of network is pretty much good. But testing accuracy is not.

So far, I've tried:

• Multi layer perceptrons (patternet) with:

1. Varying learning rate.
2. Varying number of layers (upto 8), with upto 200 neurons in each layer (further neuron increase causes overfitting).
3. Adding different regularization values.
4. Different training algorithms (traingdx, traingda, trainscg).

• Radial basis networks, with many spread constants.

but no luck so far.

Any suggestions on network selection / data preprocessing for such cases before feeding it to neural network? Thanks in advance !!