There is something wrong with the learning process of networks with more than 1 hidden layer.

I have a small dataset, 500 lines with 10 inputs and 2 outputs each (all inputs/outputs are already between 0 and 1).

When I run a simple NN with 1 hidden layer, it gets highly optimized in a few epochs and error drops gradually until it reaches a plateau (like expected).

Ex:
var myPerceptron = new Architect.Perceptron(10, 10, 2);
var myTrainer = new Trainer(myPerceptron);
var trainingSet = [
{ input: [0,0,0,0,0,0,0,0,0,0], output: [0,1]},
{ input: [1,1,1,1,1,1,1,1,1,1], output: [1,0]},
...
}

myTrainer.train(trainingSet, {
rate: .1,
iterations: 10000,
error: .0005,
shuffle: true,
log: 100,
cost: Trainer.cost.CROSS_ENTROPY
});

But if I change it's structure to 2 hidden layers, it never gets optimized, even with hundreds or thousands of epochs.

var myPerceptron = new Architect.Perceptron(10, 10, 10, 2); // small change here
var myTrainer = new Trainer(myPerceptron);
var trainingSet = [
{ input: [0,0,0,0,0,0,0,0,0,0], output: [0,1]},
{ input: [1,1,1,1,1,1,1,1,1,1], output: [1,0]},
...
}

myTrainer.train(trainingSet, {
rate: .1,
iterations: 10000,
error: .0005,
shuffle: true,
log: 100,
cost: Trainer.cost.CROSS_ENTROPY
});

There is something wrong with the learning process when you add extra hidden layers. One extra layer should, at worst, make the learning process slower, but not interrupt it at all.

I can add the source-code or input/output data if needed.