It's very likely you have used a command for pattern recognition, something like
Your NN is meant to be used for classification.
Generally speaking, by means of training given inputs and targets, NN finds values of weights and biases to find the best approximation of mapping inputs into outputs, that is something like non-linear regression. The more complex structure, the better approximation but also greater tendency to over fitting. First, your NN maps R-800D dimensional inputs into R-3D output (hidden layer) and then R-3D input (output from a hidden layer) into R-2D output - the actual output from the network.