The training data is always supplied in a vector form to the MLP/X control. The training is done in a stochastic update manner with the data supplied. It is also possible to batch train the network in the sense of going over the same data multiple times in a batch run, but the updating is still done for each data point, not in the "update once per data set batch style".

There are different learning algorithms available for use in MLP/X. These include variations of backpropagation such as plain vanilla backprop, as well as a second order Extended Kalman filter based algorithm (Multiple Extended Kalman Filter Algorithm - MEKA) that is capable of providing fast learning algorithm convergence.