This is performed in SCENT exactly as in the other DNTNs as described in section 2.2.
We use a learning rate of 0.05. We have not found it necessary to vary this setting or
any details of the learning schedule in any of the experiments reported here. The input
vectors are presented to the tree in epochs, but the order of presentation within an epoch
is random and different for each epoch. In general the number of epochs needed for
convergence was found to be small (of the order 15-35 epochs).