
168 ◾ Biological Computation
5. Repeat all the previous steps for the next input pattern in the training
set.
6. Repeat all the previous steps for the whole training set for the next
epoch.
Unlike in the simple perceptron case, for the backpropagation algo-
rithm we don’t have guaranteed convergence. us, the obvious ques-
tion is when we should stop training the network—or how many epochs
to run. We can choose among several halting criteria: halt when the
error on the training set is small enough; halt when the size of the
weight updates for each epoch is small enough; halt when the network
gives satisfactory results on the t