Modules that do the weight update in batch mode (see Section 3.1.2 for explanation). All the statements in these modules have all the input patterns as columns of the matrices in the equations. Therefore although the contents of the file seem to be not very different from the incremental mode version, the functionality is different.
Here, all the inputs are presented to the network in a single command, and the
outputs of the network is obtained corresponding to all patterns in a matrix.
Then, the remaining weight update rules make use of this matrix to come up with
the total errors for a whole epoch, always working with matrix variables until
updating the weights and biases, where the matrices are collapsed by adding
all columns together due to the nature of matrix multiplications (see deltaBiasL2
actOutsL1', where ' stands for the transpose
of the matrix).