neuralnetworks.error-fn
cross-entropy
(cross-entropy input thetas activation-nodes output lambda)
Suitable for classification problem. Note this cost function will not work well with hyperbolic tangent sigmoid function due to the presence of negative number
mean-squared-error
(mean-squared-error input thetas activation-nodes output lambda)
Suitable for linear regression problem (curve fitting)
regularization-cost
(regularization-cost lambda input-count thetas)
thetas is a vector of matrices