This paper presents a novel multi-layer perceptron neural network arch
itecture selection and weight training algorithm for classification pr
oblems. The MLP iterative construction algorithm (MICA) autonomously c
onstructs an MLP neural network as it trains. Experimental results sho
w the algorithm achieves 100% accuracy on the training data, the same
or better generalization accuracies as Backprop on the test data, whil
e using less FLOPS. Moreover, relaxation of the hidden layer nodes imp
roves test set recognition accuracies to be greater than that of Backp
rop. Furthermore, seeding the Backprop algorithm with the hidden layer
weights from MICA is demonstrated. The MICA seeding improves the effe
ctiveness of Backprop and enables Backprop to solve a new class of pro
blems, i.e., problems with areas of low mean-squared error.