Neural networks, simply stated, are machines that learn. A neural network consisting of several neurons, which are modeled after the human brain, uses synaptic weights, or interneuron connection strengths, to store acquired knowledge, which can be used in the future. One form of neural networks, feedforward networks, uses multilayer perceptrons, which are layered sets of neurons. This type of network contains an input layer, one or more hidden layers, and an output layer. Further, these multilayer perceptrons can be trained in a supervised manner using the back propagation algorithm, which uses the output of a forward pass through a multilayer perceptron network with fixed synaptic weights to adjust the weights based on an error correction rule. The forward and backward passes are repeated until improvement in the training (or test) set is no longer realized. The error rate that this type of network produces can be further reduced by a method known as boosting. Boosting uses weak learners from the previous training set to construct a new training set for each round of forward and backward passes. Again, forward and backward passes are repeated until improvement in the training (or test) set is no longer realized.

For this paper, schemes for backpropagation alone and for backpropagation with boosting are constructed using the MATLAB technical computing language. Two architectures are used to test the alogrithms. Three variations to each architecture are tested several times to obtain statistically significant results. Comparisons between the two architectures reveal that backpropagation with boosting performs drastically better than backpropagation alone. Also, as the test variation becomes more difficult, the degree of improvement that boosting provides over backpropagation alone decreases.