The use of multilayer perceptron (MLP) neural network (NN) as human chromosome classifier was studied. The MLP NN classifier was optimized in the sense of learning rate, momentum constant and training cycle, for the chromosome data. The MLP classifier learning curves were examined by measuring probability of correct test set classification for an increasing size of training sets. Only 10-20 examples were required for the MLP NN classifier to reach its ultimate performance disregarding the number of features used. To compare the results to relevant theory, we have calculated the entropic error (loss). The empirical dependence of the entropic error on the number of examples is highly comparable to 1/t function that is a learning curve.
|Title of host publication||WORLD CONGRESS ON NEURAL NETWORKS-SAN DIEGO, - 1994 INTERNATIONAL NEURAL NETWORK SOCIETY ANNUAL MEETING, VOL 3|
|State||Published - 1994|