Barycentric Correction Procedure - Efficient Threshold Unit Training Method for Constructive Algorithms

Herve Poulard and Said Labreche


Short Description

The BCP is an effective procedure for training threshold neurons. The classification is performed using geometrical properties of the patterns. Experimental results have shown that the BCP is more effective than the perceptron variants such as the pocket algorithm and the thermal perceptron. It is supposed to improve the learning speed of constructive algorithms and also provide better generalization.

Reference

Barycentric Correction Procedure, Herve Poulard and Said Labreche. Submitted for review to the IEEE Transactions on Neural Networks