Matrix-Pattern-Oriented Ho-Kashyap Classifier with Early Stopping
Zhe Wang, Songcan Chen, Zhisong Pan, Xuelei Sherry Ni · 2009
Matrix-pattern-oriented Ho-Kashyap classifier has been demonstrated to have a superior classification performance to its vector classifier. However, it is found that the matrixized classifier takes a large computational complexity for convergence in some cases. To overcome the disadvantage, this paper introduces the early stopping technique into the matrixized Ho-Kashyap classifier and presents a matrix-pattern-oriented Ho-Kashyap classifier with early stopping named MatHKES. The presented MatHKES adopts early stopping as a new regularization technique instead of adding a regularization parameter in the criterion. The proposed algorithm achieves: 1) a less running time; 2) a competitive or better classification performance; 3) an avoidance of over fitting in training.