Conference paper

Appeared in: MIWOCI 2011, Mittweida Workshop on Computational Intelligence

Editor list: Schleif, Frank-Michael; Villmann, Thomas

Publication year: 2011

Pages: 5-15

URL: https://www.techfak.uni-bielefeld.de/~fschleif/mlr/mlr_06_2011.pdf

Conference: 3rd Mittweida Workshop on Computational Intelligence

Title of series: Machine Learning Reports

Number in series: 2011, 06

Abstract: 

By combining very steep squashing functions and normalization as parts of the cost function of generalized learning vector quantization (GLVQ) and its descendants, vector label misclassification gets directly minimized in the limit of classseparating sigmoids towards step functions. To cope with the resulting difficult optimization problem a switch from standard stochastic gradient descent to a quasi- 2nd order Newton batch optimization scheme is proposed. Results for weighted squared Euclidean distance (GRLVQ) and adaptive matrix metrics (MRLVQ) are faster obtained and usually show a better class discrimination than traditional implementations of GRLVQ and MRLVQ. Code is available online.

Harvard Citation style: Strickert, M. (2011) Enhancing M|G|RLVQ by quasi step discriminatory functions using 2nd order training, in Schleif, F. and Villmann, T. (eds.) MIWOCI 2011, Mittweida Workshop on Computational Intelligence. Mittweida: University of Applied Sciences. pp. 5-15. https://www.techfak.uni-bielefeld.de/~fschleif/mlr/mlr_06_2011.pdf

APA Citation style: Strickert, M. (2011). Enhancing M|G|RLVQ by quasi step discriminatory functions using 2nd order training. In Schleif, F., & Villmann, T. (Eds.), MIWOCI 2011, Mittweida Workshop on Computational Intelligence. (pp. 5-15). University of Applied Sciences. https://www.techfak.uni-bielefeld.de/~fschleif/mlr/mlr_06_2011.pdf