A comparative study on mirror image learning and ALSM
Toru Wakabayashi, Meng Shi, Wataru Ohyama, Fumitaka Kimura · 2003
In this paper, the effectiveness of a corrective learning algorithm MIL (mirror image learning) is comparatively studied with that of ALSM (average learning subspace method). Both MIL and ALSM were proposed to improve the learning effectiveness of class conditional distributions. While the ALSM modifies the basis vectors of a subspace by subtracting the autocorrelation matrix for counter classes from the one of its own class, the MIL generates a mirror image of a pattern which belongs to one of a pair of confusing classes to increases the size of the learning sample of the other class. The performance of two algorithms is evaluated on handwritten numeral recognition test for IPTP CDROMI. Experimental results show that the recognition rate of the subspace method is improved from 99.05% to 99.37% by ALSM and to 99.39% by MIL, respectively. Furthermore, the recognition rate of the projection distance method is improved from 99.13% to 99.35% by ALSM and to 99.44% by MIL.