| Unsupervised learning for nonlinear synthetic discriminant functions (2009) | |||||||||||||||
Abstract | |||||||||||||||
| It has been shown in previous work5'12 that the family of filters which includes the minimum average correlation energy (MACE) filter7 can be formulated as a linear associative memory (LAM)3 preceded by a linear pre-processor which changes depending on the optimization criterion. We have presented a methodology by which the MACE filter and other synthetic discriminant function (SDF) filters can be extended to nonlinear processing structures9 (i.e. nonlinear associative memories) resulting in improved performance with respect to generalization and out-of-class target rejection10. Our earlier focus was towards developing efficient training algorithms for computing a nonlinear discriminant function without changing the linear pre-processor. In this paper we discuss a nonlinear pre-processing method based on concepts of information theory. We show a simple unsupervised method by which input images can be nonlinearly transformed onto a maximum entropy feature space. | |||||||||||||||
Details der Publikation | |||||||||||||||
| |||||||||||||||