Design of a cerebellar model arithmetic computer with adaptive resolution and its application to nonlinear signal processing

Abstract

AbstractA cerebellar model arithmetic computer (CMAC) with adaptive input space quantization is designed. the quantization intervals are compressed by the segmentation of weights in the region where the function to be learned changes rapidly, whereas those are extended by the integration in the region where the function changes slowly. the adaptive resolution technique reduces required memory space and improves learning speed. Results of simulation studies on nonlinear signal processing problems are reported. A storage procedure for the weights and new criteria for the integration and segmentation are also reported.