A novel neurophysiologically-inspired self-organizing cerebellar memory framework

Abstract

The human cerebellum is a major brain construct that facilitates the learning and acquisition of motor and procedural skills. Computationally, the cerebellum functions as an associative memory with stable, fast and efficient learning based on supervised error-correction. The multi-layered Cerebellar Model Articulation Controller (CMAC) neural network is a classical computational model of the human cerebellum. CMAC possesses strengths such as fast training, local generalization and ease of hardware implementations. This subsequently motivates its prevalent use in engineering applications. However, several drawbacks are associated with the CMAC network. They are: (1) the curse of input dimensionality; (2) a constant output resolution; (3) the generalization-accuracy dilemma; and (4) convoluted network computations. These drawbacks are fundamentally due to the uniform quantization of the CMAC memory surface, where the CMAC computing cells are regularly spaced (allocated). Two main approaches have been used to resolve these deficiencies:' multi-resolution discrete and fuzzy quantization of the CMAC memory space. However, the solutions are suboptimal and they introduced high operational complexity to the CMAC network.