An Association Memory Mapped Approach of CMAC Neural Networks Using Rational Interpolation Method for Memory Requirement Reduction

摘要

Cerebellar Model Articulation Controller (CMAC) is a table look-up neuron-computing technique. It has good generalization capability and convergence speed of learning is very fast. The CMAC technique is implemented by a memory mapping process through an association memory to the actual memory as the mapping function of table look-up model. In other words, the CMAC technique is a huge waste of memory learning model because it requires recording memory mapping information using extra association memory and needed given some actual memory to record weights information. In this paper, a novel association memory mapped approach of CMAC neural networks is proposed for association memory requirement reduction. Our proposed method uses a continued fraction of rational interpolation method to predict the mapped address of association memory. Our method only needs to record few coefficients of the continued fraction for each association memory mapping. From experimental results show that our proposed method can reduce effectively the association memory requirement and can also keep a good learning accuracy. Also this method provides a more flexible association memory mapped approach and it is useful for CMAC hardware implementation.