The CMAC (cerebellar model arithmetic computer) neural network, an alternative to backpropagated multilayer networks, is described. The following advantages of CMAC are discussed: local generalization, rapid algorithmic computation based on LMS (least-mean-square) training, incremental training, functional representation, output superposition, and a fast practical hardware realization. A geometrical explanation of how CMAC works is provided, and applications in robot control, pattern recognition, and signal processing are briefly described. Possible disadvantages of CMAC are that it does not have global generalization and that it can have noise due to hash coding. Care must be exercised (as with all neural networks) to assure that a low error solution will be learned.>

A learning control technique that uses an extension of the cerebellar model articulation control network developed by J.S. Albus (1975) is discussed, and results of real-time control experiments that involved learning the dynamics of a five-axis industrial robot (General Electric P-5) during high-speed movements are presented. During each control cycle, a training scheme was used to adjust the weights in the network in order to form an approximate dynamic model of the robot in appropriate regions of the control space. Simultaneously, the network was used during each control cycle to predict the actuator drives required to follow a desired trajectory, and these drives were used as feedforward terms in parallel to a fixed-gain linear feedback controller. Trajectory tracking errors were found to converge to low values within a few training trials, and to be relatively insensitive to the choice of control system gains. The effects of network memory size and trajectory characteristics on learning system performance were investigated.>

The authors describe a cerebellar model arithmetic computer (CMAC) neural network and its use in learning the inverse function necessary for deconvolution and nonlinear inverse filtering. Simulations are described that use random noise, telegraph, or bit string signals as inputs to linear and nonlinear systems to generate the signal to be inverse-filtered. Results are shown for linear systems with decaying sinusoidal impulse responses and nonlinear systems with memory having saturating nonlinearities. Examples with low noise and testing (nontraining) results with new random sequences are shown. The results show considerable promise.>

In this paper, a method of classifying objects is reported that is based on the use of autoregressive (AR) model parameters which represent the shapes of boundaries detected in digitized binary images of the objects. The object identification technique is insensitive to object size and orientation. Three pattern recognition algorithms that assign object names to unlabelled sets of AR model parameters were tested and the results compared. Isolated object tests were performed on five sets of shapes, including eight industrial shapes (mostly taken from the recognition literature), and recognition accuracies of 100 percent were obtained for all pattern sets at some model order in the range 1 to 10. Test results indicate the ability of the technique developed in this work to recognize partially occluded objects. Processing-speed measurements show that the method is fast in the recognition mode. The results of a number of object recognition tests are presented. The recognition technique was realized with Fortran programs, Imaging Technology, Inc. image-processing boards, and a PDP 11/60 computer. The computer algorithms are described.