A learning algorithm for a CMAC-based system and its application for classification of ultrasonic signals

Abstract

We present a Cerebellar Model Articulation Controller (CMAC) neural network architecture that is suitable for continuous learning, by providing stability and plasticity to the network, which is difficult to achieve in the majority of the neural networks. We also present two learning algorithms, namely, (a) a learning process and (b) a re-learning process. The learning process involves transforming an input into a set of codes such that similar inputs are coded as overlapping codes, while dissimilar inputs are coded as disjoint codes. Each category is represented by a set of codes stored in a memory as a part of the training. An input is classified to a category if the majority of its codes are in common with the codes corresponding to that category. In the re-learning process, we extract the statistical information about each category in terms of the frequency of appearance of a code and the number of signals classified to the category. Based on this statistical information, new exemplars (ideal signals representing each category) for each category are created. Based on these new exemplars the network is trained using the learning process. The trained net is then used to classify ultrasonic signals obtained from a damaged sample. The classification results are compared with the classification results obtained from the standard C-scan imaging techniques and it is found that the classification by the CMAC system with a re-learning algorithm is 92% accurate.