LEARNING WITH FORGETTING: AN APPROACH TO ACHIEVE ADAPTIVE NEURAL NETWORKS (original) (raw)

Match Adaptive Resonance Theory (MART2) is developed as a modified version of self -organising Adaptive Resonance Theory (ART2) neural network for Arabic alphabet recognition. The new model does not utilise bi-directional synapses, match-reset loops and vigilance parameter. Novel subsystem is added to select the winning F2 node conserving competitive learning concept applied to reset wave. It relies on different sequence of operations of ART2 algorithm, but the classification of the input patterns remains unchanged. In the new architecture, algorithm execution takes almost equal time for each input pattern to be clustered and it has a new strategy in accessing an appropriate node in F2 without having bottom-up connections, generated from F1 to F2. However, top -down connections play an important role in matching and resonance. MART2 classifier of Arabic letters signals is implemented. The raw input signals are segmented and preprocess ed depending on two criterions, amplitude average and zero crossings rate which determine voiced and unvoiced frames. Fast Fourier Transform (FFT) is used to transform the signals from time domain to frequency domain. The most important features of the letters are extracted to reduce data size. The reduced data are then presented to MART2 for training and classification. ART2 and MART2 are employed for clustering Arabic letters. Experimental results show that the new algorithm of MART2 generally exhibit s faster learning, better clustering performance, lower error level, an improved recognition ability and more accuracy; even without a need of bottom up connections, match -reset loops and a vigilance parameter. That is, a major advantage of a flexible adaptive resonance theory neural network.