An improved hardware-realisable learning algorithm for pyramidal feed-forward pRAM based ANNs

Proposes a hardware-realisable training algorithm, modified from that proposed by Guan et al. (1992). Probabilistic random access memory (pRAM) based artificial neural networks (ANNs), trained using the improved algorithm (which lets the network itself decide the output coding it should use for classification), managed to easily overcome the hard learning problem facing architectures that contain hidden layers. Also, lower percentages of noisy training were needed to achieve similar or better results than those obtained using earlier algorithms without increasing the training time needed. Pattern similarity problems can be overcome by letting the network decide the codes. Initial simulation results indicate much quicker training times have been achieved with better generalisation. Further investigation is necessary to optimise the algorithm and to investigate the optimum number of hidden layers or units to be used in layers