The role of excitatory and inhibitory learning in EXIN networks

We propose modifications for the learning rules of Marshall's EXIN (excitatory+inhibitory) neural network model in order to decrease its computational complexity and understand the role of the weight updating learning rules in correctly encoding superimposed and ambiguous input patterns. The MEXIN (modified EXIN) models introduce mixtures of competitive and Hebbian updating rules. In this case, only the weights of the unit with highest activation are updated. Hence, the MEXIN networks require less computation than the original EXIN model. A number of simulations are carried out with the aim of showing how the models respond to overlapping, superimposed and ambiguous patterns.