Performance improvement of the BSB-Eidos neural network

Attempts to improve the performance of the BSB-Eidos neural network are presented. BSB-Eidos is a fully connected, unsupervised network that uses both hebbian and anti-hebbian learning to correct some of the flaws of the original BSB model. We found that changing the amplitude of anti-hebbian learning has a significant impact on the network´s learning speed, recall speed and recall accuracy. An optimal ratio between the gain coefficients of hebbian and anti-hebbian learning was found to be K_h/K_h^-=25. We also found that changing the network´s output function during recall can lead to substantial improvements in both recall speed and recall accuracy