Figure 7.

Comparative statistics of the AuGMEnT variants on performance on the 12AX task. Barplot description of the learning behavior of the three networks on the 12AX task according to the convergence criterion given by Alexander and Brown (2015). After 100 simulations, we measured the fraction of times that the model satisfies the convergence condition (left) and the average number of training trials needed to meet the convergence criterion (right). Although training dataset consists of 1,000,000 trials, the standard AuGMEnT network never manages to satisfy the convergence criterion, while the leaky (blue) and hybrid (red) models have similar convergence performance with a learning time of about 30,000 trials.