Figure 6.

Learning convergence of the AuGMEnT variants in the 12AX task. Minimization of the TD loss function during training on the 12AX task. (A) All networks show a good decay of the mean squared TD error, but they seem to converge to a non-zero regime and, in particular, the base AuGMEnT network (green) maintains a higher mean squared TD error level when compared to leaky control (blue) and Hybrid AuGMEnT (red). (B) Mean squared TD error associated with only potential Target cues X and Y. (C) Mean squared TD error related to only Non-Target cues.