Figure 7
Adaptive time scales increase memory capacity. Three different model variations were tested on a memory capacity task. The number of time steps the input had to be remembered varied from 5 to 40. (A) The model using standard Elman units had the lowest memory capacity, performing better than chance up to memory lengths of 10 time steps. Both the model with global and local rate constants performed better than chance up to memory lengths of 30 time steps (dashed line represents chance level, which is defined as predicting the average over all time steps). (B, C) The models with adaptive rate constants learned slower rate constants with increasing memory lengths, increasing the capacity to maintain memories over longer time scales (error bars represent standard error over 20 repetitions).
