Fig. 4: Scaling behaviors of the memory capacity.
From: Reservoir-computing based associative memory and itinerancy for complex dynamical attractors
A Algebraic scaling relation between the number K of dynamical patterns to be memorized and the required size Nc of the RC network for various tasks and different memory coding schemes. Details of each curve/task are discussed in the main text. The “separate Wout” and “bifurcation” tasks are shown with prediction-horizon-based measures, while all the other curves are plotted with the region-based measure. B Examples of how the success rate increases with respect to N and comparisons among different coding schemes, where the success rate of accurate memory recalling (using the region-based performance measure) of three coding schemes on the dataset #1 with K = 64 is shown. All three curves have a sigmoid-like shape between zero and one, where the data points closest to the 50% success rate threshold correspond to Nc. All data points in the scaling plots are generated from this type of curve (success rate versus N) to extract Nc. The one-hot coding is the most efficient coding of the three we tested for this task. C Comparisons among different recalling performance measures. The three different measures on the dataset #1 with a one-hot coding have indistinguishable scaling behavior with only small differences in a constant factor.
