Figure 9.

Number of learning samples per reservoir for two different architectures applied to the same input. Every step where training of the ESN readout was activated is counted as a learning sample. (A) Learning samples for 8 ESNs trained on 9 different input patterns. (B) Learning samples for 20 ESNs trained on the same 9 input patterns. The results show that when the pool of ESNs is bigger that the number of features present in the input, only a necessary subset of ESNs from the pool is used to learn these features. The remaining ESNs are not trained, and can be used to learn new features from future inputs.