MSC: | 68T05 | Learning and adaptive systems |
68T10 | Pattern recognition, speech recognition | |
68Q32 | Computational learning theory |
Abstract:
This article develops approaches to generate dynamical reservoirs of ESNs
with desired properties reducing the amount of randomness. It is possible
to create weight matrices with a predefined singular value spectrum. The
procedure guarantees stability (Echo State property) and minimizes the impact
of noise on the training process. The final reservoir types were already
known in the literature but related input weights were out of the scope of
previous articles. But our experiments show, that well chosen input weights
can improve performance.