Options
2010
Conference Paper
Titel
Echo state networks with sparse output connections
Abstract
An Echo State Network transforms an incoming time series signal into a high-dimensional state space, and, of course, not every dimension may contribute to the solution. We argue that giving low weights via linear regression is not sufficient. Instead irrelevant features should be entirely excluded from directly contributing to the output nodes. We conducted several experiments using two state-of-the-art feature selection algorithms. Results show significant reduction of the generalization error.