Vytenis Sakenas

• In complex pattern recognition tasks data usually exhibits many local distortions which significantly disturb the recognition process. A method for extracting temporal features from a signal that are invariant to these distortions is presented in this report. The idea is to use Echo State Network to generate a rich high-dimensional representation of data. Temporal features are then extracted by finding projections of the high-dimensional representation that are minimally influenced by the selected distortions while still carrying most of the information about the underlying signal required for the performed task. The algorithm performance is analyzed on synthetic signals as well as on high-dimensional handwriting data for shift and scale distortions. It is shown that the algorithm is capable to extract a low dimensional feature set from a reservoir which is invariant to the selected distortions and relevant to the performed task.