070110 Jaeger

The Bernstein Center for Computational Neuroscience Freiburg

Announcement for the next

BCCN Seminar

Dr. Herbert Jaeger

Fraunhofer Institute for Autonomous Intelligent Systems
Sankt Augustin
Germany

"ECHO STATE NETWORKS - INTRODUCTION, MEMORY PROPERTIES, EXAMPLES"

The talk introduces the "Echo State Networks" (ESN) approach to the supervised training of recurrent neural networks. ESNs are characterized by two properties: (1) a large recurrent neural network is used as a "reservoir" of excitable complex dynamics; this network is not changed by learning; (2) only the weights of output connections are learnt. An important aspect of ESNs is their long dynamic short term memory span. The basic theoretical insight is a randomly connected structure, such as an ESN, has a memory capacity which is in the same order as that of structures explicitly designed for short-term memory, for instance a delay line. Numerous examples are presented, e.g. chaotic time series prediction, text modeling, event detection for robots, speaker recognition, digit recognition, motor control, instantaneous music motif learning. The selection of examples can be made ad hoc according to wishes from the audience.

Host: Prof. Ad Aertsen