Towards Self-organized Online Extraction of Invariances Using a Hierarchy of Multiple-timescale Reservoirs

Real-world processes often generate data that have regularities on many different timescales. A robot operating in such an environment can benefit from extracting these regularities in order to build models for prediction of future events, and for decision making. A number of methods exist which enable this kind of feature extraction, but they are usually computationally very expensive, prohibiting their use in an online learning setting. We present an approach based on a hierarchy of reservoir networks with different timescales and argue that this setup is well suited to detect regularities and invariances in the sensory stream of a humanoid robot. In line with ideas from reservoir computing, many connections of the hierarchy are randomly initialized and remain fixed; only a few connections are trained by unsupervised, local learning rules. Training for the prediction of future sensory data is done by a recursive least squares method enabling efficient online learning. Preliminary results using synthetic data are presented, and future issues are outlined.