MuSeq recurrent oscillatory self-organizing map. Classification and entrainment of temporal feature spaces

Mauri Kaipainen, Cognitive Science Program, Indiana University,
Pantelis Papadopoulos, Indiana University, Computer Science Department, Cognitive Science Program,
Pasi Karhu, Cognitive Science Program, University of Helsinki


A recurrent oscillatory self-organizing map is introduced. This is an architecture that assumes oscillatory states, assigned to all units, indicating their "readiness-to-fire" or "exhaustment", and constitute an internal timing mechanism. The architecture feeds the vector of all these states back onto the input layer, making the units recognize and adapt to not only input but also to the entire history of activations. The self-organized map so equipped is capable of detecting sequences that are consistent over time in the input flow. The timing mechanism translates these sequences into real-time periodicities to which the model can automatically entrain itself. The network is shown to distinguish between sequence endpoints that differ only with respect to their temporal context and to entrain to the periodicity of simple data, provided that the initial wavelength is set close enough to the data periodicity. The trained network can also correct its wavelength after it is manual! ly set to 50% and -30% of its original value.