Multistate combination approaches for liquid state machine in supervised spatiotemporal pattern classification

Liquid state machine is a recent concept whose aptitude for spatiotemporal pattern recognition tasks has already been demonstrated. It consists in stimulating an untrained spiking neural network with input streams, creating complex dynamics that form the liquid state. An external function, the readout, is trained to map the liquid states into the desired outputs. In this paper, different avenues are explored to improve the classification performance of the readout. First are compared the membrane potentials and the firing rates of the neurons as two different liquid state representations. We also propose a new liquid state representation based on the frequency components of short-time membrane potential signals. Tests on synthetic and real data reveal that the frequency-based representation gets higher recognition rates than by using membrane potential or firing rates. Finally, we show that the combination of the different liquid states can improve the classification performance on spatiotemporal data.