Statistical spiking model for real world pattern recognition applications

This paper introduces a spike based pattern recognition method and applies it to a real world seismic event recognition application. The processing unit - building blocks - of the developed method is statistical and considers both temporal frequency of spikes and their timings. Upon a significant statistical change of the temporal timings of the input spikes, compared to the model it has been trained for, the state of the processor changes and an output spike is generated. For pattern recognition applications, we first generate spike trains by utilizing filter bank decomposition and then pulse width modulation. Second, dynamic programming is employed to decode underlying class(es) of the temporal spike trains to which they belong. The model and approach of this study has been applied for seismic event detection and recognition of vehicle vs. human footsteps vs. everything else. The system showed over 97% performance on the classification of the above mentioned events.