Discrete Synapse Recurrent Neural Network with time-varying delays for nonlinear system modeling and its application on seismic signal classification

Discrete Synapse Recurrent Neural Network (DSRNN) using fully Recurrent Neural Network (RNN) structure and Extended Kalman Filter (EKF) algorithm for its training is improved with time-varying delay in its recurrent connection. An additional shadowing network is employed and learned to choose appropriate time delays at the right time in order to increase the memory depth inside the recurrent connection efficiently. As a lumped nonlinear model in capturing temporal dynamics related between input and output sequences, DSRNN with time-varying delay is applied to a task of seismic signal classification to discriminate footsteps and vehicles from background which are recorded in the deserts of Joshua Tree, CA. Even though the smaller sized network was trained from a smaller set of training data due to slow convergence in training, the proposed classifier showed 0.6% false recognition rate for the recognition of human footsteps, 0.8% for vehicle, and 0.0% for background. The models were able to reject quadrupedal animal´s footsteps (in this study a trained dog). The system rejected the dog´s footsteps with 0.1% false recognition rate.