Non-linear Systems Analytical Method

---

  Goal: Our goal is to quantify accurately the nonlinear dynamic characteristics of the neuronal activity over time (on-line) in order to detect reliably systemic changes caused by exposure to chemical-biological agents.   Progress: We have begun employing a pseudorandom, broadband stimulating pattern (i.e., Poisson process) and advanced modeling techniques suitable for the practical identification of nonlinear dynamic systems from stimulus-response data. We have begun using the Volterra model class to simulate this nonlinear dynamic stimulus-response relationship as it is the only generall model class with any realistic prospects for reliable estimation using real biological, i.e., noisy, data. We recently have developed a new, practical methodology for the purpose of estimating the Volterra model from noisy data which employs an artificial neural network architecture termed the Laguerre-Volterra network (LVN). The novelty of the LVN is in the use of a Laguerre filter-bank for input pre-processing and the employment of polynomial activation functions in the hidden units (instead of conventional sigmoid functions).