A neural network (NN) filter/target-tracking system has been developed as reported in Pap et al. (1992). The design accepts and inputs signal data to a noise/target classifier which uses spectral estimation techniques to distinguish noise from real targets. In that design, the NN is used to calculate the coefficients of an auto regressive linear predictive filter. The current evolution of that design invokes the use of Lagrange multiplier methods to incorporate known characteristics of the noise vs. signal. A (linear) Hopfield NN is used to perform the constrained optimization to solve for the filter coefficients. This algorithm has been demonstrated on real stochastic data. The filter resulting from this process succeeds in reducing the noise, whose structure was learned by the NN. Not only did this approach reduce structured noise without target attenuation or the addition of a 'ghost' signal, but it also lowered the base level of the resultant signal significantly. The overall concept has been tested and validated using real data on a workstation and the hardware NN implementation has been validated. This concept has been tested on the AAC Multiple Instruction Multiple Data (MIMD) Neural Network Processor (NNP) hardware. Each processor runs at 140 million connections/sec with 8 K neurons. An expanded version of the system performs a total of a billion plus connections/sec. Unlike classical SIMD NN architectures, which are really general purpose array processors, this MIMD system architecture was custom designed for NN applications.
展开▼
