An improved multilayer neural model and array processor implementation

The authors present a method for obtaining faster learning in a multilayer neural network. The key ingredient is the concept of floating positive/negative thresholds used in the output neurons to interpret the output states. In a traditional multilayer perceptron, the output state is 1 or 0, depending on whether the activation value exceeds the fixed target threshold or not. The proposed approach determines the state of an output activation by comparing the difference between the output activation and the two floating positive/negative thresholds. If the output activation is closer to the positive threshold or negative threshold, then the output state is 1 or 0 respectively. The iterative learning process completes whenever it is decided that an activation value is closer to its target threshold. Simulation results show that the learning iterations for the model are a hundred times fewer than for the traditional multilayer perceptron. Mapping this multilayer neural net onto a ring systolic array maximizes the strength of VLSI in terms of intensive and pipeline computing